Analyzing player positioning and movement in badminton matches is crucial for understanding match outcomes. Traditional post-match analysis often requires significant time and manpower, leading to inefficiencies. This study introduces an automated method for analyzing the movement and positioning of badminton players using recorded video footage. By leveraging player movement and positioning data, our approach aims to predict the outcomes of points during a match. The workflow consists of four key components. First, players are detected in the video using the YOLOv7 network model. Next, player coordinates are extracted frame by frame. These coordinates are then transformed into a standardized perspective using homography transformation. Finally, the data is visualized as heatmaps using a Gaussian distribution and fed into a 3D convolutional neural network (CNN) for model training. The goal is to evaluate how a player's movement and positioning during specific rallies influence wins or losses. Our approach focuses on individual players, offering deeper insights into their strategies and adaptations compared to their opponents. This method enhances players' awareness of their movement and positioning, providing valuable tools for improving performance.

Deep learning has emerged as a pivotal technology across various domains, demonstrating remarkable performance. However, its susceptibility to security threats, particularly adversarial samples, poses a significant concern. These samples can manipulate inputs slightly, deceiving models such as those used in image or speech recognition and potentially leading to incorrect predictions. This undermines the reliability of deep learning in critical applications. In this paper, we propose an effective method utilizing Autoencoder to detect and intercept audio adversarial attacks before they are input to speech recognition models. The proposed approach first uses clean audio data to train the Autoencoder model, then isolates adversarial samples from clean ones by comparing their features against normal features encoded in the Autoencoder. Our method does not require prior knowledge about the target automatic speech recognition model or attack methods. Experimental results show that it can effectively discriminate adversarial attack samples from clean ones with high accuracy.

The use of high-performance computing devices to detect anomalies in public spaces is expensive and impractical. Innovative improvements, especially hardware acceleration in embedded systems, are significantly impacting the deployment of deep learning in the real world. In terms of edge computing, the Jetson Platform by NVIDIA is one of the leading platforms that provide optimal energy efficiency, required accuracy, and throughput when running deep learning algorithms. Therefore, we offer the use of anomaly detection on edge devices to perform intelligent monitoring. Deploying an AI-powered anomaly detector at the edge allows initial analysis to be executed at the captured image’s location, reducing communication overhead and enabling a fast security response. On the edge device side, we prepared an I3D (Inflated 3D CNN) model that can detect anomaly events. To be specific, the edge device used NVIDIA® Jetson Orin NanoTM 8GB. The model output the probability of anomaly events, along with the confidence percentage and estimated occurred frames. Finally, for real-time analysis, we built a GUI-friendly web interface through flask framework in order to demonstrate the products to the user. To verify the quantitative results, we evaluated our system using Area under Curve (AUC) and accuracy. The resulting experiments showed an average prediction AUC of 81% for the public UCF Crime dataset. The demo of our web-app anomaly detector can be found here https://youtu.be/X8iVyxkzsxA.

Practical deployment of convolutional neural network (CNN) and cryptography algorithm on constrained devices are challenging due to the huge computation and memory requirement. Developing separate hardware accelerator for AI and cryptography incur large area consumption, which is not desirable in many applications. This article proposes a viable solution to this issue by expressing the CNN and cryptography as generic-matrix-multiplication (GEMM) operations and map them to the same accelerator for reduced hardware consumption. A novel systolic tensor array (STA) design was proposed to reduce the data movement, effectively reducing the operand registers by$2\times $. Two novel techniques, input layer extension and polynomial factorization, are proposed to mitigate the under-utilization issue found in existing STA architecture. Additionally, the tensor processing element (TPE) is fused using DSP unit to reduce the look-up table (LUT) and flip-flops (FFs) consumption for implementing multipliers. On top of that, a novel memory efficient factorization technique is proposed to allow computation of polynomial convolution on the same STA. Experimental results show that Cryptensor achieved 21.6% better throughput for VGG-16 implementation on XC7Z020 FPGA; up to$8.40\times $better-energy efficiency compared to existing ResNet-18 implementation on XC7Z045 FPGA. Cryptensor can also flexibly support multiple security levels in NTRU scheme, with no additional hardware. The proposed hardware unifies the computation of two different domains that are critical for IoT applications, which greatly reduces the hardware consumption on edge nodes.

In recent years, ensemble methods have shown sterling performance and gained popularity in visual tasks. However, the performance of an ensemble is limited by the paucity of diversity among the models. Thus, to enrich the diversity of the ensemble, we present the distillation approach-learning from experts (LFEs). Such method involves a novel knowledge distillation (KD) method that we present, specific expert learning (SEL), which can reduce class selectivity and improve the performance on specific weaker classes and overall accuracy. Through SEL, models can acquire different knowledge from distinct networks with various areas of expertise, and a highly diverse ensemble can be obtained afterward. Our experimental results demonstrate that, on CIFAR-10, the accuracy of the ResNet-32 increases 0.91% with SEL, and that the ensemble trained by SEL increases accuracy by 1.13%. Compared to state-of-the-art approaches, for example, DML only improves accuracy by 0.3% and 1.02% on single ResNet-32 and the ensemble, respectively. Furthermore, our proposed architecture also can be applied to ensemble distillation (ED), which applies KD on the ensemble model. In conclusion, our experimental results show that our proposed SEL not only improves the accuracy of a single classifier but also boosts the diversity of the ensemble model.

