This PDF file contains the front matter associated with SPIE Proceedings Volume 13250, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Image segmentation technology is constantly advancing, especially in its application in medical diagnosis and treatment. Many segmentation tasks are based on the U-Net structural model method in convolutional neural networks, but this structure constrains feature extraction, and its global modeling ability still needs to be improved. Specifically, traditional convolution operations cannot capture feature information at different scales, leading to limited localization of local detail features and lower precision in global feature extraction. Based on these issues, we propose a new structure called RSFormer. The RSFormer architecture combines the Transformer's ability to extract crucial segmentation features in the main branch with a supplementary fully convolutional branch to address its limitations in full-size prediction, thereby enhancing its overall performance and applicability. By fusing the result features of the two branches, we can ultimately predict the segmentation map of h×w. Our method demonstrates its performance in terms of mDice, mIoU, and mPrecision metrics on three datasets benchmarks. These results indicate that the proposed RSFormer model has superior performance on multiple datasets.

In recent years, digital archaeology, as a new field of archaeological research, is accelerating the innovative combination of digital technology and archaeology and providing unprecedented opportunities for archaeological research. However, the current digital archaeology also faces some shortcomings, such as high technical threshold, complicated operation and low technical application. In this regard, this paper will start with the archaeological field investigation, analyze the application of three-dimensional image reconstruction technology in the archaeological field, and on this basis, develop a visual display system with the help of virtual reality technology, providing archaeologists with a brand-new research tool. Practice has proved that with the aid of Agisoft Photoscan 3D reconstruction software, the collected multi-angle images and photos of archaeological remains can be transformed into a real 3D model. Then, a virtual display system is developed based on Unity3D to complete the high-precision reproduction of archaeological remains, cultural relics and models. The whole process intuitively reflects the practical effect of virtual reality three-dimensional image reconstruction technology in the field of archaeology, and provides some technical reference for the later protection and presentation of cultural relics.

With the advent of the digital age, the demand for data storage has surged, especially in scenarios where disk arrays serve as backend storage systems. Disk arrays, with their high reliability and performance, have become the ideal choice for data storage. This paper introduces a novel data allocation strategy, Pdbs, aimed at addressing the issue of uneven data distribution within disk array storage systems. By preemptively identifying and processing duplicate data during the backup process of data streams, the Pdbs strategy achieves a more uniform distribution of data, thereby enhancing the parallelism of data access and the read performance of the backup system. Experimental results demonstrate that, compared to the round-robin placement strategy, the Pdbs strategy significantly improves load balance and accelerates the data recovery process.

In the process of automatic berthing, the Marine Autonomous Surface Ship ( MASS ) needs to perceive its own state relative to the berth in real time and accurately to avoid collision. The camera mounted on the drone can act as the 'eyes' of the ship, and extending its "arm" allows for a clearer view, while there is no obvious accurate feature on the berth for drone visual recognition. Therefore, this paper proposes a MASS berthing state perception method based on UAV visual recognition ArUco code. The method consists of three parts : camera calibration, ArUco code detection and berthing perception. Firstly, the camera calibration is carried out by the calibration method proposed by Zhengyou Zhang. Secondly, the ArUco code is identified by image gray processing, thresholding, Gaussian filter illumination compensation and other methods. Then, a berthing state perception model is proposed, including the calculation methods of bow and stern berthing distance, approaching speed, forward speed and ship berthing angle. Finally, a test platform was built and a berthing experiment was carried out. The accuracy of the method is verified by real ship test, which meets the requirements of real-time and accuracy for berthing state.

Many algorithms have been developed for no-reference image quality assessment (NR-IQA). While various problems will arise in the process of designing the network structure, such as feature loss, gradient disappearance and insufficient generalization ability and so on. This work introduces a deep multi-scale residual CNN for NR-IQA to address the issues. The significant improvement in our model's effectiveness is attributed to the optimization of traditional residual networks by eliminating unnecessary modules. The residual structure to form a deep network, which consists of several residual blocks with skip connections. Residual block contains convolution kernels of different sizes to adaptively extract image features, which are merged with each other to obtain effective image information. Then, use the output of each multiscale residual block as the hierarchical feature of global fusion, so as to retain as much image information as possible for a final quality prediction. An auto-encoder based method is introduced to reduce the dimension of local feature in multiscale residual block to improve training efficiency. Experiments show that the model is superior to the existing state-of-the-art methods in the prediction of various distortion types in the LIVE dataset. The results of cross dataset validation on the TID2013 and CIQA datasets also reflect the better performance of the model. Finally, in order to prove the generalization ability of the model, we generate Gradient Class Activation Maps (Grad-CAM) to compare with VGG16 and Resnet18 for feature visualization. The results of our experiments reveal that our model possesses strong robustness and excellent generalization capabilities.

Computer-generated holography (CGH) has undergone significant advancements in recent years. However, there are limited CGH methods designed for volume data. The realization of holographic visualization of volume image data with precise depth perception remains a challenging task. In this work, we propose a layered angular spectrum CGH method tailored for volume image data to achieve neuronal holographic visualization with accurate depth perception. The experimental results demonstrate that the proposed method enables holographic visualization of neuronal volume images with accurate depth perception. Moreover, we harness GPU parallel processing for hologram generation, achieving satisfactory temporal performance.

The accurate alignment of multimodal remote sensing images is of paramount importance for a wide range of applications, including environmental monitoring, change detection, and resource management. These tasks require robust registration techniques capable of handling the inherent variations in image appearance due to different sensor modalities. In this study, we introduce a novel approach that leverages synthetic data to improve the performance of image registration networks for multimodal remote sensing imagery. Our method involves the generation of synthetic image pairs through color-space transformations applied to synthetic data, which are then utilized to fine-tune a neural network specifically designed for image registration tasks. By training on these synthetic pairs, the network learns to cope with the complex variations between different sensor modalities. The fine-tuned network demonstrates state-of-theart (SOTA) performance when applied to real-world remote sensing images, showcasing the effectiveness of synthetic data in enhancing the network’s ability to generalize across diverse imaging conditions. This advancement represents a significant step forward in the field of remote sensing image analysis, offering a robust solution for the precise registration of multimodal datasets.

As an outstanding method of rapid measurement, called coded aperture compressive snapshot imaging of moving target shows the unique advantages in overcoming the problem of quickly position of high-speed targets. This technology adopts an advanced coded aperture modulation designing strategy, which allows that the capture of high resolution and multiframes moving target images can complete in a single exposure imaging process. That greatly improves the imaging system's ability to capture high-speed targets, also compressive sensing (CS) can improve the performance depending on the advantage of high-resolution recovered from the subsampling. Furthermore, that benefits the precise object identification. This paper proposed a coded aperture compressive snapshot imaging (CACNSI) model for moving target based on the CS sampling theory. That also constructs the optical prototype, which uses the coded aperture to modulate the moving scene and detects modulation of multi-frame in a snapshot, and the high-resolution multiple frames can be reconstructed using CS algorithm. The model is validated by the simulation experiment. Through experimental data, it is confirmed that the mathematical model has outstanding performance in reducing noise, enhancing image sparsity features, and improving imaging quality. Further experimental results show that the CS method can significantly improve the tracking accuracy of high-speed moving targets and strengthen the target localization and recognition capabilities. In summary, the mathematical model of single exposure imaging of encoded aperture moving targets is not only innovative in theory, but also proves in practice that it has a significant contribution to the improvement of imaging quality in complex dynamic scenes.

During magnetic resonance imaging (MRI) scanning, the patient's movement may introduce motion artifacts, leading to a decrease in image quality and potentially affecting diagnostic accuracy. To address this issue, this work proposes a new retrospective motion correction method called SUPMARNet. This method is based on Unet++, retaining its framework structure and replacing the convolutional layers of its encoder and decoder with Swin Transformer Block. Proposed model integrates the features of Unet++'s multi-layer feature fusion with Swin Transformer's capacity for capturing long-distance dependencies, thereby optimizing the utilization of distinct features from each layer and enhancing the model's capacity to recognize and eliminate motion artifacts. The experimental results demonstrate that the model can accurately recognize and remove motion artifacts while preserving image details, outperforming existing comparative methods.

A new neural network model called Tensorial Restoration Neural Network Model (TRNNM) is proposed, which can repair the noise, blur, and information missing of digital textual images. It has a different structure from traditional classification convolutional neural networks which have deconvolution layer inside to perform dimensionality enhancement and inpainting on the repaired image data. After experimental verification of the proposed network model, the processed handwritten images are showing good restoration performance, which is much better than some traditional methods.

Vessels as the main carrier of ocean transportation, when the ship is completed with intelligent transformation, the target detection and classification recognition problems of such ships will have an important impact on maritime traffic and national defense security. However, to enhance target detection performance, it often requires the use of larger and deeper neural networks, which in turn consume a substantial amount of computational resources and thus energy. This is particularly troubling for non-nuclear-powered ships that have limited energy resources. To address this issue, this paper presents an improved ship image classification method, the Dynamic Convolutional Network and Triplet Attention Mechanism Integrated ResNet (DY-ResNet-TA). Through theoretical analysis and experiments based on a ship image dataset, the proposed method has been proven to significantly improve model performance with a minimal increase in computational resource consumption.

Early diagnosis of oral cancer is particularly important as it can significantly increase the chances of successfully treating skin cancer and recovering. However, Histopathology image analysis (HIA) traditionally used to perform is time-consuming and requires expert knowledge. Therefore, we propose an oral cancer Multi-class segmentation method called Sc-deeplabv3+. Our model introduces a module called SCConv into the deeplabv3+ model, which uses convolutional neural network (CNN) to compress spatial and channel redundancy between features to highlight important features. We evaluated our model on the ORA-N dataset. The experimental results demonstrate that our proposed model can effectively obtain good performance in the task of oral cancer segmentation. The average delivery of the dice coefficient is 82%, which is 2.7% higher than the baseline deeplabv3+, verifying that our method outperforms existing methods.

Deep convolutional neural networks have achieved remarkable progress in medical image segmentation, yet these large-scale models often require extended inference times. There is an urgent need for lightweight models, and knowledge distillation has emerged as a pivotal focus of current research. This paper introduces a novel knowledge distillation framework, designated as ISKD. Unlike traditional methods, we introduce an innovative approach to address this complex challenge: refining features from the perspectives of feature subspaces and inter-class similarity to facilitate student models to more effectively absorb the structural and contextual information from teacher models. Specifically, we propose to first learn a set of reconstruction basics within the feature space, then select corresponding basics within the signal subspace, and project the input data into this space for feature denoising. Furthermore, this approach employs KL divergence to compute the similarity matrix between class distributions, thereby offering a more precise assessment for distillation. A large number of experiments have been conducted on two famous medical image datasets, Synapse and Flare2022, to verify the effectiveness of this method in DSC, and it is superior to most existing knowledge distillation methods.

