This PDF file contains the front matter associated with SPIE Proceedings Volume 11519 including the Title Page, Copyright information, and Table of Contents.

Shadow detection is an important part of scene understanding tasks. This paper proposes a network, named Shallow Triple Unet, using shallow Unet as a unit for shadow detection. The network structure is intuitive and the number of parameters is small. With the techniques of hierarchical supervision and results fusion, it can achieve a good shadow detection effect. In order to prove the effectiveness of the network, we performed experiments on popular SBU datasets and compared them with networks such as patched-CNN, stacked-CNN, scGAN, and DSC. The results prove that our network is the best among them, with a BER index of 5.45%. In addition, we also performed ablation experiments to verify the role of various parts of the network. Experiments show that all the techniques we use have significantly improved shadow detection results.

Railway patrolling inspection train has been widely used for railway infrastructure safety monitoring. Cameras are mounted on the train, which can capture the image of the overhead contact power line system for defect detection. In the catenary support device of overhead contact power line system, the insulator can keep the catenary equipment insulated from other equipment. Defect detection of insulators is extremely important to railway safety. In recent years, some achievements have been made in defect detection on railway system based on computer vision. We propose an insulator localization algorithm and insulator defect detection algorithm using deep convolutional neural networks. Firstly, the insulator localization network based on Rotation Region Proposal Network (RRPN) can be used to locate insulator area in catenary support device images by using rotated bounding box. Rotated bounding box can effectively eliminate unnecessary background in localization results. After that, based on the insulator localization results, a Faster R-CNN based insulator defect detection network was used to detect defect of insulator. This method can effectively detect defect of insulator and solve the high false positive defect problem.

Detecting people's presence, monitoring their flows, and their activities, counting how many persons are in a specific place can be strategic goals in different contexts, providing useful insights for different purposes, including those ones related to the management of staying quality in indoor environments. In particular, having information about the actual and current occupancy of a specific room, in specific hours, could be strategic in providing interesting and helpful information for smart building management. In fact, this information could be needed to adequately set the Heat, Ventilation and Air Conditioning (HVAC), the alarm, the lighting systems, and other management issues also. In this context, the Internet of Things paradigm, together with the diffusion of the availability of sensors and smart objects, can provide significant support in monitoring and detecting daily life activities in various situations. Moreover, advancements and specific analysis in image processing can play a strategic role in guaranteeing and improving accuracy, whenever cameras are involved in these situations, to get pictures from the monitored environments. In this paper, we present a people counting approach we have defined and adopted to monitor persons' presence in smart campus classrooms, which is based on the use of cameras and Raspberry Pi platforms. Such an approach has been improved thanks to specific image processing strategies, to be generalized and adopted in different indoor environments, without the need for a specific training phase. The paper presents some evaluation tests we have conducted, showing the accuracy of our approach.

Facial component and landmark detection have many applications in many facial analysis tasks. In this paper, a semisupervised method for this task is proposed to detect facial components and landmarks. Different from other facial detectors algorithms, our model without extra input solve the occlusion problem by detecting the visible facial components. Firstly, we propose a data augmentation method based on the Deep Convolutional Generative Adversarial Network to generate a large amount of semi-supervised training data. Then, a semi-supervised learning model based on Region-based CNN is responsible for multi-task facial component and landmark detection by training on the generated semi-supervised training data. During training, facial component regions and landmarks are used as supervised training data, while unsupervised training data only contains component bounding box. Experimental results illustrate that the proposed model can handle multi-task facial detection, and outperforms the state-of-the-art algorithms.

With the popularity of commercial unmanned aerial vehicles (UAVs), people have easy access to UAV. However, people’s privacy and safety can be threatened if UAV flies at airports, private yards, etc. It is important to be able to detect the illegal UAV accurately and promptly on these vulnerable sites. However, motion blur, occlusion and truncation occur frequently due to fast movement of UAV. It is hard to make correct predictions because of the small size of UAV in images. In this paper, we propose an anchor-free one-stage method for UAV detection. The method eliminates the anchor boxes that are used in most existing detectors, which makes our method simpler and more efficient. We improve the detection accuracy in the following two ways. First, a new multi-scale feature fusion method is proposed to enhance the semantic information exchange between different scales. Second, a reasonable loss function is adopted to increase the proportion of small UAV’s loss. Experimental results validate the effectiveness of our improvements and our proposed detector achieve a superior performance.

Road diseases detection plays an important role in road maintenance. It’s the basis for promoting the healthy development of transportation infrastructure. In this paper, we focus on the method of automatic road diseases detection based on digital image, present a road diseases identification algorithm based on probability superposition edge detection network. Based on the original probability map, we add the probability generated by special disease features. The results of experiment show that the accuracy of this algorithm is significantly higher than the traditional method, especially for the recognition of special diseases.

Abnormal behavior detection in surveillance video is a pivotal part of the intelligent city. Most of the existing methods only consider how to detect anomalies, with less considering to explain the reason of the anomalies. In this work, we investigate an orthogonal perspective based on the reason of these abnormal behaviors. We propose a multivariate fusion method that analyzes each target through three branches: object, action and motion. The object branch focuses on the appearance information, the motion branch focuses on the distribution of the motion features, and the action branch focuses on the action category of the target. The information that these branches focus on is different, and they can complement each other and jointly detect abnormal behavior. The final abnormal score can then be obtained by combining the abnormal scores of the three branches. In the action branch, we also propose an action recognition module using inter-frame information to solve the multi-target and multi-action recognition in the surveillance video, which is not utilized before in the anomaly detection field. The proposed method outperforms the state-of-the-art methods and also can explain why the target is detected as an anomaly.

Road potholes affect comfort, safety, traffic condition and vehicle stability. Accurately detecting these potholes is vital for assessing the degree of pavement distress and developing road maintenance plan accordingly. This paper proposes a simple and effective pothole detection method based on 3D point cloud segmentation. Using binocular stereo vision to acquire 3D point clouds, fitting the pavement plane and then eliminating it from the 3D point clouds of road scene, we could roughly extract the pothole. K-means clustering and region growing algorithms were adopted to extract the potholes precisely. The experimental results demonstrate that our proposed method has a very good segmentation effect on scenes involving plane and target object.

In order to detect kidney cancer automatically from abdominal UCT (unenhanced CT) or CECT (contrastenhanced CT) images at an early stage, a promising approach is to use deep learning techniques with convolutional neural networks (CNNs). However, there still seem to be several challenges in detection of kidney cancer. For example, it is necessary to cope with the wide variety of abdominal CT images. In this paper, as an empirical study, we attempt to construct a CNN that detects kidney cancer well from abdominal CT images, with a special focus on how visual explanations produced by Gradient-weighted Class Activation Mapping (Grad-CAM) help us to construct such a CNN.

