This PDF file contains the front matter associated with SPIE Proceedings Volume 13169, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Sports analytics is a field of study that utilizes camera and sensor data to monitor the athlete’s performance and health to optimize the player's strategy and increase the success rate. Coaches rely on analytics to scout opponents and optimize play calls in gameplay. With the advancement in artificial intelligence, accessible and in-depth data collection has been enabled. The well-grounded technique for performance evaluation in sports analytics is Human Pose Estimation (HPE). Our focus is on real-time action recognition in combat sports like boxing. Existing state-of-the-art deep learning models are heavily parameterized, so can’t be used in real-time in any low-end devices. Apart from this, fine-grained classification in highly dynamic activities in sports are typically performed using sensors only. Our proposed Machine Learning based pipeline provides real-time fine-grained solution for 14 boxing punch types of classification using RGB video only. Our approach includes the implementation of three novel and generalized motion dynamics features that encode velocity as well as acceleration of the pose sequences., 1) Unified-Axis Angular Encoding (UAE), 2) 2D Motion Dynamics Descriptors (2DMDD), 3) Fifth-order Angular Encoding (FAE). We employed classical machine learning algorithms I.e., Support Vector Machine (SVM), Random Forest (RF), and K Nearest Neighbours (KNN) to make a lightweight model and test it on YouTube videos. The average accuracies of pipeline using the proposed features are found to be 55%, 92% and 84% for UAE, 2DMDD, and FAE respectively. Using KNN, we have achieved 99% accuracy on 10-fold cross-validation by using FAE features.

The task of recognizing license plates from photographs or videos is extremely relevant nowadays. The development of transport links between cities and countries has led to the need to recognize vehicles from different regions. However, license numbers have their own characteristics for each country, such as differences in symbols, their position and font, which complicate recognition. A universal automatic recognition system must cope with these features. Our previous studies have proven the effectiveness of a model trained specifically for a target country, which for effective recognition requires an additional decision for choosing a country or region. The data used is a dataset of images of license plates mainly from countries that previously belonged to the CIS. We propose a vehicle country recognition model from license plate image based on siamese network model with triplet loss function and negative sampling technique. This model is significantly superior to a solution based on a convolutional network in terms of recognition accuracy, achieving 0.9651 value, as well as time spent on training.

When applying the tubing ultrasonic testing technology to evaluate the contact condition of the metal-to-metal sealing surface of premium connection, it is necessary to judge the ultrasonic reflection amplitude image of the manual contact interface to observe whether there is sealing defect. At present, the images of phased array ultrasonic testing results usually require professionals to rely on technical knowledge to judge, the analysis has low efficiency and strong evaluation subjectivity. Therefore, there is an urgent need for an intelligent method to identify the location, range and type of sealing defects in ultrasonic images accurately and efficiently, so as assisting or replacing manual operations. Aiming at the problems of heavy reliance on data quality and inflexible segmentation effect during the process of identifying the sealing surface region using the original Mask R-CNN network, the approach described in this paper enhances the model by incorporating the Segment Anything Model(SAM) and employs prompts to guide the object detection model in generating masks that fulfill various criteria. Experiments show that the method adopted in this paper can not only correctly identify the sealing surface location, but also, compared with the original Mask R-CNN network model, it can output a segmentation mask that meets the demand according to the prompts of different segmentation criteria, and the obtained sealing surface region is closer to the theoretical segmentation region.

A major e-commerce platform like Shopee offers thousands of macro categories for listing various products. Product price comparison shopping is a challenge for the e-commerce platform. Image classification helps customers accurately search for products from a large number of images, and it also helps them find the best price deals. To develop an effective image classification model for this use case, we utilized Convolutional Neural Networks (CNN) with transfer learning in our study. We sourced our image dataset from Kaggle (Shopee - Price Match Guarantee) and selected 645 e-commerce images from the top 20 classes, each containing the highest number of images. We chose MobileNetV2, EfficientNetB3, and InceptionV3 to compare the efficiency of pre-trained models. Our experimental results revealed that EfficientNetB3 and MobileNetV2 outperformed InceptionV3 in the validation set. Moreover, the time required to train all layers of EfficientNetB3 was approximately 8 times less than that of MobileNetV2.