In recent years, increasing the depth of field of optical imaging system has been a research hotspot in the field of applied optics. This is because the optical imaging system with large depth of field can be applied in many fields. For example, it can be applied to machine vision, biological or medical microscopic imaging, optical remote sensing and other fields. However, the hardware structure of the traditional zoom system is complex, the zoom process is cumbersome, and the zoom range is limited. It is difficult to meet the increasingly complex and diverse industrial application requirements. In order to obtain reliable images quickly and efficiently, new zoom imaging methods must be introduced. In this study, a system of increasing depth of field based on image clarity is designed by using liquid lens. The system includes two parts : hardware and control. The hardware part integrates the liquid lens into the optical system, eliminating the use of mechanical devices. It only needs to use the camera, ordinary lens and liquid lens to realize the function of the traditional mechanical zoom device, greatly improving the image acquisition efficiency and imaging quality, and effectively overcoming the limitations of the traditional lens depth of field. The control part mainly consists of three parts : image acquisition, autofocus and image fusion. The system is specially used for large quantities of object detection with height difference. An experimental device was built with a liquid lens, a common lens and a camera used in the laboratory. In the study, a large depth of field imaging method is designed according to the zoom characteristics of the liquid lens. The focus images of different regions are obtained by using the sharpness evaluation function, and the large depth of field image is obtained by fusion. Large depth of field imaging experiments are carried out in different scenes. The results show that the proposed method overcomes the limitation of the depth of field of the lens, effectively expands the depth of field of the lens imaging, and realizes the focused imaging of objects with height difference. It has potential application value in industrial detection, medical diagnosis and many other directions.

With the rapid development of the Internet, especially the extensive application of deep learning, change detection has been successfully applied in many fields, but the pursuit of greater feature information content has led to an increase in memory and computing power requirements. To solve this problem, a Lightweight Siamese attention Residual Network (LSRNet) is proposed in this paper to reduce the memory and computational power requirements, embed the Siamese Fast Small Attention (SFSAttention) to filter out feature information with less relevance, and then use feature fusion of channel dimensions to ensure feature richness and reduce the number of parameters of the model. The residual network module is introduced to extract the entire feature information and obtain the true change graph. Under the condition of ensuring accuracy, the number of parameters in the LEVIR-CD and CCD datasets is reduced by 3.7 M and 4 M, and the number of FLOPs is reduced by 2.5 G and 2.9 G, respectively.

This paper explores the dependencies between inputs and outputs in deep neural networks under adversarial conditions from the perspective of information theory, particularly focusing on the dependencies on low-frequency and high-frequency patterns. We propose and validate a novel adversarial defense strategy that enhances the robustness of deep learning models by separating and optimizing the mutual information of low and high frequencies. Experiments demonstrate that this method effectively improves the model's defensive capabilities against complex adversarial attacks.

In order to better capture image information and focus more on key areas of images, this paper proposes an image description method based on the encoder-decoder framework that integrates the Selective Kernel (SK) mechanism with a soft attention model. The SK attention adaptively adjusts the weights of convolution kernels at different scales, allowing the model to autonomously select the most suitable kernel for feature extraction based on the content of the image, thus better handling multi-scale information in images. Additionally, this paper integrates a soft attention model, which calculates attention weights for different regions within the image, enabling the model to focus more on areas relevant to the current descriptive vocabulary when generating descriptions, resulting in more accurate and detailed descriptive statements.

With the development of the information technology era, the exchange of information between people is gradually increasing, and digital images, as the most commonly used information carrier, need more protection. The encryption method based on chaotic systems for digital images often has broad application prospects in the field of image encryption due to its similarity with cryptography. This paper improves a new three-dimensional discrete hyperchaotic system based on the traditional one-dimensional logistic chaotic mapping and analyzes the chaos of the system. When encrypting images, the commonly used image encryption techniques are scrambling and diffusion; This paper first pre-processes the image, then uses an improved Arnold map to scramble the image, and finally diffuses the pseudo-random sequence generated by the three-dimensional discrete hyperchaotic system with the scrambled image. Through experimental analysis of image histograms, image information entropy, and other data, it has been proven that the improved system has a good encryption effect on image encryption.

Using deep learning technology to detect individual bamboo trees in UAV images is a valuable research field. Similar to other object detection tasks, in bamboo detection scenario, data augmentation can greatly increase the number and appearance diversity of bamboo samples, making the trained model has good generalization ability and performs better when predicting bamboo trees. Since different data augmentation operations have different effects on improving the performance of the model, just treating them equally and randomly selecting and combining some of them when training bamboo detection model will not achieve the best result. In this study, we collected 50 km2 of UAV forest aerial images and labeled more than 10,000 bamboo trees as the dataset. Then we analyzed a variety of data augmentation operations suitable for individual bamboo detection scenario and added them into candidate operations. On this basis, we proposed a novel data augmentation operations combining strategy, assigning different weights to different operations and dividing them into different categories, according to their degree of improvement of bamboo detection results. The experimental results shown that compared with the baseline YOLOv5x model without any data augmentation, all selected candidate data augmentation operations could improve the detection result to a certain extent. The proposed operations combining strategy achieved an impressive performance, reaching 91.1%, 85.1% and 79.1% in recall, precision and mAP respectively. Compared with commonly used RandAugment combining strategy, in our strategy, recall was increased by 4.47%, precision was increased by 1.55%, mAP was increased by 6.89%. In order to further verify the generalization ability of the model, we applied strategies to new bamboo forest, the results showed our combination strategy also had advantages in all recall, accuracy and mAP.

Early detection of skin cancer significantly improves the five-year survival rate for patients. However, early-stage malignant tumors exhibit very subtle lesions in the skin, making their symptoms less noticeable. Professional doctors require multiple biopsies and tissue extraction to diagnose the type of lesion. Existing machine learning methods struggle to focus simultaneously on spatial detail and metadata semantic features, resulting in lower recognition accuracy in skin lesion images. To effectively represent spatial and metadata features and prevent issues such as misclassification of easily distinguishable images due to an overemphasis on detail, a new skin lesion classification network called Metadata-P2T-Net has been proposed. This method uses a P2T-based visual backbone network and incorporates a Multi-Head Attention Feature Fusion Module (MHAFFM), enhancing the model's capability to model global context and dynamically weight visual features with metadata information, highlighting metadata-related visual regions. Additionally, a Multi-Dimensional Fusion Residual Module (MFRM) has been introduced, which effectively integrates visual, fusion, and metadata features to improve the model's classification performance. Experiments conducted on the ISIC2019 dataset show that the Metadata-P2T-Net achieved a Matthews correlation coefficient (MCC) of 0.9086, an accuracy of 93.69%, a precision of 93.89%, a recall of 86.77%, and an F1-Score of 90.18%. Compared with other methods, these metrics significantly improved, demonstrating the effective extraction of detailed and metadata features and providing a new basis for the diagnosis of dermatoscopic images.

This study examines the use of UAV technology to capture images of construction crane bolts and process them for denoising and enhancement. Traditional image acquisition methods often struggle to capture clear bolt images due to the variability of the construction site environment. To address this issue, this study employs UAVs for automated image acquisition, resulting in significantly improved efficiency and image quality. To tackle potential noise and illumination issues in the resulting images, this paper presents a range of image pre-processing techniques. These include median filtering, mean filtering, Gaussian filtering, and efficient denoising methods such as the C-BM3D algorithm. Additionally, histogram equalization and image enhancement strategies based on the Retinex theory are employed to further optimize image quality. The experimental results indicate that both the C-BM3D algorithm and the improved Retinex algorithm have a significant effect on enhancing image clarity. This study’s findings are important not only for improving the safety and efficiency of construction sites but also for advancing image processing technology.

Remote sensing image change detection technology is rapidly advancing under the impetus of deep learning. In this study, a remote sensing image change detection method based on hybrid backbone and high and low frequency attention modules is proposed for the problems of insufficient extraction of feature components of traditional convolutional neural networks in this field of remote sensing image change detection. The method adopts a hybrid trunk network, and through the attention feature fusion module, the features of the two branches of convolutional neural network and Transformer are fused, which can take care of the extraction of local features and global information. Further, this study integrates high and low frequency attention modules to refine the high frequency details and low frequency background information in the image respectively. The implementation of this method significantly improves the quality and depth of feature extraction. Ultimately, the ability to discriminate and extract the location information and features of interest is strengthened by the coordinate attention module, which improves the recognition accuracy of local details. After extensive experimental testing and validation, it is confirmed that the proposed model achieves a significant improvement in performance compared to existing change detection models.

Clinical practice significantly benefits from the precise delineation of medical images, which aids in the precise identification of conditions, the formulation of therapeutic strategies, and the ongoing surveillance of illnesses. However, current approaches to segmentation are predominantly designed for particular image patterns or specific disease categories. Given the infrequent occurrence and the broad spectrum of mediastinal tumor types, achieving an accurate diagnosis solely through visual inspection presents substantial challenges. This design is a deep learning medical image segmentation system based on MedSAM, which has processed a series of custom datasets and trained them based on pre trained models to enable semantic segmentation and detection of lesion areas.

We propose Cross-ViT GAN, a novel image restoration model that addresses the limitations of CNNs by integrating Cross-ViT into GANs. The Cross-ViT GAN includes both generators and discriminators. In the generator, we replace the original normalization layer with a self-modulation layer, allowing the intermediate feature map to adapt to the input noise vector during backpropagation, thus enhancing generated image quality. The discriminator incorporates Cross-ViT, dividing the input image into large and small branches and using cross-attention to exchange features between different scales, enhancing image features for better recognition. We use three datasets, CelebA, Palces2 and CIFAR-10, to evaluate the performance of the model. The experimental results show that the evaluation metrics PSNR:26.48, SSIM:0.923, IS:4.11, and FID:1.65 of the Cross-ViT GAN model in the CelebA dataset outperform those of other comparative models, which proves the effectiveness of the Cross-ViT GAN model.

Coronal holes are the primary source of high-speed solar wind, and the energetic particle flow produced by solar storms can impact the magnetic field environment near Earth, sometimes causing geomagnetic storms. Starting with coronal hole images, this paper explores how to extract coronal hole characteristic parameters from SDO/AIA and SDO/HMI Magnetogram of solar extreme ultraviolet images captured by SDO satellites, and then predicts geomagnetic disturbance by focusing on magnetic field characteristics and physical characteristics of coronal holes through HMI Magnetogram. After segmentation, alignment and other image preprocessing operations, this paper proposes and quantifies 11 characteristic parameters, such as Magnetic Unipolarity, Average Grayscale Value, Coronal Hole Center Location and Area, to establish statistical relationships between coronal holes and geomagnetic storms. Based on the Vision Transformer model, this paper adds the manually extracted magnetic field physical characteristics, which reduces the prediction error RMSE by 0.004, and increases the Correlation Coefficient by 0.034, which enhances the interpretability of the deep learning prediction model. It provides a research basis for the accurate prediction of geomagnetic storms.

Composite images generated through simply merging objects with background images often demonstrate a pronounced lack of realism due to the lack of shadows, an important visual element. Therefore, shadow generation plays a critical role within the domain of image synthesis. Existing shadow generation methods primarily employ Generative Adversarial Networks (GANs) as the underlying framework. However, GANs often overlook critical guiding features in background images due to mode collapse, leading to poor shadow generation quality. To address this concern, we have designed a two-stage shadow generation method based on Latent Diffusion Model (LDM). The non-adversarial training process allows the diffusion model to effectively avoid mode collapse. Specifically, in the first stage, we introduce a non-local attention mechanism and design a focal loss function based on Gaussian activation to generate high-quality shadow masks. In the second stage, we treat the shadow shading task as a special style transfer task to extract shadow shading guide features from background images, and design a partitioned dynamic weight loss to shade the masked areas as shadows. We demonstrate the effectiveness of our model through a series of experiments on the DESOBA dataset.