Table detection is a crucial step in many document analysis applications as tables are used for presenting essential information to the reader in a structured manner. It is a hard problem due to varying layouts and encodings of the tables. This paper uses Faster R-CNN with the feature pyramid structure as the main network structure to detect the table.It presents a deep learning-based solution for table detection in document images. In order to adapt to different shapes of tables, we classify the tables and preprocess the text with Run Smooth Length algorithm and open source OCR tools. In contrast to most existed table detection methods that only applicable to PDFs, our method can detect document images, which also applies to PDF (because PDF format can be automatically converted to pictures). To evaluate the effectiveness of our method, we tested on ICDAR, UNLV and TableBank datasets, and achieves F1-measure of 92.59% in UNLV, which is higher than some previous methods.

In complex traffic scenarios, there are many problems such as inaccurate vehicle positioning, low recognition rate, and missed detection of vehicles for instability of the traffic environment. This paper proposes an improved Faster R-CNN algorithm for accurate vehicle detection. We use feature maps of depth images to supplement vehicle details by adding a depth channel into the detection model. When training the model, we add a hard sample mining strategy. We evaluate our newly proposed approach using KITTI dataset. The experimental results show that our proposed approach has a significant improvement in recognition accuracy by 5%.

Constant false alarm rate (CFAR) detection is an important method which is widely used in radar automatic detection. However, the performance of CFAR detection degrades dramatically in non-homogeneous environment because of the existence of interference targets and clutter in the reference window of the test cell. To the end, an intelligent multistrategy fusion (IMSF) CFAR detection algorithm is proposed in this paper. By combining the preprocessing results of FOCA-CFAR, BOCA-CFAR and OSVI-CFAR, IMSF-CFAR can utilize more suitable independent, identically distributed reference cells than conventional CFAR method, which means better detection performance. The simulation results reveal that the IMSF-CFAR maintains stable detection performance both in the homogeneous and heterogeneous environments.

As the number of vehicles soared, demand for parking spaces became more intense. Although there are many horizontal parking spaces located outdoor, many existing methods cannot detect horizontal parking spaces effectively due to the complex environment of outdoor and the long span of horizontal parking space. In this paper, a method based on vehicle-mounted fisheye image is proposed for outdoor parking space detection. While a car goes through a parking space at idle speed, fisheye images are captured by vehicle-mounted fisheye camera. Image processing algorithms are used to remove noise and classify light intensity. Then, different grayscale thresholds are used to enhance image contrast and Canny operator is utilized to detect contour information of the parking space line. Next, Hough transform is performed to detect straight-line segments. Angles of the line segments are calculated to determine whether the segments are perpendicular to each other. If right angle is detected in two consecutive frames, the second frame is selected as signal of starting or ending position of parking space. Different distances between space line and wheels are selected to verify that the method has good adaptability to distance. Experiments show that the proposed method can detect parking line effectively and meet real-time requirement.

Positron emission tomography/Computed Tomography (PET/CT) scanning is viewed as one of most effective technologies for lung tumor diagnosis. However, with increasing application of PET/CT scanning, massive medical images were generated, and radiologists are suffering an increasing workload. Hence, it could effectively curtail the workload of radiologist by developing the Computer-aided Diagnosis method based on Artificial Intelligence technology. In this paper, a novel approach for lung tumor detection using PET/CT imaging based on multiscale and multimodality Mask Region-based Convolutional Neural Network (Mask R-CNN) was proposed. We firstly produced three Mask R- CNN models to extract lung tumor candidate. Of the three models, two were trained using PET images with two different scales, and the third one was trained using CT images. Each training dataset included 400 lung tumor slices. Then these three models were combined using method of intergrated learning. 160 axial slices, including 50 PET slices and 50 CT slices with lung tumor and 30 PET normal slices and 30 CT normal slices, were employed for testing. The F- score, precision and recall of novel method were 0.94, 0.92, and 0.98, respectively. The experimental results showed that the proposed method has the capbility of detecting lung tumor effectively and identifying a healthy chest pattern

Boundary and edge cues are very useful in improving various visual tasks, such as semantic segmentation, object recognition, stereo vision, and object generation. In recent years, the issue of edge detection has been revisited, and deep learning has made significant progress. The traditional edge detection is a challenging two-category problem, and the Multi-category semantic edge detection is a more challenging problem. And we model the edge detection of cultural relics and classify the pixels of cultural relics. To this end, we propose a novel end-to-end deep semantic edge learning architecture based on ResNet. Then, we proposed an adaptive class weighter for this problem to supervise the training. The results show that the proposed architecture is superior to the existing semantic edge detection methods in our own design of cultural relic edge detection performance.

In recent years, the object detection technology based on deep learning has made great breakthroughs, greatly improving the detection accuracy. However, most of the existing deep learning detection models are designed for multi-class object detection in natural scenes, which may lead to over-fitting when applied in structured specific railway scenes. Secondly, in order to meet the real-time detection requirements of high-speed comprehensive detection train with a speed of 350 km/h, the detection speed is put forward with extremely high requirements, and the existing deep learning model is difficult to meet the timeliness of high-speed detection. In this paper, we propose an optimized structured regions fully convolutional Networks (SR-FCN), which change the multiple small objects detection problem into single structured region location problem. The structured prior information of rail track is fused into the various processes of deep learning network including that sample construction, proposal region generation, network building and loss function constraint. By optimizing the regional proposal network as well as the anchor’s traversal number, the locating speed of the railway objects is greatly improved, and the locating error caused by local missing or background interference is avoided, which improves the robustness of detection. The experimental results show that the proposed SR-FCN network can not only achieve a high detection accuracy up to 99.99%, but also maintain a fast detection speed, which can meet the real-time detection at the high speed of 350 km/h.

Low-light images are not suitable for direct use in computer vision tasks due to the low visibility of the images. The existing low-light image enhancement methods usually produce colour distortion and noise amplification. This paper proposes a low-light image enhancement method based on multi-feature learning. Our method is different from most methods that decompose the image into two parts: an illumination image and a reflection image. In our learning model, these features are designed based on the pixel level, which makes the model concise and ensures colour fidelity. Our network learns three categories of image features: global features, local features, and texture features. A loss function part based on SSIM is used to ensure that multiple features are extracted effectively. Furthermore, a loss function part based on Sobel is designed to suppress noise and protect the image details. Subjective and objective experimental results demonstrate the effectiveness of our approach for low-light image enhancement.

Currently, image-to-image translation tasks have been extensively investigated by Generative Adversarial Network (GAN) ¹ and Neural Style² . However, the model of the GAN can easily lead to network collapse, making the generation process free and uncontrollable. Neural Style is more dependent on the selection of features layers and loss weights. Therefore, these may cause the network model to fail to produce deterministic results we wanted, and applications are very limited. To resolve these problems, we propose an improved transfer model based on CycleGAN and an image preprocessing method for the particularity of training datasets. In this paper, we convert the synthetic face images generated by simulator into more realistic face images. Besides, this process is reversible. Depending on your tasks and goals, you can choose from synthetic images to real images or from real images to synthetic images. This is valuable and challenging in computer vision and computer graphics. In the results, we compare our model with the most recent GAN models from qualitatively and quantitatively. We demonstrate the state-of-the-art effectiveness of the proposed method for generating high-quality images. You can see more results and discussion of our model on this website https://github.com/qiqi7788/dataset-and-result.