Extracting 3D features from point cloud data is the central component for the automatic repair system of ancient ceramic fragments developed by us. This paper presents a local 3D point clouds descriptor for individual point on various surfaces based on the distances among the point’s neighborhood, which providing feature extracting service to the system. The descriptor use the distance between the point with its k-neighborhoods, which are chosen by a adjustable search radius, to identify the surface type of the ceramic fragments. The descriptor then utilize and maintain the existed local distance information from some selected point to generate feature vectors for accurate point matching which will ease the restoration work. To validate the algorithm, experiments on real data are performed and compared with other similar descriptors. The results show that our descriptor can provide reliable identification performance and also do not introduce any local information loss. The descriptor are also a non-geometric deformation operation, so it has the potential to be a commonly used intermediate algorithm in many point cloud applications.

State of the art methods for anomaly localisation in product images take a patch based approach that models an anomaly patch in an image as an outlier to a distribution of normal image patches. These approaches require the availability of sufficient normal and sometimes even abnormal product images. In this work we present a zero/few-shot anomaly localisation method, where, given an image and a set of textual prompts describing the possible anomalies, the solution can localise the defects in the image in the form of an anomaly map. We present a few approaches for identifying prompts optimised for improving localisation performance. We benchmark our proposed method with a recently proposed WinCLIP method and shows promising results using metrics like F1-max, AUROC, AUPR etc.

Deep neural networks (DNNs) have found wide applications in various domains. However, recent research has revealed the vulnerability of DNNs to adversarial examples. Existing adversarial attack methods can easily mislead the models. Moreover, while traditional denoising techniques effectively deter certain attacks, they have limitations. To address this, We propose Text-guided Diffusion Model Purification (TGDP), an adversarial perturbation purification method based on diffusion models.This method preprocesses input images to purify adversarial perturbations. TGDP employs Protogen x3.4 (Photorealism) Official Release as a diffusion model for conditional image generation. During the generation process, text information is incorporated to enhance control over the diffusion model instead of relying entirely on the internal randomness of the model. By iteratively adding Gaussian noise to disrupt the adversarial examples and reversing the noise addition process to restore the image afterward, we can completely eliminate carefully crafted perturbations, achieving the purification objective. Extensive experiments on the ImageNet dataset against common adversarial attacks demonstrate that TGDP outperforms other defense methods applied to the ImageNet dataset.

Deep learning-based machine reasoning and visual question answering models achieve a near-human performance on their respective datasets; however, their performance dramatically drops under domain shift suggesting that models fail to generalize to the level of human-like reasoning. In this paper we present a new CLEVR-like dataset consisting of images-question pairs to evaluate the visual reasoning capability of deep models. The objects in the images are arranged in a way that the first half of the question is ambiguous and multiple answers seem to be correct up to this point; however, the second half of the question clarifies the situation and makes the whole visual question-answering (VQA) task unambiguous, and a unique answer can be reported. Therefore, deep models during their reasoning process need to handle ambiguousness in their neurons. They can handle this either via graph (or tree) traversing in the search space with using back-tracking technique or via refining a candidate set of possibly correct answers by iteratively eliminating incorrect ones upon some reasoning calculations. We call this data-set CLEVR with Back-Tracking Database, CLEVR-BT-DB. It consists of 2,500 images and 10,000 questions in the same format as the standard CLEVR, and it is available at https://huggingface.co/datasets/Aborevsky01/CLEVR-BT-DB site. The code to generate additional data is available at https://github.com/AFigaro/CLEVR_BT_DB site. We tested MDETR method, a recent deep model for VQA from Meta Research, it achieved an accuracy of 99.7 % on the Standard CLEVR dataset; however, it achieves an accuracy of 28.01 % on our CLEVR-BT-DB dataset.

Virtual YouTubers (VTubers) offer significant growth potential as a new type of content creator. However, the financial aspect poses a hurdle for aspiring VTubers. This article proposes a cost-effective solution by utilizing generative models to create full-body images of 2D VTuber characters. Notably, studies have achieved remarkable results using Generative Adversarial Networks (GANs), including Deep Convolutional GAN (DCGAN) and Style-based GAN 2 (StyleGAN2), for anime face generation. To address the lack of image synthesis systems for full-body anime characters, experiments were conducted with DCGAN and StyleGAN2 on the Danbooru dataset. The results demonstrate that StyleGAN2 models outperform DCGAN, yielding superior Fréchet Inception Distance (FID) scores of 25.06, 28.03, and 24.52, compared to DCGAN's 159.21 FID score. This research contributes to reducing the cost associated with becoming a VTuber and offers insights into generating 2D full-body anime characters for VTubers.