With the rapid development of deep learning, road traffic target detection has become increasingly popular in intelligent video surveillance and vehicle-assisted driving. However, traditional road traffic target detection algorithms have limitations in terms of generalization ability and recognition rate. To address this issue, this paper proposes a road traffic target detection model named LKF-YOLO. To begin with, we propose a novel and efficient down-sampling module to replace the existing down-sampling structure in YOLOv8. Furthermore, by incorporating KaernelWareHouse, we have successfully mitigated the substantial computational expense of C2f modules in YOLOv8 while enhancing the model’s representative capacity. Finally, we have incorporated Focal Modulation Networks to replace SPPF, thereby augmenting the model’s capacity to assimilate contextual information. The experimental results demonstrate that the mAP50 index of LKF-YOLO has increased by 2.3%, the recall rate has increased by 4.7%, and both parameters and GFLOPs have decreased. These findings indicate that our improvement strategy effectively enhances the detection accuracy of the model and reduces its size. This study provides an efficient and accurate solution for road traffic target detection, with broad application prospects.

The objective detection of traffic signs in scene images is a research problem that entails identifying traffic sign targets within a given visual context. In recent years, this pursuit has found extensive applications in road scene monitoring systems and autonomous driving technologies. Consequently, the optimization of traffic sign object detection holds paramount significance. This study focuses on establishing a model for real-time detection of traffic signs in the urban environment of Shanghai. Leveraging a dedicated traffic sign dataset, training was conducted utilizing the SSD (Single Shot Multibox Detector) neural network for object detection. The results indicate an achieved error rate of approximately 5%. With few exceptions related to challenging lighting conditions or instances where the size of the traffic signs proved to be excessively small, the model demonstrated favorable outcomes across the majority of images.

Industrial defect recognition is a key part of product quality control. Small-scale defect recognition has become a difficult task in defect detection due to the weak and easy loss of recognizable features. To address this problem, this paper proposes an end-to-end small-scale defect detection algorithm based on RT-DETR (Real-Time DEtection TRansformer) model. First, we feed a shallow texture feature layer into the efficient hybrid encoder and optimize the cross-scale feature fusion module with RepC3 block to extract richer texture features for small-scale defects; Second, we utilize IoU-aware Query Selection for feature screening to ensure the reliability of the features input to the decoder for prediction; Finally, we conduct comparison experiments on the self-made WM-FSD dataset and HRIPCB dataset, and the experimental results demonstrate that our proposed method for small-scale defect recognition in different scenarios has obvious advantages in terms of precision, recall, and average precision.

To address rainwater interference in visual imaging systems for water navigation, exploration, and rescue by proposing an improved QSMD network based on QSNet. To mitigate channel information loss, we combine guided astrous convolution with CBAM, enhancing the model's spatial representation and feature expression while optimizing computational efficiency and maintaining high accuracy. Experimental results confirm the effectiveness of this improvement.

In this letter, we present an autonomous navigation system driven by dual target points, designed for goal-directed exploration of unknown environments through multi-agent deep reinforcement learning. The system identifies candidate target points from the environment that could serve as potential navigation directions and selects the best two waypoints based on available data. Guided by these waypoints, robots are directed towards global targets. Utilizing multi-agent reinforcement learning, a movement strategy is trained in a simulated environment using LiDAR sensor data and pose information, enabling two robots to move towards the currently selected target points. Experiments demonstrate that our method effectively conducts target point exploration, and the collaborative exploration by two mobile robots significantly enhances the efficiency of global target point discovery.

Raveling on asphalt surfaces can lead to the occurrence of diseases such as loose potholes on the road surface. Monitoring and real-time warning of the development status of pavement raveling helps formulate road maintenance plans in time and effectively curb the occurrence and development of pavement diseases. To realize the intelligent identification of asphalt pavement raveling, the LS-40 portable 3D surface analyzer was used to collect the image data of asphalt pavement raveling and normal asphalt pavement, and the recognition effects of traditional image classification and recognition models supported vector machine (SVM), random forest and convolutional neural network (CNN) were compared. The results show that the random forest model has a good accuracy of 82% for the recognition of pavement raveling images based on the traditional machine learning model on the high-precision texture data dataset, and the convolutional neural network has an accuracy of 93%, which is 11% higher than that of random forest. The results verify the efficiency of the convolutional neural network in the identification of pavement raveling diseases.

Unmanned aerial vehicle (UAV) aerial imagery typically presents challenges such as complex backgrounds, small-sized targets, and wide fields of view, making object detection more difficult. So, this paper proposes a novel method that improves upon the YOLOv7 algorithm. Firstly, it introduces the BiFormer attention mechanism to help the model effectively capture the correlation between target objects, enabling better identification and localization of small objects, thus providing more accurate object detection results. Secondly, to address the issue of standard convolution's inability to share parameters, the RFAConv convolution module is embedded in the E-EALN module to extract spatial features of the receptive field, further enhancing the model's performance. Lastly, the boundary box regression loss function is improved to optimize cases where predicted boxes have the same aspect ratio as ground truth boxes but completely different width and height values, a problem existing boundary boxes cannot address. It demonstrate that the YOLOv7 algorithm haves a 1.5% added in mAP0.5. Additionally, it is better suited for tiny target detection tasks.

In order to solve the problem of unclear foreground targets and false foregrounds after performing target detection tasks in foggy video images with wide fields of view, we propose an improved video target detection algorithm for the CSLBP operator to detect foreground targets in different scenes. The improved CSLBP operator is used to extract the texture features of the wide-field image, and the foreground target pixels are judged combined with background modeling. This method is not sensitive to changes in indoor lighting or outdoor sunlight, can effectively solve the shadow problem, and more completely extracts foreground moving targets. It also has obvious effects in bad weather such as foggy weather, and has good robustness.

The development of an automated soybean seed counting tool is instrumental in predicting yield automatically before harvest. Traditional counting methods are labor-intensive, prone to errors, and time-consuming. To facilitate rapid and accurate detection and counting of soybean seeds, as well as to enhance the breeding speed and level of soybeans, this paper proposes a lightweight YOLOv4 model named SMDC-YOLOv4. Built upon the original YOLOv4 architecture, MobileNetV1 is employed as the backbone network to streamline the model. The Spatial Pyramid Pooling (SPP) structure is enhanced into the Spatial Pyramid Pooling—Fast (SPPF) structure to improve model speed, and depth-wise separable convolutions are applied to the feature fusion network. Activation functions are replaced with the more comprehensive Mish activation function to enhance model stability. Additionally, the CBAM attention mechanism is integrated into the PANet network to further enhance model recognition accuracy. The improved SMDC-YOLOv4 model achieves an accuracy of 79.19% and a recall rate of 30.06%, with an average precision (mAP) of 52.33% and an F1 score of 0.44. This lightweight model demonstrates outstanding detection performance, striking an optimal balance between soybean seed detection efficiency and recognition accuracy.

In recent years, the application of automobile driver assistance system has improved road traffic safety. However, the capability of traditional target recognition model is limited in the complex and changeable situation of actual road, which often leads to the confusion of pedestrian and vehicle target recognition results. In order to realize the joint detection of pedestrians and vehicles, a deep neural network model suitable for pedestrian and vehicle detection is established based on the object detection method of fast regional convolutional neural network. Aiming at the problems of frequent false detection and missing detection, poor detection effect of small-size targets, complex and changeable background environment, etc., a variety of network improvement schemes such as edge extraction of pedestrian and vehicle images, multi-layer feature fusion and multi-target candidate region input are designed respectively, so as to improve the detection effect of pedestrian and vehicle targets.

The identification information on industrial X-ray films is crucial for film classification and management. In order to solve the problem of low detection accuracy and high missed detection rate, this paper proposes a model for X-ray film identification information recognition, named XRF-YOLOv5 (X-ray Film-YOLOv5), based on Adaptive Spatial Feature Fusion (ASFF). Experimental findings indicate a notable enhancement in the efficacy of the XRF-YOLOv5 model in detecting identification data on industrial X-ray films. In practical application scenarios, compared to the original YOLOv5 model, it exhibits higher recognition accuracy, higher confidence in recognition, and greater robustness. Therefore, XRF-YOLOv5 meets the basic requirements for industrial X-ray film identification information recognition.

The conventional image recognition method in electric power altitude working scene mainly uses DPM (deformable part model) deformation component model, which is vulnerable to the changes in the micro scale of personnel, resulting in poor image recognition performance. Therefore, a method of image recognition of personnel hazard sources in the electric power aerial work scene considering the masking effect is proposed. That is to say, the identification characteristics of the hazard source image of the personnel in the electric power aerial work scene are extracted, and the image recognition algorithm of the hazard source image of the personnel in the electric power aerial work scene is designed considering the masking effect, thus completing the hazard source image recognition. The experimental results show that the designed electric power aerial image recognition method has high recall rate, accuracy rate, average intersection/merge ratio, accuracy mean and F1 value, which proves that the designed image recognition method has good recognition effect, reliability and certain application value, and has made certain contributions to improving the quality of electric power aerial work and reducing the risk of electric power operation.

Aiming at the semantic complexity and the difficulty of capturing long-distance dependencies in the task of named entity recognition in Chinese resume text, a BLDC (BERT-BiLSTM-Dense-CRF)-based named entity recognition method for Chinese resumes is investigated. The model first uses the pre-trained BERT-Base-Chinese model to obtain the semantic information of the text, then learns the contextual information of the sequence through the BiLSTM layer, followed by mapping the output of BiLSTM to the label space through the Dense layer, and finally uses the CRF layer to consider the dependencies between the labels for sequence annotation. Experimental results on public datasets show that the method outperforms other models in terms of precision, recall, and F1 value, and effectively improves the performance of named entity recognition in Chinese resume text.

Aiming at the issues of low object detection accuracy and false detection of small object in complex background, a modified remote sensing image detection algorithm based on YOLOv8 is proposed. Use of the RFA convolution kernel enhances the extraction of network space features and their associated global information, and further bring in the Deformable Attention mechanism to obtain a more flexible dynamic attention model. The Powerful-IoU loss function is introduced in the regression loss function, which improves the convergence speed of the network and the object detection accuracy. On the RSOD and DOTAv1.0 datasets the mAP of the improved algorithm is increased by 1.5% and 2.3% respectively.

This study proposes a lightweight YOLOv8 detection model to tackle the performance and efficiency challenges in detecting small targets on printed circuit boards (PCBs) with surface defects. Firstly, the CA (Coordinate Attention) mechanism is integrated into the core of the model to enhance its capability in processing PCB image features, thus boosting the model's robustness and generalization. Secondly, DepSepConv (Depthwise Separable Convolution) is employed in the neck part to replace the original C2f module, significantly reducing computational requirements and parameter count, while improving detection performance across various defective targets. Additionally, an enhanced Efficient_Detect network structure is introduced to further trim model parameters, computational complexity, and enhance resource utilization. Through experimental validation, the proposed method achieves a 96.5% mAP50, outperforming the original YOLOv8n model by 2.4%. The model's parameters, GFLOPs, and weight sizes are reduced by 39.4%, 44.4%, and 39.3%, respectively, with current parameters standing at 1.74 M, GFLOPs at 4.5, and weight sizes at 3.7 M. Furthermore, the detection speed reaches 141 FPS, indicating its potential for real-time detection. This research presents an effective solution for scenarios with limited computational resources and holds promise for application in the field of circuit board defect detection.