This work presents an automatic pixel-based approach for image segmentation in hemispherical images obtained by a stereoscopic imaging device with fish-eye lenses for forest inventories purposes. In this way, the objective of the proposed image segmentation process is the identification of tree stems. To that end, four geometrical features were obtained for each pixel taken into account the intensity, the local color variance in both radial and tangential directions, and the greenness ratio. Thus, the proposed approach is an automated method based on a global image classification based on the four aforementioned features followed by a constrained region growing process. As a result, each pixel is classified as belonging to Sky, Foliage and Stem texture classes. The quality of the segmented images was successfully evaluated against two segmentation strategies previously used for similar purposes, by using three well-known image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Feature Similarity Index Metric (FSIM). The results obtained in this work confirm the relevance of this automatic proposal.

Here we propose a novel image inpainting model DFG-GAN, which can effectively alleviate the artifacts problem when the missing region area is too large. Unlike other image inpainting models, our model can transfer the image inpainting task into a GAN task when the mask fills the total image. Apart from that, we also take advantage of the extra class label information to tell what kind of the damaged image is. The more information feed in, the better result shall be. Experiments on several publicly available datasets demonstrate the advantage of the proposed method over existing approaches, regarding both visual fidelity and margin texture.

Khao Dawk Mali 105 (KDML105), internationally known as "Jasmine Rice", is one of the most famous and major commercial rice in Thailand. Physical appearance of polished grain is one of key factors that influences rice’s price. One of major traits is a degree of chalkiness. Rice breeding scientists thus make a great effort in reducing grain chalkiness in order to meet market quality and match consumer preference. The routine task in breeding process is visually inspection of chalkiness level in rice. Since human visual inspection is slow, subjective, and not consistent over a long period, we propose to use global thresholding methods to automatically segment chalk area in order to improve speed of chalkiness inspection and provide objective and consistent results. However, due to the characteristics of rice chalk that causes several difficulties in a thresholding mechanism, we thus proposed a new method for improving histogram in order that the global threshold value can be computed easier. The proposed histogram improvement method has several desirable advantages, such as very low computational cost, efficiently dealing with the problem of low contrast, and insensitivity to size and location of objects. The effectiveness of the proposed histogram improvement method is evaluated using four well-known global thresholding methods on a real 96 chalky grain images with different degrees and variety characteristics of chalkiness. The accuracy of segmented chalk area was verified by comparing it with human segmentations produced by rice researchers. Experimental results demonstrate that the proposed histogram improvement method can significantly improve the segmentation results.

High-quality depth information is urgently required with their increasingly wide application in many real-world multimedia fields. However, due to the limitation of depth sensing technology, the captured depth map in practice usually owns low resolution and poor quality, which limits its practical application. As we all know, consistency between high-quality color images and low-quality depth maps achieves good effects in depth super-resolution. But the edge inconsistency also limits the recovery of depth map. Inspired by the geometric relationship between surface normal of a 3D scene and their distance from camera, we discover that there are more consistency between surface normal map and depth map in the edge areas. Meanwhile, surface normal map can provide more spatial geometric constraints for depth map reconstruction, for both of them are special images with spatial information, which we called 2.5D images. In this paper, we propose a unified framework, Normal Data Guided Depth Map Restoration with Edge-Preserving Smoothing Regularization (NDEPS) method, via joint spatial domain and gradient domain regularization, one characterizing the relationship between surface normal data and depth in the spatial domain and another edge-aware constraint in the gradient domain. The proposed NDEPS method formulates a constrained optimization problem that can be solved by an iterative conjugate gradient(CG) algorithm. Extensive quantitative and qualitative evaluations compared with state-of-the-art depth recovery methods show the effectiveness and superiority of our method.

Recent deep learning approaches have shown significant improvements in the challenging task of image inpainting. However, these methods may generate blurry output and distorted textures. In this paper, we propose an efficient end-to-end two-stage network for image inpainting. In the coarse stage, we employ residual dense block (RDB) as well as short and long skip connections to fully leverage and exploit features from all convolutional layers and give a globally rough reconstruction. In the refinement stage, we propose a local and global residual network based on channel and spatial attention block (CSAB) to adaptively weigh both channel-wise and spatial-wise features focusing on more meaningful information, and generate a locally fine-detailed image. Experiments on Paris StreetView and DTD textures demonstrate the effectiveness and efficiency of our method. Results show that our method outperforms the baseline techniques quantitatively and qualitatively.

The advantage of the adaptive filter is that it can automatically adjust the parameters in the filtering process when the statistical characteristics of the input signal are unknown. Therefore, it can achieve better results when applied to image processing than traditional filtering methods. For pepper and salt noise, the algorithm in this paper is based on judging the difference of image gray level and using the fuzzy function to process the image. Through the comparison of normalized mean square error (NMSE) and peak signal to noise ratio (PSNR), the advantages of the proposed algorithm over the traditional adaptive filtering algorithm in preserving image details are verified.

OSVOS is one of the best algorithms in video semantic segmentation of single-target. CBAM module is proposed in object detection and classification. This module can improve the performance of the network without adding extra computation. In view of this, this paper proposes an end-to-end network AttnVSS based on attention and VGG-16. CBAM is embedded in the shallow layer of our network. The network is first pre-trained on ImageNet, and then full connection layer is removed and fine-tuned to meet the segmentation requirements. Through the ablation experiment on DAVIS dataset, it is proved that CBAM can be embedded into semantic segmentation networks. At the same time, it can help the network to allocate "attention", focus on the most meaningful part of the input, to achieve faster and more accurate segmentation.

The initial assessment of any skin disease is usually made by visual inspection of doctors and skin specialists. Further tests may be recommended such as biopsy and pathological examination for a more accurate diagnosis. With the use of skin disease identification system, the diagnosis of infected skin is readily attainable without undergoing biopsy and pathological examination. The infected skin disease image is identified using SIFT algorithms with local features and KNN classifier. The skin disease that will be identified are acne, psoriasis, eczema, rashes, hives, warts, tinea versicolor (an-an) and unknown skin disease. The system was confirmed to be efficient in identifying the aforementioned skin diseases. Identification of infected skin images is accomplished by K-Nearest Neighbors (K-NN) algorithm which shows an accuracy of 90% in functionality testing.