Machine vision and AI-based techniques hold significant promise for automating the analysis of extensive sports video datasets to uncover longitudinal performance trends. This study introduces an innovative pipeline tailored for the analysis of lengthy top-view boxing training session videos, recorded in uncontrolled natural settings and featuring multiple athletes. Our primary focus lies in capturing the movement patterns of boxers within the ring. Within this research, we present Histotracker, an intelligent rule-based tracking module that connects segmented objects across frames using cosine similarity. Distinguishing itself from existing trackers, this module possesses the unique ability to backtrack and correlate frames with the highest association to maintain continuous tracking information. When compared to various standard approaches, our proposed Histotracker demonstrates remarkable results, boasting a MOTA score of 0.95 In approximately 69% of the total bout videos, there were no occurrences of Identity Switching or Identity Update. These findings hold immense promise for advancing the application of automated video analytics in diverse combat sports.

The fundamental data structure of our proposed scheme is a pixel pair. The pixel values are adjusted through simple yet well-designed equations depending on different conditions. The computation method provides an easy recovery of the original values for the pixel pair as well. For some applications, it’s important to extract the hiding data as well as to recover the original media. In order to apply to a wider range of applications, more and more reversible data hiding (RDH) schemes are introduced or enhanced in the recent decades. Therefore, the two goals of the novel scheme proposed are to increase embedding capacity and to recover original media easily. For vector quantization (VQ) compression, it’s a well-known and efficient method to compress images. A well-trained codebook is crucial to maintain the visual quality of recovered images as good as possible. An index table is generated based on the codebook to full-fill the purpose of compression. Most of studies utilize index tables as the embedding carriers for VQ or VQ-variant compression methods. There are a few state-of-the-art schemes make use of codebooks to embed secret and our proposed scheme extends the study on the codebook embedding. The codewords in a codebook are divided into pixel pairs and secret bits are embedded by adjusting the pixel values of these split pairs to form a stego codebook. During extraction, secret bits are extracted first out from the stego codebook and a recovered codebook is obtained before recovering VQ compressed images.

Credit card fraud has been increasing with the rise of cashless payments, making it difficult to identify fraudulent transactions among thousands of normal ones. Machine learning algorithms can help address this issue by classifying transactions as either fraud or non-fraud in a dataset. For this project, we utilized a highly imbalanced dataset from Kaggle containing European cardholder data. While the dataset is mostly clean, we needed to balance it for training and testing purposes through undersampling. We developed six classification models to differentiate between fraudulent and non-fraudulent transactions and evaluated their performance based on accuracy and F1 score. The KNN model outperforms the other models for the dataset we used in our experiment.

As edge computing continues to evolve, addressing the inherent challenges of low stability in edge devices becomes imperative for optimizing routing efficiency. This paper introduces a novel approach to tackle this issue through the application of Reinforcement Learning (RL) for routing optimization in dynamic edge computing environments. The core problem revolves around the need for an algorithm capable of selecting edge devices strategically to minimize latency. To validate our proposed solution, we conducted experiments utilizing Oliver30 and ry48p datasets. Comparative analysis against traditional heuristic algorithms, such as Simulated Annealing (SA) and Ant Colony Optimization (ACO), demonstrates the superior performance of our RL-based model. The results highlight the effectiveness of leveraging advanced machine learning techniques to enhance routing efficiency in challenging edge computing scenarios.

The major mainstream method for human ocean exploration is to use sound waves to measure water depth. The paper investigates the application of multi-beam bathymetric surveying technology in ocean depth measurement. Traditional single-beam sonar technology faces challenges in survey line design due to dense track lines and lack of data. To address the challenges of multi-beam survey line design below the sea surface, this paper establishes a mathematical model for calculating coverage width and overlap rate through geometric analysis of the beam’s coverage area on the sloping surface. Considering the non-isobath movement of the beam center along the sloping surface, the paper discovered the equivalent relationship between the cross-sectional angle and the horizontal plane. Therefore, a mathematical model for the depth coverage width of multi-beam survey lines was established based on the slope angle and the lateral angle. Based on previous research and relevant references, this paper designed a complete survey line scheme that can fully cover the ideal test area while maintaining a low overlap rate, and verified its rationality.