Installation of warning balls on transmission lines can prevent pedestrians and vehicles from approaching the wires and reduce safety hazards, while prolonged operation and the impact of weather can cause loose damage to the warning ball, so it is necessary to dismantle and replace the warning ball in time. In order to meet the demand for efficient installation of warning balls by robots instead of human beings, this paper researches the accurate recognition of warning balls and bolts, and proposes the improved YOLOv7 target detection algorithm to recognize warning balls and bolts. The improved YOLOv7 algorithm is formed by embedding the ECA channel attention module in the backbone module, adopting the feature fusion network of WBFPN, and choosing a more advanced loss function EIOU Loss. The experimental results show that the mAP value of the improved YOLOv7 algorithm reaches 94.7% and the FPS reaches 57.76f/s, which proves that the improved algorithm can satisfy the recognition requirements during the installation process.

Due to the difficulties such as the large proportion of small targets in UAV aerial images, an improved aerial image target detection algorithm is proposed in view of the low performance of the current mainstream algorithms in aerial image target detection tasks, and the high false detection rate and missing rate. First, the RFCBAM convolutional module is introduced into the YOLOv8s backbone network, which enhances the spatial awareness of the model without increasing the computing load. Secondly, the lightweight up-sampling operator CARAFE is introduced, which has a larger perceptual range and can fuse semantic information in a larger receptive field without introducing too many parameters and computing burden. Then, a P2 detection head was added to detect smaller targets, while a P5 detection head was removed for more efficient detection of aerial image targets. Finally, WIOUv3 is introduced as a new regression loss function to improve the learning ability of the model to cope with poor samples. Experiments show that compared with YOLOv8s, the improved model in this paper improves MAP50 by 6.5% and MAP50-90 by 4.4% in Visdrone2019 data set, and the number of parameters decreases by 28.1%.

Traffic sign recognition plays a key role in the field of assisted driving and autonomous driving, helping drivers and autonomous vehicles accurately capture important road information. At present, in the process of traffic sign recognition, the recognition of small traffic signs is still challenging. In order to further improve the performance of the small traffic sign recognition algorithm, the MND-YOLOv5 algorithm is proposed. Firstly, the value of the detection model is enhanced through the K-Means re-concentration anchor value. And replace the model's main network part to ODCONV to enhance network feature extraction capabilities; Secondly, for the neck of the model, the cross-layer cross-connection is added, and the BIFPN_Concat is introduced to reconstruct the Concat connection to supplement the semantic information of the feature map in space and enhance the detection of small targets; At last, NWD is introduced into the detection head to replace the CIOU_ loss function to solve the sensitivity of the IOU to small target position deviations. The experimental results on the dataset TT100K show that the accuracy P of the MND-YOLOv5 algorithm is 87%, the recall rate is 79.8%, and the average accuracy (mAP@0.5) is 85.3%, which is 4.7%, 7.7% and 5.9% higher than that of YOLOv5s, respectively, and has better detection performance.

In a variety of UAV application scenarios, such as agricultural monitoring, urban safety and traffic management, high-precision image object detection is crucial for real-time analysis and decision-making. Traditional object detection algorithms are often ineffective in dealing with small objects in complex environments. To this end, this paper proposes an improved YOLOv8 model that integrates a NAM (Normalization-based Attention Module) attention mechanism and a Bidirectional Feature Pyramid Network (BiFPN) specifically to optimize the object detection performance of UAV images. The NAM module enhances the model's spatial attention capability by utilizing a normalization process that This enables the model to adaptively emphasize important feature regions in the image, which significantly improves the recognition accuracy of small-scale objects without adding too much computational burden. BiFPN further optimizes the flow of information between different scales, which improves the efficiency of feature utilization and the overall performance of detection. Extensive experiments on the VisDrone2019 dataset show that the improved YOLOv8 model improves the mAP metrics by 11.3% compared to the original model, and performs especially well in scenes with complex backgrounds and the presence of multi-scale objects. This study not only validates the potential application of NAM in UAV vision tasks, but also provides a new technical path for the implementation of deep learning models in real monitoring systems.

At present, unmanned aerial vehicle (UAV) has widely used in tremendous tasks including aerial photography, plant protection, micro-selfies, express transportation, disaster rescue and wildlife observation. However, traditional planning and detection methods is basically precisely impossible in the aerial situation with unstable and complexity inputs. Additionally, existing mechanisms only consider to dispose one issue in a single model that means to utilize a machine to response the path planning and another model to achieve object detection, which may cause numerous system costs and pressure the loads of UAVs. Therefore, we utilize the deep neural network method to train dual neural networks and separately achieve the planning and detection for UAV, which can significantly reduce the system costs and enhance the effectiveness of working. From our comparison results with traditional machine learning algorithms, we can directly observe that our proposed method can increase the detection accuracy and provide multiple reasonable path planning strategies with acceptable system costs.

In order to effectively identify the damage characteristics of simply supported beam, a damage identification method for simply supported beam combining autocorrelation standard deviation vector and BP neural network is proposed. The damage simulation of simply supported beam is carried out through numerical simulation to obtain the strain response under different damage conditions, and then the autocorrelation of the strain response is processed to obtain the autocorrelation standard deviation vector damage index, and the autocorrelation standard deviation vector before and after the damage of the simply supported beam is used as the input parameter of the neural network to establish the damage identification model. Based on the fiber optic grating (FBG) sensing network technology, the strain response of simply supported beam is obtained, and the feasibility of the damage identification model is verified through experiments. The experimental results show that this method can accurately recognize the damage location and damage degree of simply supported beam.

With the development of agricultural intelligence, robots replacing manual labor in agricultural production has become an inevitable trend. For apple picking robots, accurate identification of apples is the primary step in their work. At present, the existing L*a*b and OHTA color spaces have problems of incomplete extraction of apple regions and incomplete differentiation of fruit and leaves. To address this issue, this study designed a color space (LPX color space) specifically for apple recognition, and based on this, proposed a complete algorithm for identifying green or red apples in a complex open environment. The experimental results show that the algorithm has high recognition accuracy in both cases of apples being shaded and unshaded, and can provide accurate orientation for the subsequent automated apple picking.

In the field of autonomous driving, object detection is the core foundation of autonomous driving environment perception, and it is also the premise for the safe operation of the whole system, and the accuracy of object detection in bad weather is also a hot research content in the entire industry. To this end, this paper proposes an improved algorithm for YOLOv5s in Blizzard scenarios. (1) The C3 module in the YOLOv5s backbone network is replaced with a C2f module with a richer gradient flow to improve the detection accuracy of small targets. (2)The attention mechanism SE was introduced to adjust the channel weight, which strengthened the extraction of important features of the vehicle. (3) The loss function C_IOU is replaced with E_IOU, which accelerates the convergence of the prediction frame and improves the regression accuracy. And the data set constructed in this paper is verified, and the results show that: Compared with YOLOv5s, the accuracy of the algorithm proposed in this paper is increased by 6.7% to 91.3%. The recall rate increased by 2% to 72.2%; The mAP@.5 value increased by 3.5% to 81.9%,The detection performance of the model in Blizzard scenarios is effectively improved.

The objective of this research is to develop an improved traffic sign recognition algorithm that can operate effectively in a variety of weather conditions. The existing algorithms have been found to have two main limitations: a slow detection speed and an insufficient detection accuracy. This study proposes an enhanced traffic sign recognition algorithm based on YOLOv8 as a solution to these issues. The experimental comparison revealed that YOLOv8 is the optimal base model for this purpose. The ADown module is employed to supplant the downsampling structure in YOLOv8, thereby optimising the YOLOv8 detection head framework with shared volumes. This approach results in a notable reduction in algorithmic parameters while maintaining accuracy, thereby enhancing the detection speed. The results of comparative experiments demonstrate that the improved algorithm exhibits an average accuracy mAP50 of up to 94.0% when utilising the same dataset. This value is higher than that of other algorithms, and the predicted accuracy is enhanced under varying weather and night conditions. Furthermore, the leakage rate is significantly reduced, aligning with the requirements of the road traffic sign recognition task.

Nowadays, face recognition technology has widely penetrated into various industries, however, there are still multiple factors that challenge the recognition effectiveness of the model. To this end, this paper innovatively introduces a novel network architecture that incorporates convolutional neural networks and Inception modules. In this architecture, we use a 1×1 convolutional layer for dimension boosting operation, while multiple 3×3 convolutional units as well as the dilated convolutions technique are employed in parallel to fully capture the feature information of face images under different receptive fields. Considering the intrinsic similarity between face images, a joint loss function combining the Softmax loss function and the Center loss function is specially designed in this study, which is used as a benchmark for evaluating the model during the backpropagation process. After a series of rigorous experimental validation, this improved scheme shows significant effect enhancement and confirms its superiority in the field of face recognition.

The objective of student classroom behavior recognition is to employ artificial intelligence technology for locating and identifying student behaviors within the classroom. Publicly available datasets in this domain are relatively scarce due to concerns about student privacy. Furthermore, current research lacks a clear standard for defining student classroom behavior and encounters limitations in modeling its influencing factors. To tackle these issues, this paper constructs a dataset named SCBD (Student Classroom Behavior Dataset) and proposes a two-stage relationship modeling network (TSRM-Net). TSRM-Net employs the YOLOv7 algorithm to locate student positions. In the first stage of relationship modeling, it utilizes an adaptive 3D convolutional neural network to automatically learn the direct relationship between students and the classroom environment. In the second stage of relationship modeling, it employs a visual attention mechanism to extract interaction information among students. Finally, behavior classification is conducted with a multilayer perceptron network that utilizes the fused features obtained from the two-stage relationship modeling. The proposed approach outperformed traditional methods, achieving the most favorable experimental results on the SCBD dataset.

6DoF (6 Degrees of Freedom, or 6DoF or 6D) pose estimation has practical applications in robot vision, 3D scene understanding and other fields. 6D pose estimation based on RGB-D images has an important role in robotic applications, however most current methods directly stitch together RGB and depth information, resulting in low accuracy of pose estimation. This paper presents a novel object 6 DoF pose estimation method based on RGB-D images, which improves on the insufficient accuracy of pose estimation. In this paper, we transform the depth map into point clouds, design an asymmetric geometric feature extraction network AGF to effectively extract the geometric information in point clouds, and combine with EfficentNet-B7 network to extract RGB features. In order to make full use of the cross-mode information, the primary and auxiliary feature interaction module MAF is proposed, which can realize the efficient fusion of RGB and point cloud features. Moreover, this paper introduces an adaptive feature contribution weight calculation module AWF, which is able to adaptively calculate the different contribution weights of the two modes to the pose estimation task, thus further improving the estimation accuracy. The experimental results show that the proposed method shows higher accuracy and robustness in the object 6 DoF pose estimation task.

This article presents a novel Faster RCNN-based approach for detecting transmission line targets, aiming to overcome limitations of traditional methods. When it comes to feature extraction, it employs ResNet as its backbone, harnessing the prowess of its intricate convolutional layers and ingenious residual connections to capture intricate and diverse features. To accommodate targets of varying scales, the Feature Pyramid Network (FPN) is integrated, effectively blending shallow and deep features. Thorough experiments have demonstrated the model's efficacy, resulting in substantial gains in detection accuracy and recall, thereby providing fresh insights into intelligent recognition of visible light inspection images captured by power transmission drones.

In recent times, due to the fast progression of artificial intelligence technology, the recognition of emotions has emerged as a trending area of study. Among many emotion recognition technologies, the method based on expression recognition is particularly critical. We propose a two-stream model that can simultaneously extract static and dynamic expression features. To verify the information representation ability of the two-stream model for expression images, we conducted emotion classification experiments based on a deep convolutional neural network model on two public dynamic expression databases (CK+ and Oulu-CASIA). Experimental results show that our two-stream model has significant advantages in classification performance compared to single static or dynamic expression classification methods, as well as classification techniques based on 3D convolutional neural networks.