Aiming at the problem that the current book call number recognition system does not consider background color interference and intercepts the book number area directly and subjectively, a method is designed to directly process the random image of books on shelf under the complex color background, so as to realize the self localization of the book number area and the identification of the book number. Filter out other color background interference, locate the call number area, and split the spine image and call number characters, and finally use the template matching method to identify characters. The experimental results show that the accuracy of the algorithm is 96.11% and the recognition rate is 90.83%. Compared with the common methods of combining color space and mathematical morphology processing, the accuracy of location is improved by 31.08%, and it is not affected by complex color backgrounds. It has strong antiinterference and high efficiency.

Periocular recognition has been gaining attention as one of the promising biometrics as it contains rich information of the ocular, skin and eyes color as well as eyebrow. Present researches of periocular recognition in the wild mainly are based on the convolutional neural networks that are equipped with standard cross-entropy loss. Label smoothing regularization (LSR) has been recognized as an effective regularization technique for generalization improvement. LSR optimizes the network based on a weighted combination of cross-entropy loss and KL divergence of uniform and network prediction distributions. In this paper, we extend LSR to Learned LSR (L2SR) by considering learned smoothen prediction distribution instead of predefined uniform distribution. L2SR outperforms LSR at reducing intra-class variation and, thus, improve the generalization. Extensive experiments on three periocular in the wild benchmarking datasets demonstrate the effectiveness and superiority of our method.

This paper presents an effective method to recognize mechanical sphygmomanometer dial value based on Hough transform. The edges of image are detected by using the canny operator on the binarized image. Geometric features in the image are obtained based on Hough transform, which includes the center of the dial and the vertex of the pointer. After connecting the center of the dial and the vertex of the pointer, a pointer line segment is obtained, and then an accurate dial value is effectively computed according to the deflection angle of the line segment. Experiments show that the method in this paper has wide applicability and can be used to display the waveform of numerical changes more accurately in the video.

Person re-identification (Re-ID) is an important technique towards the automatic search of a person’s presence in a surveillance video or security systems. Applying incremental learning techniques to accelerate the online training speed with ever-increasing data is desired and critical for Re-ID. As an incremental learning algorithm, Incremental Kernel Null Foley-Sammon Transform (IKNFST) method significantly reduces the computational complexity while holds the accuracy. However, with ever-increasing person samples within the same category, the corresponding growth of dimensions makes it difficult to update the online model. To address the issue, we propose to maintain constant update speed by constructing Reduce Set (RS) expansions during online updating. The key idea is to firstly extract new information brought by newly-added samples and integrate it with the existing model by Incremental Kernel Principal Component Analysis (IKPCA) scheme for further Reduce Set (RS) compression. And the compressed samples and the corresponding model are then input to Kernel Null Foley-Sammon Transform (KNFST) algorithm for generating an updated model. Extensive experiments have been carried on three public datasets, including Market-1501, DukeMTMCReID and CUHK03. The results show that our proposed method beats the state-of-the-art IKNFST by a big margin.

The atmospheric scattering phase function represents the angular distribution characteristics of atmospheric scattering, which is a key index to understand the influence of atmospheric molecules and aerosols on climate. The lateral scattering lidar has become one of the effective methods to study the scattering coefficient and scattering phase function of the near ground atmosphere. Based on Monte Carlo simulation algorithm, this paper simulates the propagation process of the lateral scattering lidar photons in the atmospheric medium, and obtains the simulation image of the lateral scattering lidar. By analyzing the principle of the lateral scattering lidar, the profile of the lateral scattering coefficient is obtained by using the curve relation between the relative light intensity and the position according to the equation of the lateral scattering lidar.

The combination of rule-based reasoning (RBR) and case-based reasoning (CBR) is used to build an expert system for the endoscopic detection of aero-engines. The aircraft maintenance manual is converted into knowledge rules, and a rule base is constructed. Typical damages and corresponding maintenance decisions adopted are taken as cases to construct a case base. The rule base and the case base together form the expert system knowledge base, which is stored in the database. At the same time, the corresponding learning mechanism is established to realize the maintenance of diagnostic rules and typical cases. When the expert system works, through the man-machine interface, the user inputs the image related information, extracts the image damage information, carries on the reasoning in the rule base, obtains the maintenance decision according to the matching rule, regarding is difficult to make the maintenance decision in the critical state damage, needs to carry on the case-based reasoning, retrieves the similar historical case from the case base, finally unifies RBR and CBR reasoning result to make the maintenance decision comprehensively.

The objective of this work is to explore the feasibility and accuracy of using a fixed-wing unmanned aerial vehicle (UAV) to obtain slope surface information data. Herein, the precision of the slope model based on UAV images was evaluated using ground control points and the point cloud data obtained by a terrestrial laser scanner. The results show that the proposed model is accurate at the decimeter level. The proposed method is reliable and practical and can be used in similar projects to obtain high-precision surface data on slopes.

State-of-the-art object detection networks have reduced the running time and get better detection results. However, for remote sensing image scenes, in a certain projection direction, the remote sensing image target will be tilted, The positioning of the horizontal bounding box used by the exit detection algorithm will cause a lot of overlap between the target bounding boxes, After using NMS (Non-maximum suppression),it will lead to the lost of the target. In this paper,we propose a Rotated Faster R-CNN(R-FRCNN) that is a target positioning method based on arbitrary angle bounding box,which can perform non-redundant positioning on the target,thus can reduce the missed detection rate. when the target is densely distributed or the angle of the object is arbitrary.Compared with traditional and state-ofthe art object detection algorithms,our approach obtain the superior performance.

At present, mature technology of images fusion is generally implemented by software. But it has been proved to have a bad real-time capability. This paper designs a new system to resolve the problems of bad real time performance on traditional method. Based upon FPGA, the system processes the infrared and low-level-light images then output images data to the display module in real time. In order to meet requirement of accuracy and optimization, this system takes a registration method based on parallel optical axis. Also, it adopts the bionic algorithm for color fusion of infrared and low Level light images. The images resultant increased by an average of 3.86% and 8.77% than Waxman algorithm in terms of mutual information and structural similarity respectively. The exceptional quality images we can obtain shows that the system is useful in images processing. The processing and transferring time of one frame image is 0.043434s in total shows that the system performs better in real-time. In addition, this system is valuable for defense and security.

Surface penetrating imaging technology reconstructs target images that conform to visual perception by interpreting radar data, which reflects the target's geometry and electromagnetic scattering characteristics directly. For surface penetration imaging, various sources of clutter cause failure of image interpretation, therefore clutter suppression is the key to obtain target information accurately. Though clutter suppression approaches aiming at the strong clutter make the target image prominent enough for imaging process, the weak clutter with energy close to the target affects the imaging quality. In this paper, a nonlinear gain control method is proposed for the suppression of weak clutter and improving the contrast of the imaging by non-linearly weakening the data below the threshold and enhancing the data above the threshold. Experiments show that the proposed method can effectively improve the imaging quality.