As artificial intelligence technology continues to advance, the utilization of deep learning in sonar image target detection is growing. This trend has garnered significant interest in maritime defense infrastructure and marine economic development as a promising research avenue. Nevertheless, existing sonar image target detection methods face challenges due to the inherent characteristics of sonar images, such as low resolution, blurred edges, and significant speckle noise. These issues often result in large model parameters and slow inference speeds. To mitigate these challenges, this study introduces a novel approach: a sonar image target detection model utilizing a Fast Large Kernel Attention Network (FLKANet). The FLKANet enhances feature extraction capabilities, reduces model parameters, and improves inference speed. Moreover, employing Large Separable Kernel Attention (LSKA) instead of Small Patch Feature Fusion (SPFF) enables the model to prioritize the shape of sonar targets over texture, thereby enhancing detection accuracy and inference speed. Experimental results on the SCTD sonar image dataset demonstrate the superiority of the proposed model in terms of parameter count, computational complexity, inference speed, and detection accuracy. These findings offer valuable insights for deploying edge-side models in future applications.

As an important hardware component in the process of EEG recognition, the brain information perception system undertakes the functions of accurate EEG signal acquisition, amplification, conversion, transmission and recognition. As the EEG signal shows the characteristics of weak in amplitude, low SNR, low frequency range and strong randomness, it can be easily affected by electromagnetic interference in the environment. Thus leads to a decrease in accuracy. Therefore, the interference sources that interfere with the brain information perception process are sorted out, the simulation is carried out for different objective factors, in order to explore the influence of electromagnetic interference in the environment on the brain information perception system. Finally, we provide reference and ideas for the interference of the brain information perception system.

Due to the challenges associated with traditional methods in reconstructing complex water images, such as low resolution, absence of key information, and significant noise, this paper presents a network integration-based algorithm for low noise super-resolution reconstruction. In order to make the reconstructed image texture clear, the implicit neural expression of the image is applied in the traditional SRGAN algorithm. We also utilize the concept of network integration to effectively capture both the surface-level and in-depth information from the image. Experimental results indicate that the algorithm we propose outperforms the current mainstream algorithms in terms of both subjective visual effects and objective quality evaluation indicators for reconstructed images.

To settle the problem of parameter mismatch under the model predictive control, a flux-linkage identification scheme is proposed for permanent magnet synchronous motor (PMSM) driven by model prediction current control (MPCC) in this paper. The electrical parameters of the PMSM will change under the influence of the external environment and temperature. The high dependence of MPCC on the model parameters will produce the current prediction error, in which the mismatch between flux-linkage and inductance has a great impact on the control effect. This paper proposes a flux-linkage identification method based on the error of the predicted current and the actual measured current, from which we extract the information containing the flux-linkage error and online compensates the estimated value of the flux-linkage in the prediction model. It largely decoupled the flux-linkage from inductance parameters and is insensitive to inductance changes. The theoretical proof and simulation results show that this method has high identification accuracy and can accurately identify the flux-linkage when the inductance parameters change.

This article studies a high-precision intelligent detection device for the inner distance of standard wheelsets in rail transit. Different measuring anvils are used to effectively avoid the impact of threaded splicing on the straightness and repeatability of the measuring rods. High precision grating sensors are used, with a resolution of 0.0001mm and an error of ± 0.018mm. Storage and data calling functions are installed, which effectively improve the accuracy, stability, and real-time recording of wheelset detection; The structure and measurement principle of the standard wheelset inner distance ruler were introduced, and the uncertainty was evaluated. The results showed that the detection device is accurate and feasible for measuring the inner distance of the standard wheelset.

Conventional fast location algorithm for small target in power operation scene tree obstacle mainly uses ECKF (Embedded Cube Kalman Filter) embedded volume Kalman filter to create a random location function, which is vulnerable to changes in random parameters, resulting in poor location performance. Therefore, a fast location algorithm for small target in power operation scene tree obstacle based on self supervised visual transformer is proposed. That is to say, the self supervised vision transformer is used to design a fast positioning framework for small targets in the power operation scene tree obstacle, thus generating a fast positioning optimization algorithm for small targets in the power operation scene tree obstacle. The experimental results show that the designed fast location algorithm for small target with tree obstacle has high location accuracy, low miss detection rate and false detection rate, which proves that the designed fast location method for small target with tree obstacle in power operation scenarios has good location performance, effectiveness and certain application value, and has made certain contributions to reducing the risk of power operation.

In response to the low efficiency, high error rates, and reliance on the experience of detection personnel in the current manual visual grading method for beef, a machine vision-based beef grading system has been designed. Furthermore, in-depth research has been conducted on its image processing algorithms. The system focuses on the eye muscle of beef as the research subject. Images of the beef are collected using an industrial camera. Subsequently, the visual software OpenCV is employed for image analysis and processing. Utilizing an improved Canny edge detection algorithm, the system detects the edge contour information of the beef eye muscle. It employs image segmentation algorithms to extract relevant information from the beef eye muscle images. Ultimately, based on the extracted information, the system grades the beef and visualizes the grading results through a graphical user interface (GUI). Compared to traditional manual inspection, this beef grading system is stable and reliable. It not only enhances the efficiency and accuracy of grading but also reduces labor costs. It possesses vast application prospects and economic benefits.

Due to the characteristics of multi-scale and complex edge shape of X-ray film weld defects, the segmentation accuracy of conventional segmentation methods is low. Aiming at the characteristics of these defects, this paper proposes an X-ray weld defect segmentation method based on deep separable convolution structure. Firstly, a multi-channel information fusion module is innovatively designed to replace the two consecutive 3×3 convolutions. Secondly, a spatial depth separation attention mechanism is added after the skip connection. Finally, a tail feature information cascade module is proposed to fuse different levels of feature information. This method can show high performance when segmenting X-ray film weld defects with multi-scale and complex edge shapes. The experimental results on the public dataset GDX-ray show that the proposed method is 6.1 % and 6.4 % higher than the classical Unet method in DSC value and precision P respectively, which effectively improves the segmentation accuracy of X-ray weld defects.

To ensure product quality, it is necessary to detect the appearance defects of products to ensure that high-quality products enter the market. Machine deep learning technology is used to extract product defect characteristics, and parameters are adjusted through small sample detection to realize the identification training of the detection system. AI algorithm is used to conduct real-time defect positioning on the production line, which can meet the requirements of product defect detection, to control the production line for correct distribution. In the production process, with the increase of defect samples, the defect samples produced in the production are directly annotated and retrained, which can gradually improve the detection accuracy of the network model and enhance the generalization performance of the network model. After the system design is completed, the production line is tested to achieve a high detection accuracy. In addition, when detecting two similar targets, we only need to change the data set and training parameters and use the transfer learning method to load the previously trained weights into the new model for training, to achieve good results.

A coordinated control strategy integrating Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC) has been designed to enhance the trajectory tracking accuracy and stability of autonomous vehicles. This strategy consists of upper and lower-level controllers, with the upper level responsible for tracking desired trajectories and vehicle speed while minimizing lateral error. The lower-level control executes actions to achieve desired steering angles and accelerations. An Extended Kalman Filter (EKF) observer is employed to update the vehicle state. Finally, hardware-in-the-loop testing experiments are conducted to validate the robustness and effectiveness of the controller, which effectively reduces lateral error and improves tracking accuracy.

Camera calibration is a critical process in computer vision applications, ensuring accurate mapping between image coordinates and real-world coordinates. We propose a method to fine-tune the calibration parameters of cameras by using machine learning technology, aiming to improve the calibration accuracy and stability. The proposed method involves data collection, data preprocessing, dataset division, model selection, model training, model evaluation, parameter refinement, and iterative optimization. Real-world data is collected and pre-processed to ensure data quality and consistency. The dataset is then divided into training and test sets for model training and performance evaluation. A machine learning model using linear regression to model the relationship between camera parameters and calibration errors. The model is trained and evaluated, and the grid search is used for cross-validation to adjust the parameters of the model to make the model better. Based on the model completed by parameter tuning, new camera parameters are input, fine-tuned by genetic algorithm, and the average error is calculated until the set error is satisfied, and the adjusted parameters are output. Finally, the HALCON measurement assistant is used to verify the output results. The results show that although the calibration data obtained by HALCON is not excellent, the calibration accuracy and stability can be improved by using the method in this paper. The research results can be used to optimize the calibration parameters of tilt-shift lenses, which is very helpful for 2D data measurement and 3D reconstruction.

This paper mainly studies the optical fiber distribution system based on machine vision, and discusses its advantages and limitations in practical application. The development status and problems of optical fiber distribution system are introduced, and the application prospect of machine vision technology in this field is discussed. The design of optical fiber distribution system based on machine vision is introduced in detail, including system structure, hardware equipment selection, software system design and algorithm implementation. In the aspect of system performance test and evaluation, the test environment and data are introduced in detail, and the experimental results are analyzed and the system performance is evaluated. Through this research, it can provide reference and guidance for the practical application of optical fiber distribution system based on machine vision.

Detection of actual pipeline cracks is an important task of road administration. Aiming at the practical difficulties of the current single means of pipeline defect inspection, a pipeline crack defect intelligent robot inspection system based on Yolov8 visual detection model is designed. The pipeline robots are mainly composed of underlying motion control, high-performance computing, and automatic recharging in terms of hardware. The equipped camera acquires real-time camera image data from the pipeline, sends it to the Yolov8 node through the robot operating system (ROS) topic, and outputs the crack location information after GPU calculation. The on-board LiDAR acquires the surrounding environment information, builds a map of the surrounding environment using the GMAPPING algorithm, and then applies the TEB path planning algorithm to realize the robot's automatic pipeline inspection navigation based on, Simultaneous localization and mapping (SLAM). This study also uses ROSJAVA to design an Android host computer system for controlling, mapping, and navigating the pipeline robot.

Liquid metal ladles serve as crucial equipment for containing high-temperature molten steel. Comprising a metal shell enveloping an inner layer of refractory material, these ladles are susceptible to cracks, thinning, or peeling of the refractory bricks, which can result in the hazardous leakage of high-temperature liquid metal, thereby jeopardizing employee safety. In this research, we propose an innovative continuous non destructive testing and monitoring system that harnesses the power of infrared thermography to assess the external temperature field of ladles. By deploying multiple infrared thermal cameras, we establish an aerial virtual infrared image inspection gate, wherein each passage of a ladle through this gate triggers a meticulous analysis of the temperature field. This analysis aims to identify any irregularities or aberrations and subsequently compares them with historical records to discern prevailing trends. Our system has been successfully implemented and employed for the diligent tracking and monitoring of ladles within steel plants, effectively mitigating the necessity for labor-intensive manual inspections and enabling streamlined unmanned operations.

As an important expression of holographic video, point cloud video is widely used in many fields. However, extremely high bandwidth demand and unbearable transmission delay are the biggest obstacles when streaming, and the complexity of its encoding and decoding cannot be ignored. This article addresses the aforementioned difficulties by using super-resolution to reduce the dependence on network bandwidth during video transmission, and incorporates it into third-party platforms to reduce transmission latency. Specifically, we only store video tiles of basic resolution in the server, with the goal of selecting the most suitable resolution for each tile within the perspective within the platform's computing power limitations. We transform it into an optimization problem and solve it through heuristic algorithms. The simulation results show that this scheme is more effective than traditional schemes.