We address the problem of single-shot color multiplexed digital holography and propose its new technical realization. To acquire a three-wavelength interference pattern, we propose to implement spectral filtration of wide-band light by means of acousto-optical tunable filter (AOTF) providing three-band operation. We describe digital image processing procedure for amplitude and phase information extraction, and confirm its effectiveness experimentally. Proposed approach may be effective for the detection and classification of structural elements without spectral or mechanical scanning.

In this paper, we propose a new design of an imaging optical system combining the advantages of stereoscopy and spectroscopy principles. The quality of spectral stereoscopic image is the key factor influencing the accuracy of threedimensional (3D) measurements and shape reconstruction. The main advantages of acousto-optical tunable filters (AOTF) in combination with stereoscopic imager are shown. We developed a new scheme of an imaging AOTF-based stereospectrometer in order to raise accuracy of 3D-measurements. According to numerical simulation, the scheme has improved image quality compared with existing ones.

In this paper, the authors discuss the results of the IA based on deep learning to improve the productivity of preventive thermographic inspections of the electrical network when compared to manual methods of analysis. The goal of this process is to use computer to capture and recognize images of hot spots from the distribution powergrid network during car’s inspection deslocation. In order to that, a special vehicle was assembled. Initially the vehicle was equipped with eight cameras to proceed inspection in both side of the road and covers the front view as well. This solution results in the capability to inspect hundreds of miles of power distribution lines without the need to stop the vehicle and without the need for a human operator.

Airborne surveillance and reconnaissance play an increasingly crucial role in military operations. Aerial video can perceive dynamic events, track active targets and provide the real-time location of targets, but the poor accuracy of geographical coordinates limits its usefulness. Registration of aerial video to geo-referenced images is an effective means to improve the location accuracy. Generally, the aerial video and the geo-referenced images to be registered were taken at different times, from different viewpoints and by different cameras. Due to these differences, registration of aerial video to geo-referenced images is a challenging task. This paper discusses the essential issues of the problem, analyzes the characteristics of the images, and shows what features work better. Then, this paper presents a two-stage image registration algorithm, which is based on decomposition of transformation space. The first stage estimates the rotation parameter by matching histograms of weighted edge orientation between aerial image and geo-referenced image. The second stage determines the translation parameters by comparing partial images using partial Hausdorff distance. This approach is computationally efficient and robust to significant differences between the images to be aligned. Experiments have been conducted on real-world data, and the results have shown the good performance of the algorithm.

Visual analytics has been in the limelight as a powerful tool to support large scale management of places, people, and activities. Harnessing the power of machine learning is possible to quickly identify critical issues across thousands of cameras. Several stakeholders have voiced concerns about privacy. Visual analytics techniques can be used in facial recognition thus enabling fine-grain user tracking. This paper addresses such privacy concerns for some specific scenarios. It explores the feasibility of visual analytics in using low-cost/low-resolution thermal cameras thus delivering context-awareness information yet protecting user’s privacy. This paper proposes a model able to classify and count humans, in indoor hallway settings, using low-resolution thermal pictures. The model is designed to work with YOLOv3 and leverages the power of deep-learning. Results show that it is possible to classify and count humans with over 90% accuracy based on the images from a low-cost 80x60 pixel thermal camera. The results were evaluated against the ground truth checked by a human agent and recorded through a regular camera. The study exposed possibilities and limits offered by low-cost thermal cameras and identifies the potential application scenarios. The dataset including both real and thermal images used for the training and the testing will be made available to the scientific community.

Burgeoning virtual reality technology allows people to experience the content of video authentically. In order to provide an immersive experience to user, virtual reality video needs higher resolution (4K and higher) and more data than traditional video, which increases the coding complexity extraordinary. In this paper, a fast intra mode decision algorithm based on the sum of region-directional dispersion is proposed. Based on the diverse degree of horizontal stretching in different latitude area, the relationship of the optimal intra prediction mode related to latitude area for coding units is obtained. A new metric named the sum of region-directional dispersion is defined to determine the mode selection interval, and the prediction mode where located the minimum sum of dispersion is used as the center mode of the selection interval. The optimal intra prediction mode is obtained by comparing it with other adjacent modes in the interval. Compared with the original reference software HM16.20, the proposed algorithm can reduce the time in coding by 28.5%, while BD-rate only increases by 1.0%.

Image classification and recognition has always been a key issue in the field of computer vision. Not only has it achieved great development, but its research space has broad prospects. This paper first explains the advantages and disadvantages of the traditional image classification methods, and then introduces the development on the neural network model. The algorithm proposed in this paper is mainly based on the LeNet-5 network, and the experiment is performed on the CIFAR-10 dataset. The improved neural network is superior to original model in classification accuracy.

Virtual reality 360-degree video is now widely used in various fields. However, the huge amount of data caused by its high resolution increases the coding complexity. While the existing video coding framework, such as HEVC, is not optimized according to the characteristics of virtual reality video, in order to accelerate the speed of intra-mode decision process in video coding, a new algorithm is proposed in this paper. Each virtual reality 360-degree video frame is divided into two regions, namely the pole region and the equatorial region. The number of candidate modes in the pole region is reduced in advance according to the statistical distribution pattern of the modes, and intra-mode in the equatorial region is determined by the cart decision tree using the texture contrast of the coding unit. The experimental results show that compared with the original HM16.20, the proposed algorithm can save the coding time by 28.3%, while only increasing the bd-rate by 0.612%.

Existing deep learning-based object removal methods produce plausible results. However, they generate unsatisfactory results when the object is large, especially in facial images due to the lack of information about the affected region. Most of these methods rely on the object information in terms of a binary segmentation map which is insufficient to provide information about the face boundary and semantics symmetric relation. To address the problem, we propose a two-stage GAN-based image-to-image translation method that exploits the face semantic segmentation instead of the binary segmentation map of the object. Specifically, our model learns a complete facial segmentation map from an input image (face image with unwanted object) in the first stage and translates that generated semantic segmentation map combined with input image into a plausible face image without the object. We also exploit the joint loss function that consists of low-level loss, adversarial loss, and perceptual loss to produce semantically realistic facial images. Experimental results show that our method outperforms previous state-of-the-art methods both quantitatively and qualitatively.

One of the significant part of CNN is feature extraction module. For computer vision, image’s patterns are extracted by filters and convolution operations in feature extraction. The CNN learns filter weights from signal data and extracted features. So, widely used filters are based on Gaussian distribution initializing, together with the low-level feature extraction in state-of-the-art architectures. According to CNN, reducing parameters and computing resources while accuracy maintaining is challenging. Most CNN compression research aims at general image which inappropriate for medical image. The MRNet dataset with MRNet CNN model for knee MRI diagnosis assisting by Bien et al. from stanfordML group, have recently been developed. This MRNet achieved AUC of 0.965 on internal ACL tear dataset classification. Based on MRNet, we modified feature extraction module to compress MRNet with measured on ACL tear within MRNet dataset. We designed with a foundation of 2 n form filter, and supplemented by MRI-cut selection. We split MRI by cuts and tested results of each cut combination together. The combination leading to the best accuracy is Coronal/ Sagittal, we used it as input dataset. Then we replaced the filters in baseline model by 2×2 filters, 4×4 filters, medical DIP 8×8 filters, with symmetric padding to eliminate shift problem in even-sized filter. The MRNet baseline (trained from scratch) got average error rate as 8.50% and our proposed got 12.94% but a number of parameters is pruned by 52.250%, a number of FLOPs is pruned by 46.145%, and requiring only Coronal and Sagittal.

Gait can be a concrete manifestation of an individual's athletic traits, but also can become an essential feature of the judgment of the physical condition. Making an accurate gait analysis has become one of the critical medical information needs. However, the measurement and analysis of gait in medical institutions are still mainly based on the visual experience of physicians, or expensive and complicated motion capture system that are not suitable for use in the home environment; and the simple wearable sensors do not provide enough information to assist medical judgment. In this research project, a novel gait analysis method based on an image recorder is proposed, which has the advantages of making the gait analysis more natural and more suitable for the home environment. A more realistic gait profile can be measured without having to be in close contact with the observed body parts like motion capture system or inertial measurement device. This research project, a gait analysis system based on the first-person video will be developed, and the characteristics of captured video will be transformed into the features of clinical medicine. This system will be used in combination with traditional clinical medicine. The experimental results show that the generated measurement information can be correctly converted for clinical medicine using, and is more suitable for home environments. Further, we will cooperate with clinicians to provide a useful evaluation of home exercise rehabilitation and predicting the longterm trend. That can give the appropriate treatment regimen matching the development of the disease.

We propose an automatic face and head deformation method that combines 3D faces with an arbitrary head model. With the rapid development of Computer Vision and Deep Learning, 3d scans of human faces are becoming easier to obtain. How to complete the scanned 3D face data and make it a complete full head model, or give the scanned 3D face different 3D hairstyles, has always been an open question. For this reason, we propose a Global Deformation Model (GDM) which is implemented by multiple iterations. By building a Global Deformation Model (GDM), a full-head 3D data with complex hairstyles could be combined with 3D face data. In this way, the scanned face is automatically completed as a full-head model. Experiments show that compared with other deformation algorithms and full-head reconstruction methods, our method has better automation and robustness. It shows good deformation results in complex 3D data. We provide an attractive solution for graphic design, Virtual Reality, 3D printing, and other industries, which can be widely used in consumer scenes.

Laser speckle is a long-standing issue which will corrupt image formation and interpretation. Over the years, various techniques have been developed to mitigate the speckle noise. Although machine learning approaches have been widely applied for image denoising tasks, there are very few of them are specifically designed for speckle reduction. In this work, we present a network specialized for reducing speckle noise in images. The training set consists of nearly 3,000 coherent- and incoherent-illuminated image pairs of a variety objects. The network is trained to learn a transformation from speckled to speckle-free images. We compare against the traditional image processing methods, and show that our learning-based approach outperforms these methods regarding both high PSNR, SSIM results and maintaining high-frequency edge features. Our data-driven approach dramatically reduces laser speckle noise by 11.71 dB, compared to a 0.17 dB reduction from non-local means filtering, a 0.1 dB reduction from median filtering and a 0.12 dB reduction from Gaussian filtering. Moreover, conventional image processing approaches reduce both laser speckle and high-frequency image features, which will result in the blurring effect. In contrast with optical speckle-denoising approaches, our method reduces cost and computational complexity. So, applications that require small and bright illumination sources with high-quality imaging can benefit from our work.

High dynamic range (HDR) image reconstruction of dynamic scenes from several low dynamic range (LDR) images captured with different exposure time is a challenging problem. Although several methods based on optical flow or patch-match have been proposed to address this problem, they are not robust enough and results still suffer from ghost-like artifacts for challenging scenes where large foreground motions exist. To this end, this paper proposes a multi-scale contextual attention guided alignment network called CAHDRNet and presents some evaluation results. In stark contrast to methods based on optical flow, this paper demonstrates that HDR reconstruction can be formulated as an image inpainting task and CAHDRNet conducts patch replacement on the deep feature maps unlike previous patch-based reconstruction methods. The contextual attention module proposed by an image inpainting work is extended in CAHDRNet with multi-scale attention rebalance to help model flexibly handle different scenes and reduce patch replacement error. Experiments on the public dataset indicate that the proposed CAHDRNet produces ghost-free results where detail in ill-exposed areas is well recovered. The proposed method scores 40.97 on test sequences with PSNR metric averagely while the best PSNR score of non-flow-based methods is 38.60. The flow-based method scores 40.95 with PSNR metric while it has 5 points better score than our result with HDR-VDP-2 metric. According to quantitative and qualitative evaluations, the proposed method outperforms all non-flow-based methods and has its merits and demerits compared with the flow-based method.

Digital holography can record and display three-dimensional information of objects, and is widely used in many fields. Due to the coherence of the light source, the digital hologram is destroyed by speckle noise, which seriously influences the quality of the reconstructed image. In order to remove speckle noise, in this paper, we combine random overlapping masks and 3D block matching filtering. This method uses only one hologram, which is simple and time-saving. The experimental results show that when the image noise is large, the proposed method can still achieve higher peak signalto-noise ratio (PSNR) and structural similarity (SSIM), especially the structural similarity, reaching 0.99, and meanwhile the relative deviation and visual effect both are excellent.

In this paper, we address the challenging task of estimating 6D object poses from a single RGB image. Motivated by the deep learning-based object detection methods, we propose a concise and efficient network that integrates 6D object pose parameter estimation into the object detection framework. Furthermore, for more robust estimation to occlusion, a nonlocal self-attention module is introduced. The experimental results show that the proposed method reaches the state-ofthe-art performance on the YCB-video and the Linemod datasets.

Hashing technology is widely used to solve the problem of large-scale Remote Sensing (RS) image retrieval due to its high speed and low memory. Among the existing hashing algorithm, the unsupervised method is widely used in largescale RS image retrieval. However, the existing unsupervised RS image retrieval methods do not consider the multichannel properties of multi-spectral RS images and the discriminability in the local preservation mapping process adequately, which make it difficult to satisfy the retrieval performance of RS data. To solve these problems, we propose an unsupervised Variational Auto-Encoder Hashing algorithm based on multi-channel feature fusion (VAEH). MultiChannel Feature Fusion (MCFF) is used to extract the feature information of image, which fully considers the multichannel properties of the multi-spectral RS image. In order to enhance the discriminability in the local preservation mapping process, variational construction process and automatic encoder are added into the learning process of hashing function, and the KL distance of the Variational Auto-Encoder (VAE) is used to constrain the hashing code. Experiments on two large public RS image data sets (i.e. SAT-4 and SAT-6) have shown that our VAEH method outperforms the state of the art.