With the acceleration of urbanization, urban demolition, construction, reconstruction and other engineering projects to promote urban waste transportation activities increasingly frequent, but because of overload, overrun, dangerous driving and other non-standard operation problems, endangering social and traffic safety. Traditional police law enforcement has problems such as management and governance leakage, incomplete coverage, and weak law enforcement. In the application of the existing detection methods based on traditional loadometer weighing and road weighing, there are pain points such as high equipment investment, small detection range, and easy damage of equipment. In order to do a better job of muck vehicle management, this paper breaks the traditional muck vehicle management and loadometer weight measurement method, and creatively puts forward no gap dynamic real-time supervision and machine vision based on deep learning. Through the accurate judgment of overload and limit of muck vehicle morphological detection, it effectively solves the problem of no blind spot, continuous and less investment in law enforcement. It provides an effective management scheme for the accurate law enforcement and violation remediation of traffic control departments and administrative departments.

Multi-organ segmentation from abdominal images is an important task. Due to the imbalance between different organ and the differences in size, shape, and contrast of different organs, it is a challenging problem in the field of medical image analysis. A powerful feature extraction model is the key to solving this challenge. TransUNet is a popular solution. MultiLayer Perceptron (MLP) is its important part. However, the existing MLP modules only operate on a single feature dimension without interaction across different dimensions. This paper proposes an improved MLP module with multidimensional interaction, named SMLP-Mixer. The SMLP-Mixer achieves multidimensional information interaction while possessing lower computational complexity and fewer parameters in the feature dimension, which can reduce the risk of overfitting. Experiments on the Synapse dataset demonstrate its effectiveness. Compared to TransUNet, the Dice coefficient exhibits a 1.98% improvement, and the HD95 (mm) shows a significant enhancement of 10.83.

The current process of marking examination papers often relies on manual correction and scoring. This mechanical fixed-answer correction and scoring process not only consumes a lot of time and energy of the markers but is also prone to errors. To address this problem, this paper proposes a fully automated framework for scanning, correcting, and marking question papers. There are three main parts in this framework. The first part is a hardware system device that can scan the paper and turn pages automatically. The second part is a handwritten character recognition algorithm, based on the Pytorch framework, which can segment and identify the different categories of characters in the test paper image. The third part is the answer checking and marking system, which calculates the total marks of the question paper through the score summation model. The proposed automatic marking model can significantly improve the efficiency of test paper correction. The experimental results demonstrate that the proposed framework achieves high character recognition accuracy and can quickly and efficiently count the test paper marks automatically.

In order to provide navigation services in environments with complex buildings or poor satellite signals, a navigation scheme that uses Ultra Wide-Band (UWB) wireless communication technology to obtain people's positions and provides the best path with a path search algorithm is proposed. The scheme is based on the UWB development board, and extends the two-dimensional positioning algorithm by measuring the distance between the base station and the tag to achieve three-dimensional positioning; and adopts optimization and extension based on the traditional path search algorithm to solve the problems of not being a connected graph in the same floor and the existence of multiple entrances to the destination. The UWB positioning algorithm, path search algorithm and 3D map are integrated to realize the 3D navigation system in the building based on UWB positioning.

This study introduces a novel method for bearing fault diagnosis, merging one-dimensional convolutional neural networks (1D-CNN) with the XGBoost algorithm. We conducted detailed experiments using the Case Western Reserve University's bearing dataset. Our approach combines 1D-CNN's local pattern analysis in time series data with XGBoost's classification prowess, resulting in substantial performance improvements over existing diagnostic techniques. Notably, our method excels in recall rate, emphasizing its significance in enhancing the efficiency and reliability of bearing fault diagnosis.

In order to ensure the safe lane changing of self-driving vehicles and at the same time improve the efficiency of vehicle traffic, a game theory-based vehicle lane changing decision model is proposed. Firstly, the interaction behavior and game relationship between vehicles are analyzed. Then the safety gain, comfort gain and speed gain are established as the game gain function, and the vehicle driving conditions in the NGSIM dataset are clustered and analyzed using Gaussian mixture model (GMM), which divides the driver styles into three categories. The anthropomorphism of the vehicle lane change decision model is improved by considering the effects of different driving styles on the vehicle lane change decision. Finally, SUMO software is used to build a hybrid scenario of human-driven vehicles and self-driven vehicles to compare and analyze with the MOBIL lane change model. The results show that the lane-changing game decision-making model in this paper has higher safety and traffic efficiency.

Super-resolution image reconstruction refers to the estimation of high-resolution image from low-resolution input image. The reconstruction effect depends heavily on the relevance between the external training set and the test image. In this paper, we propose a novel fusion SR method, which makes use of the self-similar property of internal samples, so that the ill-condition external training set can also reconstruct the image robustly. Specifically, we define relevance quantitative index and calculate relevance scores for parallel feature channels. Based on the principle of consistency maximization of low correlation regions, we reasonably synthesize the results of parallel scores, classify image patches, and determine the fusion proportion of internal and external samples adaptively. Experimental results show that the proposed algorithm can effectively alleviate the visual defects caused by ill-conditioned training set. It is worth noting that under the condition of insufficient training samples, our method perform better than the reconstruction method based on deep learning.

Object detection plays a crucial role in intelligent transportation systems. However, the presence of numerous uncertainties in object detection within real traffic scenes poses significant challenges to achieving accurate and real-time detection. To address these challenges, this paper introduces a thermal radiation traffic detection algorithm based on an enhanced version of YOLOv8. Specifically, a novel DC_C2F module is proposed, which not only enlarges the effective receptive field of the convolution kernel but also reduces computational complexity. Additionally, the introduction of the MBConv module enhances the feature extraction and detection capabilities of the YOLOv8 model. Experimental results demonstrate that the proposed method surpasses several other prominent algorithms when applied to challenging thermal infrared radiation datasets, showcasing its robust object detection capabilities.

The large-scale reconnaissance task allocation problem of multiple unmanned aerial vehicles (UAVs) is researched. In order to simplify the problem, the task allocation process is divided into two independent stages: the cluster stage and the reconnaissance sequence allocation stage. In the first stage, a k-means cluster algorithm based on the azimuth angle is developed to group a large number of targets. The traditional wolf pack algorithm (WPA) does not fit the task allocation problem. In the second stage, an improved WPA is developed to decide the reconnaissance sequence of every group of targets. The improved WPA mainly modifies three behaviors in the traditional WPA, namely, wandering behavior, summoning behavior, and besieging behavior. Two simulation experiments are carried out. In the first experiment the algorithm in this paper is applied. In the second experiment the traditional k-means cluster based on the distance is applied to group the targets. The comparison of experiment results validates that the algorithm in this paper is more effective and outstanding.

Underground cable ducts are widely used to protect and place wires and cables, but the environment where they are located is narrow and occluded, which increases the difficulty of manual inspection and makes maintenance work tedious. In response to this issue, this paper designs an unmanned inspection system for underground cable ducts based on machine vision and Internet of Things (IoT) technology. Firstly, the locally trained YOLOv2 target detection model is deployed in Kendryte K210 embedded platform; then based on the temperature and humidity sensor and WiFi module, the temperature and humidity data are uploaded to the AliCloud IoT platform in real time through the MQTT protocol, and real-time monitoring the internal condition of the pipeline. Test results show that the inspection system can effectively complete the unmanned inspection task of cable pipelines, simplify the inspection process, and provide convenience for the staff.

In autonomous driving systems, trajectory prediction is crucial for enhancing roadway safety and diminishing the likelihood of accidents. However, over time, the evolution of the trajectory becomes more and more uncertain and unpredictable, increasing the complexity of the problem. To address these problems, this paper innovatively proposes a trajectory prediction framework named DistTF, which skillfully integrates Transformer model and knowledge distillation. DistTF uses Transformer model to characterize vehicle-to-vehicle interactions and mine deep timing patterns to train a high-performance teacher model. Subsequently, through knowledge distillation, the knowledge of the teacher model is effectively transferred to the smaller student model, so as to maintain considerable prediction accuracy under the premise of ensuring fast reasoning and low parameter quantity. Experimental results on the NGSIM dataset show that DistTF achieves the highest accuracy among the four algorithms participating in the comparison. Comprehensive comparative experiments verify the effectiveness of the proposed method.

Taking the electronic speed control system of single cylinder diesel engine for generator set as the research object, with the increasing complexity of the control system and the improvement of the speed control performance, the control effect of traditional PID control in the speed control system is slightly insufficient. Therefore, based on the CPO optimization algorithm to optimize the traditional PID control, this design proposes a CPO-PID control algorithm to optimize the parameters of the speed control system. In MATLAB/ Simulink environment, the process of starting and suddenly adding and unloading step disturbance of the speed control system is simulated, and the speed control effect under different control algorithms is observed. The results show that, compared with traditional PID control, the CPO-PID control system can reach the expected value quickly and stably, and has small overshoot and strong anti-interference ability. The speed control performance has been significantly improved.

In order to improve the segmentation accuracy of 3D point cloud model in feature ambiguous region, the unsupervised clustering algorithm based on surface fusion features combines the depth residuals with the normal deviation angle to form the fusion features with good descriptive properties and significant differentiation. Adopting the point cloud local density median proximity points as the candidate points for point cloud region growth, using point cloud cluster features instead of individual point features, for the problem of feature ambiguity in the edge region, the unattributed points are reclustered, so as to improve the cluster segmentation in the way of extended growth; experiments are conducted on the selected model dataset, and the results show that the method can achieve the reasonable segmentation of the complex point cloud model, and the under-segmentation situation has a obvious improvement on the under-segmentation situation.

As the core component of the power system, the normal operation of the substation is very important to the reliability of the power supply. As one of the key elements to connect substation equipment, the maintenance of the bolt fastening state has an important impact on the safety and stability of equipment. The purpose of this study is to propose an algorithm based on deep learning, which is used to detect the fastening state of substation bolts in real time to ensure the normal operation and maintenance of equipment. In this study, we collected a large number of substation bolt state data and designed a deep learning model. Firstly, we preprocessed the data with Gramian Angular Field (GAF) and Piecewise Aggregate Approximation (PAA). Then, we put forward a diagnosis model of the acoustic emission signal of substation bolts based on two-dimensional time series and Attention Mechanism (AM), and then use this model to capture the time series information of bolt state. The experimental results show that after 100 iterations, the diagnostic accuracy of the model has reached the highest, with an average accuracy of about 98%. Our deep learning algorithm can highly accurately identify the fastening state of bolts, which provides an effective guarantee for the safety and reliability of substations, and also has the ability to monitor in real-time.

Efficient container transportation management relies heavily on accurate and timely detection of container box numbers. Recognizing the limitations of current manual recording and traditional detection methods, we present an enhanced YOLOv5-based algorithm. This algorithm incorporates a spatial channel attention mechanism to bolster target feature extraction and undergoes iterative training for refinement. Experimental results showcase a remarkable average detection accuracy of 98.5% and a swift detection frame rate of 54 FPS. In comparison to conventional methods, our approach significantly improves detection performance, enabling real-time capabilities suitable for diverse operational environments in container transportation management.