This paper addresses the image compressed sensing recovery problem. To improve the recovery quality, instead of using a fixed dictionary that is generally a universal one trained in an off-line manner for sparse representations of image patches, we adopt an adaptive dictionary learning strategy. Inspired by the monotone fast iterative shrinkagethresholding algorithm, a dictionary learning algorithm is introduced in this work. Also, we abandon the classic method that breaks an image into fully overlapping patches, and propose a new overlapping patches extraction method, which decreases the number of patches and saves much run-time, while achieves similar recovery qualities.

A number of studies demonstrate that illumination is an important factor impacting the performance of computer vision tasks and illumination normalization can improve the performance of other visual analysis algorithms. At present, there are few methods aiming to illumination normalization of color face with deep learning. For this reason, we put forward a novel and practical deep fully convolutional neural network architecture for illumination normalization of color face. Comparing with the current methods based on deep learning, our approach does not need to input identity and illumination label. We preserve the identity by a well-designed generator and content loss. Experimental results show that the proposed method achieves favorable illumination normalization effect under various lighting variances and preserves identity effectively.

FPGA is popular for real time image process attributed to its flexibility and parallelism. In this paper, a parallel fast median filtering algorithm is implemented by FPGA. The times of comparison can be reduced by half compared to the bubble sort method. To keep more border information of image, an efficient implement of border management is proposed in detail. Meanwhile, a speed up method presented is more effective in the real-time application. The proposed method shows good performance in eliminating noise. The low resource consuming and the low latency is meaningful for real time image processing application.

To achieve registration of multi-sensor images by utilizing complementary information, this paper proposes an iterative image registration method based on scale invariant feature transformation (SIFT) and extended phase correlation (EPC), named as SIFT_IEPC. The reference image and the sensed image are pre-registered by SIFT and a geometrical outlier removal method. Overlapping regions corresponding to the reference image and the rectified sensed image are partitioned to block image with equal size, and the extended phase correlation is used to estimate the translation parameters between each block pairs, which are used to tune the matched feature point pairs in the block. The tuned feature point sets are used to update the registration parameters between the reference image and the sensed image. Repeat the process of EPC matching and feature tuning until terminate condition is satisfied. Experiments on three pairs including simulated and real remote sensing images are conducted to evaluate the performance of SIFT_IEPC. The comparison experiments demonstrate that SIFT_IEPC can apparently increase the accuracy of image registration.

Since the cost of labeling data is getting higher and higher, it is essential to develop some algorithms for practical applications especially when only a small number of labeled images are available. We deal with this by designing a novel model and show that with a small amount of unlabeled samples added to the training set the classification result can be effectively promoted. In our method the query image is collaboratively represented by the whole training samples, whether they are labeled or unlabeled. There are two regularization terms. Rather than getting the sample sparsity, the first regularization term, a variant Trace Lasso norm, forces the group sparsity, which means the query image is presented by a small number of groups, say one or two groups, and in each group the training samples are fully used. This sparsity between groups and density within a group is preferred to the aim of face classification. The similarity matrix is used in the second term to obtain the pseudo-labels for the unlabeled training samples, thus the whole samples are classified into different groups, each one corresponding to a class. Augmented Lagrange Multiplier scheme and Alternating Direction Method of Multiplier are given to solve the problem numerically. Some experiments are conducted to show the performance of the proposed work.

Kernelized Correlation Filter (KCF) algorithm has been successfully applied in object tracking. By finding the maximum confidence in the search region, the candidate target block is determined by KCF with constant size. When the scale of the target changes during its actual movement, the traditional KCF algorithms often fail to track the target. Moreover, it is difficult to judge whether the object is missing due to the lack of self-adaptive threshold regulation scheme in the KCF tracking. In order to solve these problems, this paper proposes a scale-adaptive target tracking algorithm based on KCF, which is mainly divided into the following steps. Firstly, the positive and negative samples of the nearest neighbor classifier are initialized by the selected target and its surrounding non-target areas. Secondly, the peak point of the spectral response of the current frame image is obtained by executing the KCF algorithm, which is the target center point. Thirdly, the scale change of object is obtained by calculating the ratio of the bandwidth of spectral response peak centered candidate regions and the target in the previous frame. Fourthly, the scaled candidate target is confirmed by calculating the sample similarity between it and the Nearest Neighbor Classifier (NNC). Finally, the positive and negative samples of the nearest neighbor classifier are updated with the confirmed tracking target and non-target respectively. Extensive experiments have been carried on four test video sequences. The experimental results show that our proposed method achieves a higher success rate and accuracy with less running time compared with the state-of-the- art methods.

With the development of Geospatial technology and remote sensing technology, a large number of remote sensing images come into application with assist of computers. Although convolutional networks have great performance in computer vision, features extracted by convolutional network doesn’t have the characteristic of rotation invariance, which means the current neural network methods can’t adapt to rotated objects. Considering the multiangle characteristics of remote sensing images, we proposed a Rotation Invariance Spatial transformation Network (RI-STNET) to extract the rotation invariance object features. RI-ST-NET combines convolutional neural networks and the Spatial Transformer Networks (STN) rotating the object to an angle which more easily to identify and is trained by means of Siamese network sharing the same weights of two network branches . Thus RI-ST-NET can adapt to the object features of different rotation patterns which then improved that effectively promote the accuracy of remote sensing retrieval. A Rotation Invariance Spatial Transformation Network combines the advantages of STN and tuple training which can catch the rotation of the same object when used in image retrieval task. A series of evaluation contrast experiments on chosen dataset demonstrate the performance of the proposed method.

This study highlights the importance of one of the most important components when prescription glasses are developed for a patient, which is the Pupillary Distance (PD). PD pertains to the distance between the two pupils. This is an essential information when having prescription glasses created to ensure that the pupils remain in the center of the glasses. This paper explores the application of the You Only Look Once (YOLO) algorithm in order to determine the distance between the two pupils measured. It has been found that the YOLO algorithm can be an efficient tool in measuring distances between the objects, which was proven by linear regression where the computed value and measured value are similar. The research has concluded that it can replicate the results from the conventional method of measuring Pupillary Distance through the use of a ruler.

In this paper, we proposed a novel method for inspection the marking label of IC chip using FIR system. Firstly, the IC chip is loaded to the inspection machine and IC chip image are captured. Secondly, Maximum Stable Extremal Region (MSER) algorithm is applied to find the characters, numbers, and company logo regions of IC chip and crop the detected marking label region. Thirdly, convert the grayscale image to a binary image and segmented the characters, numbers and company logo using connected component labeling algorithm. We observed some IC chip is tilted when moving to marking inspection station and effect to the position of the characters, numbers and company logo are shifted. Thereby, we need to orientate them by using Normalization. Finally, the recognition system based on DCT and FIR system. The example of the correct extracting rate of IC marking inspection was 94%.