To meet the demand for higher detection speed and detection accuracy in the current leather defect detection program. This paper proposes the GEI (GhostNetV2-EMA-InnerIoU-YOLOv9) leather defect detection model to solve the problems of lacking defection samples and improving detection speed and accuracy. Specifically, GhostNetV2 is used in the backbone of the YOLOv9 algorithm instead of the existing convolutional method. Then, the proposed model introduces the EMA attention module to improve the small target feature extraction capability. Finally, this paper uses Inner-IOU instead of CIOU to achieve faster and more efficient regression results. The experimental results show that the method proposed in this paper reduces the parameters by 9%, increases the FPS by 2.5% and achieves 94.7% mAP compared to YOLOv9.

With the rapid development of 3D modeling and immersive simulation technologies, the accuracy of indoor environment reconstruction is receiving increasing attention. This paper proposes an indoor scene reconstruction method that preserves surface discontinuities. The method utilizes the prior of surface discontinuities to construct a step function taking values of 0 and 1 and employs the smooth approximation of this function as weights to establish an adaptive weighted least squares fitting model for normals. Given the non-convexity of the model, we present an iterative scheme that uses the depth generated from the previous iteration (with the initial depth set to zero) to fix the weights, converting it into a convex quadratic functional for solution. Experimental results demonstrate that our method can recover indoor scenes more accurately than existing technologies while preserving surface discontinuities, particularly performing well in handling large planar sections of indoor environments.

Cylinder liner is one of the important parts of engine. The quality of cylinder liner directly affects the safety and energy consumption of engine. Most of the dimensional inspection of cylinder liner has been automated, but the surface defect inspection is still carried out by a large number of personnel. In recent years, researchers began to use image detection technology for surface defect detection, but the detection effect is still not widely used in industry. In this paper, it studies the burr cylinder liner, which is also a kind of difficult in the surface inspection of cylinder liner. The surface defects of the cylinder liner, such as pit, slag inclusion, residue, coarse burr, sparse burr. This paper mainly studies the surface defect detection of cylinder liner based on 3D vision. Experiments show that this method successfully realizes the automatic surface detection of cylinder liner, and is applied in the industrial field.

Wheat detection is significant in agriculture and can help farmers achieve efficient planting management and precision agriculture. Traditional wheat detection methods usually rely on hand-designed feature extractors and classifiers, but the ability and generalization of feature representations limit their performance. In contrast, deep learning-based methods can automatically learn feature representations and perform end-to-end training on large-scale datasets, with more vital representation ability and generalization. In this study, we propose a PVT-based detection method for wheat. PVT is a visual model combining a Transformer and feature pyramid, simultaneously capturing global and local information in images. The experimental results show that the PVT model is adaptable and can be generalized in wheat detection. It can provide accurate detection results of wheat ears and maintain stability when dealing with different light conditions and wheat growth cycles.

Eliminating defects of missing strips and mislabeling is an important task for cigarette enterprises to ensure product quality. There are various shortcomings in the defect detection methods used in the past, and machine vision technology as a new detection means has been widely used in various production processes in the tobacco industry. In order to suppress the noise interference and better detect the edges in the cigarette strip image, this paper designs a set of image acquisition and analysis system for the detection of defects in cigarette boxes based on the Canny edge detection algorithm, which achieves the purpose of counting by matching the contour with the region of interest, and achieves a good effect of defect detection.

Steel profiles is widely used in many different fields and is a fundamental raw material in the industrial production process. However, surface defects often occur during the steel production process due to differences in testing environments and production raw materials. Therefore, it is crucial to detect surface defects in steel profiles. However, with the progress of society, science and technology, traditional machine vision inspection methods are increasingly difficult to meet the requirements of modern industrial production in terms of accuracy and real-time performance in the actual inspection process. In this context, this thesis provides an in-depth analysis of deep learning-based defect detection algorithms and improves the state-of-the-art YOLOv5 object detection algorithm according to the characteristics of steel profiles surface defects in order to improve the detection speed and accuracy. First, the steel dataset is regrouped using the K-means algorithm based on IOU measured distances to generate different sets of anchor frames. Second, MixUp was integrated into the Mosaic data extension to reduce overfitting and improve model generalization. Then, the network structure was improved and an attention module was added to further improve the ability to extract network features.

For the vehicle edge computing platform, a huge model is difficult to achieve the requirements of real-time detection, which faces the challenges of high computational load and bad detection rate. In this paper, an improved YOLOv8n detection method is proposed, and a dynamic deformable convolution block (DDCNv2) is proposed in the YOLOv8n backbone network to improve the detection accuracy of the algorithm. A dynamic convolution module (KWConv) is introduced in the YOLOv8n. In addition, a multi-scale detector module is designed to reduce the number of parameters. We use the PASCAL VOC dataset and the MS COCO dataset. The results show that compared with the existing YOLOv8n, the detection accuracy of the proposed model is improved by 5.1%, the FLOPs is reduced by 69.51%, and the model parameters is reduced by 8.64%, which proves the effectiveness and superiority of the proposed method.

This paper proposes a frame-sorting module-FrameSORT and fuses it with YOLOv5 to create a novel visual detection and localization method, YOLO-FS, which can effectively and accurately detect table tennis ball moving targets, aiming at the omission and misdetection problems of YOLOv5 target detection in continuous image sequences as well as the poor robustness of traditional methods in detection. The Frame-SORT module uses the frame difference method and extended Kalman filtering for target detection and localization of frame images that are not identified by YOLO. The experimental results show that compared with other commonly used table tennis ball detection methods, this method achieves 94.6% detection accuracy in the task of video sequences, maintaining YOLO real-time detection while improving robustness, and has better detection accuracy and application prospects.

The task of spatiotemporal action detection plays a pivotal role in various domains such as video surveillance, medical fields, and sports analytics, necessitating real-time accuracy. One of the key challenges in spatiotemporal action detection lies in extracting temporal features. While many methods employ an additional branch for temporal feature extraction, effectively integrating the information obtained from these two branches remains a significant challenge. In this article, we propose a new framework for real-time action detection. P-YOWO, which conducts spatial and spatiotemporal modeling through two distinct branches to achieve action classification and bounding box prediction. In this endeavor, we improve the spatial modeling ability of the network for the current frame by improving the 2DCNN branch of the YOWO network structure. However, YOWO falls short in effectively leveraging spatial feature information. To address this limitation, we introduce a spatial attention mechanism in the fusion of branch information. This mechanism applies spatial self-attention to bolster the integration of spatial and temporal features in the spatial dimension. Through experiments conducted on the UCF101-24 dataset, our model exhibits an improvement of 4.8% in both frame-level and video-level average precision compared to the original model. The experimental results validate that our enhanced approach effectively explores spatial semantic information in feature maps, thereby augmenting the detection performance of the model.

Electroplated diamond wire saw is widely used in photovoltaic, wafer, and other industries in the field of slicing processing, but in the sawing process the twist of the wire saw will have an impact on the quality of the slicing process and the life of the wire saw, to detect the twist of the electroplated diamond wire saw, the electroplated diamond wire saw twist detection method based on machine vision is proposed. A high frame rate CCD camera is used to acquire the surface image of the electroplated diamond wire saw, and the half-cylindrical image of the wire saw is expanded to mention the coordinates of its feature points, and the twist angle of the wire saw is calculated by the change of the position of the same feature point between the neighboring frames of the image. The results show that the maximum relative deviation of the twist angle of the electroplated diamond wire saw based on machine vision is 3.8%, and the relative deviation of the average value is 1.8%, which is capable of realizing high-precision, high-efficiency, and low-cost detection of the twist angle of the electroplated diamond wire saw, and providing a detection basis for controlling the twist of the electroplated diamond wire saw.

Existing target detection algorithms perform well in common road scenarios, however, in real complex road environments, where a large number of targets and small targets are prevalent, it is still a challenging task for pavement defect detection methods to achieve the ideal state. For this reason, this paper proposes a pavement defect detection network based on multi-scale image fusion. The network uses multiple detection heads and an almost parameter-free TripletAttention module to reach the interaction between different dimensions. To cope with the high-density target detection problem, we propose a YOLOv7 network structure combined with LSK spatial attention module and introduce a new multi-scale selective fusion module, MSF, to extract image features at different scales. Finally, this paper incorporates the activation function Mish() to enhance the model's ability to recognize overlapping and occluded objects. Experiments conducted on the dataset UDTIRI show that the improved method in this paper improves the overall AP by 4.5% and mIoU by 4.6% relative to YOLOv7 target detection algorithm for baseline networks. The ablation experiments and visualization performance demonstrate the rationality and superiority of this paper's method.

In view of the missing detection and false detection of the current pedestrian detection model in the dense pedestrian scene, a dense pedestrian detection algorithm based on Yolov7-tiny was proposed. In order to solve the problem of missing pedestrian target features in dense pedestrian detection scenarios, the dynamic convolution module and the improved ROELAN module are used in the backbone network to make the network focus on the important features of dense pedestrians through dynamic convolution, which effectively reduces the impact of feature loss on model detection. In order to solve the problem that small target pedestrians are easy to be missed in dense pedestrian detection scenarios, the BIFPN idea, CARAFE upsampling operator and the improved Rep-ELAN-B module are introduced to improve the detection performance of the model by using the small target feature information in the medium and low-dimensional feature maps. The average detection accuracy of the improved model reached 83.4%, which was 2.7% higher than that of the baseline model. The experimental results show that the proposed algorithm can be better applied to dense pedestrian detection scenarios.

Foreign object intrusion on highways is a significant cause of major traffic accidents. However, existing foreign object detection models, which prioritize detection accuracy, commonly suffer from issues such as model complexity, a large number of network parameters, and poor real-time performance. Moreover, these models are typically deployed on server-side platforms and are not suitable for the vehicle-road cooperative foreign object detection domain, which requires wide coverage and high real-time requirements. To address these challenges, this paper proposes a fusion adaptive weighted vehicle-road cooperative foreign object detection model. The model employs depth separable convolution to construct the backbone network, reducing model complexity and parameter count to enhance detection real-time performance. Attention mechanisms are introduced to capture features of local target regions, thereby improving feature extraction capability. By adaptively weighting and fusing vehicle and roadside smart lamppost detection information, detection accuracy is further enhanced. Experimental results demonstrate that the proposed foreign object detection model achieves a computational load of 7.70 GFLOPS, parameter count of 11.72 million, frame rate of 38.56 fps, and detection accuracy of 79.88%. Compared to the YOLOv4 model, the proposed model reduces parameter count by 81.8%, computational load by 87.3%, and improves real-time performance by 31.9%. As a result, it can effectively be applied in the vehicle-road cooperative domain.

Distributed acoustic sensing systems can obtain information about road vibration caused by vehicle driving vibration on highways. By characterizing the vehicle driving vibration data, a highway vehicle driving vibration detection scheme based on ridge-LASSO regression is proposed. Firstly, using the random forest's out-of-bag error criterion for vibration signal multi-feature selection, to prove the validity of the feature quantity used to characterize the highway vibration signal, and then the multi-feature quantity values of multiple measurement areas are composed into a data matrix, and then finally the ridge-LASSO regression algorithm is used to detect the vehicles travelling on the highway to obtain the vehicle's trajectory. The experiment proves that the method can effectively detect vehicles on the highway

In the current facility management and maintenance industry, utilizing Unmanned Aerial Vehicles (UAVs) for crack detection holds tremendous potential. However, cracks captured at high altitudes occupy a minimal proportion in images, posing a significant challenge for detection models to effectively extract features. This paper focuses on concrete crack identification and proposes an improved YOLOv5-R network based on the YOLOv5 model and RepNCSPELAN (Re-parameter, CSPNet, ELAN) structure, aiming to enhance the detection accuracy of surface cracks on concrete structures in high-altitude environments. The model primarily optimizes the neck section, where results from multiple convolution modules across channels are fused and embedded into the Cross Stage Partial (CSP) structure, enhancing the contribution of feature information. Additionally, the CSP structure combined with standard convolution forms the generalized efficient layer aggregation network (GELAN) structure, effectively integrating low-dimensional and high-dimensional information and enhancing model gradient flow. The proposed network effectively detects cracks in high-altitude images, providing robust technical support for the infrastructure management and maintenance industry.