Discriminative correlation filter-based trackers get desired performances, but also suffer from unwanted problems such as boundary effect and overfitting. SRDCF tracker uses a spatial penalty matrix to alleviate the effect of boundary effect. STRCF introduces a spatial temporal regularized term to the objective function, which limits the change of filters of two continuous frames. However, both SRDCF and STRCF cannot address the problem of overfitting when targets’ appearance changes dramatically. Our work is an extension of SRDCF tracker and STRCF tracker. We use the observations of previous frames to constraint the filter, so that it is able to mitigate the overfitting problem. Furthermore, we apply incremental principal component analysis (IPCA) to extract the principal information of the previous frames' observations. The objective function can be efficiently solved using the alternating direction method of multipliers (ADMM). Our tracker and 9 state-of-the-art trackers are tested on the OTB-2015 benchmark. The experimental results demonstrate its satisfactory performance.

Hashing is an important technique branch of image retrieval due to its satisfactory retrieval performance with high retrieval speed and low storage cost. Deep supervised hashing methods, which take advantage of the convolutional neural network and the supervised information, have shown better performance than other kinds of hashing methods. However, previous deep hashing methods do not consider the noisy data, which generally exist in large-scale labeled datasets and mislead the learning algorithm. In this paper, we propose a novel robust deep supervised hashing (RDSH) method, in which a robust pairwise loss and a quantitation loss are used to supervise the learning procedure. The quantitation loss guides the CNN to output binary codes. The robust pairwise loss for similarity-preserving learning is designed based on generalizations of the exponential and logarithmic functions. By adjusting its parameters, the robust pairwise loss can exhibit special properties, including tail-heaviness and boundedness, which results in the robustness of the learning procedure to noisy training data. To verify the robustness of the RDSH, we conduct experiments on CIFAR10 with different noisy levels, in which RDSH shows better robustness than other deep supervised hashing methods. Experiments on standard CIFAR-10 and NUS-WIDE datasets show that RDSH outperforms other baselines.

Model compression reduces the computation costs of an over-parameterized network without performance damage, and channel pruning is among the predominant approaches to compress deep neural networks. In this paper, we propose a novel approach called MetaAMC combining meta learning and Auto ML for automatic channel pruning of very deep neural networks. It leverages meta learning and reinforcement learning to provide the model compression policy. Compared to not only the original AMC and MetaPruning but also the state-of-the-art pruning methods, we have demonstrated superior performances on MobileNet V1/V2 and ResNet-50.

In the daily driving process, road potholes pose a great threat to traffic safety. However, the actual road potholes are often irregular and the background is complex. So, it is difficult to accurately measure the volume of potholes. In this paper, a method for measuring the volume of road potholes based on three-dimensional point clouds is presented. First, binocular vision is used to obtain 3D point cloud data of the pothole, and the segmented pothole point clouds are projected onto the coordinate plane established by the road surface. Then, we triangulate projection points on the coordinate plane. Finally, restore these points to the true elevation points to generate triangular prism, and calculate the volume of a single triangular prism one by one. The accuracy and effectiveness of this algorithm are verified by experiments.

Orthogonal matching pursuit (OMP) has gained remarkable achievements in the domain of Sparse Subspace Clustering (SSC) for image clustering. However, current methods based on OMP improves the clustering accuracy by adding additional operations, which increase computational complexity. In this paper, a novel SSC algorithm with one-way selective orthogonal matching pursuit (SSC-OWSOMP) is proposed to improve the clustering accuracy without increasing the computational complexity in the SSC-OMP-based methods. In our SSC-OWSOMP, a one-way selective module is designed to avoid mutual selection among data points, which can enrich the information used for clustering without adding additional operations. Experimental results demonstrate that, with the SSC-OWSOMP, not only the clustering accuracy can be improved but also the time complexity be kept, also the SSC-OWSOMP is suitable for the data sets with high sample density.

This article aims to design a regression model to successfully predict clinical scores based on features provided by multimodal magnetic resonance (MR) images. We propose a multimodal MR predictive analysis pipeline with nonlinear support vector regression (SVR). Firstly, feature selection methods are elaborately chosen according to different feature characteristics for each modality. For features extracted from structural MR images and diffusion MR images, Support Vector Machine-Recursive Feature Elimination with Correlation Bias Reduction (SVMRFE+CBR) is applied to select predictive features and eliminate the high intra-correlations exist in the original features. For functional features, sparse coding (SC) considers multiple features’ combination for group predictiveness but cost less computation. After feature selection, the single radial bias function (RBF) kernels are calculated based on the modalities’ respectively selected features. A multi-RBF kernel is calculated by weighted-sum of single RBF kernels, and the kernel finally serves for the multimodal SVR model. Our proposed framework is tested on an online schizophrenia (SZ) dataset with 171 subjects from two study cohorts using 10-fold cross-validation (CV). The models are trained by the subjects’ clinicalrelated scores in Positive and Negative Syndrome Scale (PANSS) to be able to give the estimated scores as precisely as possible. The experimental results show that our proposed model can successfully predict clinical scores. Further comparative test results show that the proposed multimodal model can improve predictiveness compared with single modal ones, and our choice of feature selection methods plays an important role in the good performance.

To make the robot automatic welding systems more automatic and accurate, the 3D weld seam extraction has become a research hot-spot. In this paper, a seam extraction method for five types of straight-line seam based on point cloud obtained by three-dimensional reconstruction of binocular structured light is proposed. Firstly, the second derivative is used to calculate the inflection point for the rough extraction of seam characteristics. Secondly, according to the shape information of the welding workpiece, the center point of the seam is detected and the least square algorithm is used to fit the seam model. Finally, the 3D welding spot position and pose estimation are solved based on the established mathematical model. The experiments are conducted under five different situations, and the average extraction accuracy of the method can reach 0.19mm. The experimental results indicate that the proposed algorithm can efficiently locate five types of seam and generate the tracking path planning to guide robot manipulation.

User authentication using face biometrics has been becoming popular in many administrative activities (e.g., airport check-in, banking security systems, immigration managements, etc.). In this process, the facial image taken from an authorized document (e.g., identity (ID) card, passport, driving license) is compared to the live face that captured directly by stationary camera or human eyes to check whether they are matched as the same person. The problem is really challenging due to cross-domain difference including age changes, low document quality, and accessories. Recently, a few efforts have been made to build automatic face matching systems based on face recognition technologies. In this paper, we proposed a novel and efficient deep face matching system called SIRFace, which is derived from the RetinaFace detection model, InsightFace embedding model, and SVM classifier. SIRFace also contributed an efficient technique using Cosine similarity to mitigate the impacts of the cross-domain difference, which has not been used in any other related works. The experimental results on a dataset of 2,865 subjects with 9,824 images (i.e., ID and selfie photos) using our method achieved an error rate of 2.34%, which is potentially used for real life applications.