To address the issues of inaccurate and incomplete obstacle detection, this paper proposes training an obstacle detection model using deep learning techniques. Given the diverse and complex nature of various obstacles, improving detection accuracy and multi-object capturing capability is crucial. An improved YOLOV8 model is proposed in this paper, utilizing object detection algorithms for obstacle detection. Firstly, a Convolutional Block Attention Module (CBAM) is integrated into the feature extraction network of the object detection backbone to enhance the network's focus on obstacle targets. Furthermore, the DeepSort algorithm is combined with the YOLOV8 object detection algorithm to enhance the capability of capturing multiple and dynamic targets. Experimental results demonstrate that the improved YOLOV8 model achieves an mAP (mean average precision) of 94.2% and an accuracy of 95.0%. Compared to the original YOLOV8 model, the accuracy has increased by 3.9%, and the mAP value has increased by 3.7%. Moreover, it outperforms several other object detection models. The research methodology presented in this paper can provide valuable insights for the field of obstacle detection.

This study is dedicated to exploring off-site detection methods to improve the efficiency and accuracy of urban traffic violation recognition, with a special focus on pedestrian red light running. By fusing the attention mechanism and the multi-scale convolutional YOLOv5 pedestrian detection algorithm, a target detection system for off-site environments is designed with high detection accuracy and fast processing speed. The system is able to quickly and accurately identify pedestrians with red-light running behavior with an accuracy of 96.31% after the key steps of preprocessing, feature extraction and target detection of traffic surveillance video images. This research successfully realizes the automatic identification of violation behavior, which is of great theoretical and practical significance for urban traffic safety.

Due to its inadequate capacity to extract visual features, the YOLOv8 algorithm exhibits low accuracy and is not robust enough to handle the issue of several types of garbage obscuring one another. In order to improve feature extraction performance, this article modifies the YOLOv8 network structure. Initially, the traditional convolution in the backbone network is replaced by full-dimensional dynamic convolution. Second, the C2f module in the neck network is replaced with the deformable convolution structure C2F-DCNV3 module. In order to more precisely identify the trash objects that obscure one another, the CBAM attention mechanism module is finally implemented. The experimental results show that the improved algorithm has higher precision and better robustness in dealing with different occlusion types. By introducing the full-dimensional dynamic convolutional module, deformable convolutional module and CBAM attention module, the accuracy is improved by 1.6 percentage points, 2.1 percentage points and 4.3 percentage points respectively, and the feature extraction ability of the model is improved.

Rapid and effective identification of flame smoke at the early stage of a fire in a home kitchen scenario is crucial for reducing casualties and property losses. In this paper, a deformable multi-scale attention-based detection model is proposed for the detection of flame and smoke targets, which has the problems of unfixed morphology and variable features. The model adopts deformable convolutional kernel and deformable attention to enhance the feature extraction ability of the backbone network for the target, and utilizes progressive pyramid network based on adaptive spatial feature fusion and efficient multiscale attention mechanism to enhance the multiscale feature fusion ability of the model. The final experimental results show that the detection accuracy of our model improves by 3.7% when compared with the best benchmark accuracy based on Yolov8s. It proves that our method can effectively detect targets such as flame smoke and prevent fires more reliably.

To address the issues of image blurriness in railway scenarios and the failure to detect small and occluded targets, an improved person detection algorithm based on YOLOv5 is proposed. The feature extraction capability of YOLOv5 is insufficient, and it lacks sensitivity to certain information. To improve these shortcomings, the Convolutional Block Attention Module (CBAM) is introduced into the original YOLOv5 algorithm to enhance feature extraction capabilities and mitigate the problem of missed detections for occluded and small targets. Additionally, the SIoU loss function is used in place of the CIoU loss function as the bounding box regression loss function to accelerate model convergence and improve bounding box localization accuracy. Experimental results show that the average detection accuracy of the model reaches 99.4%, and the detection speed is 41 frames per second, enabling rapid and accurate detection of foreign objects on the railway, meeting the requirements for real-time target detection.

To address the challenges of data security, labor-intensive data annotation, and limited data availability in photovoltaic panel defect detection, a federated contrastive self-supervised pre-training method FCSP is proposed in this paper for photovoltaic panels defect detection. We utilize federated learning to collaboratively pre-train feature extraction backbone in a data-security manner to achieve better model initialization weights. Subsequently, each client loads the federated pretrained weights and conducts supervised fine-tuning with local labeled data for defect detection model training. Experimental results demonstrate that, compared to independent training locally, FCSP achieves an improvement on defect detection, thereby validating the effectiveness of the proposed method.

Considering that current intrusion detection methods have low accuracy and long detection time when network traffic is classified, an improved neural network intrusion detection method based on support vector machine is proposed. Convolutional neural network extracts network traffic locally and deeply, and bidirectional gated recurrent unit extracts network traffic time sequence features. The two methods are combined to form a comprehensive feature extraction method with temporal memory function. Finally, the extracted features are classified by support vector machine instead of SoftMax activation function. According to the experimental results on the NSL-KDD dataset, the accuracy of CNN-BiGRU-SVM model is 99.67%, which is 25.35% higher than that of CNN-BiGRU-SoftMax model, effectively improving the accuracy of network traffic detection.

The existing methods for detecting abnormal human behavior suffer from the large size of parameters, a lack of capacity to extract spatial-temporal features effectively and also exhibit imbalances between positive and negative samples, as well as between difficult and easy samples. To cope with these problems, this paper improves SlowFast by taking in attention mechanism and changing loss function. Firstly, use grayscale video frame clips as input data on the fast path to reduce GFLOPs effectively. Another improvement involves swapping out the original Non-local modules with ANN modules, enhancing the capability to capture spatial-temporal features while also decreasing the parameter count. Then, use Focal Loss to classify the fused feature map, addressing the issue of imbalance between positive and negative samples, as well as the challenge of classifying difficult and easy samples. The effectiveness and superiority of this method were ultimately verified through the AVA dataset and actual scene videos.

This paper proposes the SFN-YOLOv8 detection model for the detection of express packaging defects. This model has the following characteristics: it uses lightweight ShuffleNetV2 as the backbone network of the model and adopts deformable convolution (DCN), thus reducing the number of model parameters. The C2F-CPCA module is proposed in the neck part to enhance the expressive ability of the network in feature representation. Wiou is used in the loss function to reduce the impact of large gradients or harmful gradients appearing in extreme samples in the data set on the loss calculation. Experimental results indicate that compared with the initial model, the SFN-YOLOv8 model has reduced its parameters by 0.15M, decreased GFLOPs by 0.8G, and improved mAP@0.5 by 1.9%.

Pig detection technology is of paramount importance in modern livestock farming, especially within large-scale breeding environments. It not only significantly enhances breeding efficiency by optimizing feeding strategies through real-time monitoring of the pigs' health and growth status but also substantially reduces the risk of disease transmission, yielding notable economic benefits for farmers. With the livestock industry's shift towards automation and intelligence, the demand for more accurate and efficient pig detection technologies has been growing. Among the technologies employed, the YOLO series of models are frequently utilized for pig detection. However, the natural variety in the postures and forms of pigs can lead to recognition issues with some postures using standard models. Therefore, this paper introduces a Deformable Attention module to the YOLOv8 model, which not only captures more global information but also effectively addresses the issue of target deformability. By adaptively adjusting the shape of the convolutional kernels, the module enhances the capture of target features. Experimental results show that the improved model detects pigs more accurately than the original YOLOv8n model and achieves a 3.2% increase in the mAP^50-95 metric.

To address the issue of detecting foreign objects invading substations, which are difficult to detect and easily obscured, this paper suggests an object recognition method utilizing YOLOv5s. Firstly, the C3 module in the backbone network is improved to enhance its feature extraction capability. Secondly, the DCNv2 convolution and SimAM attention mechanism are used to improve SPPFCSPC and serve as the model's spatial pyramid pooling module, allowing the model to better utilize the feature information between the lower and upper layers, thereby improving model performance. Lastly, F-CIOU is used as the model's loss function to balance the issue of significant differences in the number of easy and difficult samples, further enhancing the model's detection accuracy. The experimental findings indicate that the enhanced model’s mean average precision (mAP) has risen by 2.1% over the baseline model, achieving a mAP of 91.9%, which suffices for real-world detection requirements.

Addressing the challenges posed by the complexity of the background in insulator images captured by drones during power system line inspection, the relatively small proportion of insulator faults in the images, and the blurring effects caused by weather and environmental conditions, this paper proposes improvements to the YOLOv5s detection model. Based on the enhanced detection model, a simulation study of insulator fault detection is conducted. By incorporating Dynamic Snake Convolution (DSC) into the network backbone and adding the GAM (Global attention mechanism) attention mechanism at the bottleneck of the network, as well as utilizing EIOU as the loss function calculation method, three approaches are employed to enhance YOLOv5s. Experimental analysis is conducted using Python, and the results demonstrate that the proposed improvements achieve a 2.1% increase in mean Average Precision (mAP) and a 1.1% increase in recall compared to the original YOLOv5s algorithm, while reducing the model parameters by 5.3%. This enhancement scheme ensures both detection accuracy and model size reduction, promising significant prospects for deployment and application in later stages of the model.

The traditional lane detection methods often have the problems of large calculation and low detection efficiency, and their detection accuracy will be seriously affected in complex scenes such as poor lighting conditions, lane damage or occlusion. The UFLD(Ultra Fast Lane Detection) algorithm solves the above problems well. However, UFLD algorithm has some problems, such as weak feature extraction ability, and it can't handle context information well. This paper proposes a module combining Split-Attention mechanism and adaptive spatial feature fusion(ASFF), which not only improves the expression ability of features, but also can selectively learn high-level semantic information to avoid the suppression of detail information by background or global context information, thus improving the model's ability to understand and segment images. The experimental results of ablation show that the detection speed of the improved algorithm is equivalent to that of the original one, and the detection effect is improved by 2.3 and 0.16% compared with the original UFLD algorithm on the data sets of CULane and TuSimple, and the robustness and generalization ability of the improved algorithm are improved.

In response to the problem of poor accuracy in wing non-destructive testing, a wing non-destructive infrared testing technology based on frequent subgraph mining algorithm is proposed. This technology combines infrared imaging technology with frequent subgraph mining algorithms in graph theory. By converting infrared images into graph structures and using algorithms to identify frequently occurring subgraph patterns in the images, convolutional neural networks and decision tree classifiers are applied to achieve efficient detection of wing defects. The experimental results show that this technology can handle the complex and diverse structures and materials of wings, and identify different types of defects by accurately matching subgraph patterns, thereby improving the recall rate of detection. In addition, with the continuous development and optimization of technology, as well as the support of big data and artificial intelligence, this technology can further improve detection performance and efficiency, providing more reliable and accurate support for safety monitoring in the aviation industry.