Existing Siamese network-based trackers rely on stable appearance features extracted from the target object. However, such features might not be available during tracking due to non-digit appearance deformation and severe occlusion, which result in drift problems. In this paper, we propose a background-augmented tracking network that incorporates background information surrounding the target to make up for missing or deformed target features during the matching process. A novel Background Relation Network (BRNet) is designed to effectively encode and match the background information surrounding candidate objects in the search region to help identify the correct target, and thus avoid tracking error. BRNet can complement the base tracker when reliable target features cannot be obtained. Experiments on the OTB, VOT, and UAV123 datasets demonstrate that the proposed method achieves superior performance over existing state-of-the-art methods while maintaining reasonable real-time speed.

Steel surface defect detection represents a challenging task in real-world practical object detection. Based on our observations, there are two critical problems which create this challenge: the tiny size, and vagueness of the defects. To solve these problems, this study a proposes a deep learning-based defect detection system that uses automatic dual transformation in the end-to-end network. First, the original training images in RGB are transformed into the HSV color model to re-arrange the difference in color distribution. Second, the feature maps are upsampled using bilinear interpolation to maintain the smaller resolution. The latest and state-of-the-art object detection model, High-Resolution Network (HRNet) is utilized in this system, with initial transformation performed via data augmentation. Afterward, the output of the backbone stage is applied to the second transformation. According to the experimental results, the proposed approach increases the accuracy of the detection of class 1 Severstal steel surface defects by 3.6% versus the baseline.

This paper presents the first experimental demonstration of a ternary memristor-CMOS logic family. We systematically design, simulate and experimentally verify the primitive logic functions: the ternary AND, OR and NOT gates. These are then used to build combinational ternary NAND, NOR, XOR and XNOR gates, as well as data handling ternary MAX and MIN gates. Our simulations are performed using a 50-nm process which are verified with in-house fabricated indium-tin-oxide memristors, optimized for fast switching, high transconductance, and low current leakage. We obtain close to an order of magnitude improvement in data density over conventional CMOS logic, and a reduction of switching speed by a factor of 13 over prior state-of-the-art ternary memristor results. We anticipate extensions of this work can realize practical implementation where high data density is of critical importance.

Various weather conditions, such as rain, haze, or snow, can degrade visual quality in images/videos, which may significantly degrade the performance of related applications. In this paper, a novel framework based on sequential dual attention deep network is proposed for removing rain streaks (deraining) in a single image, called by SSDRNet (Sequential dual attentionbased Single image DeRaining deep Network). Since the inherent correlation among rain steaks within an image should be stronger than that between the rain streaks and the background (non-rain) pixels, a two-stage learning strategy is implemented to better capture the distribution of rain streaks within a rainy image. The two-stage deep neural network primarily involves three blocks: residual dense blocks (RDBs), sequential dual attention blocks (SDABs), and multi-scale feature aggregation modules (MAMs), which are all delicately and specifically designed for rain removal. The two-stage strategy successfully learns very fine details of the rain steaks of the image and then clearly removes them. Extensive experimental results have shown that the proposed deep framework achieves the best performance on qualitative and quantitative metrics compared with state-of-the-art methods. The corresponding code and the trained model of the proposed SSDRNet have been available online at https://github.com/fityanul/SDAN-for-Rain-Removal.

Background subtraction is an example of a moving object detection technique that uses machine vision systems. Conventional moving object detection methods need complicated thresholds for background modeling to address changes in illumination. This paper proposes a novel background modeling approach without thresholding based on a bit-planes method, which fully utilizes color characteristics through spatial and temporal-based improvement. The proposed idea is effective and efficiently solving for shadow disturbance and brightness changes. We evaluate our proposed method using several challenging indoor and outdoor sequences from the CDNET 2014 dataset. The experiments show that the proposed idea typically achieves a higher rate of detection accuracy than those of the current state-of-the-art approaches.

To date, deep learning has been widely introduced in many fields, including object detection, medical imaging, and automation. One important application that uses deep learning based object detection is detecting defects by simply evaluating the image of an object. Such systems must be accurate, robust and efficient. Single-stage and two-stage object detection are two main approaches used in defect detection systems. A revised version of the popular object detection method called single shot multi-box detector (SSD) and the residual network (ResNet) offer a two-stage method to automatically detect defects with higher precision but has shown room for improvement with regard to speed performance. Therefore, in this paper, we propose a fully automatic pipeline for detecting defects, especially on steel surfaces. A novel transformation of the two-stage defect detection process into a more efficient single-stage detection process was introduced by utilizing a state-of-the-art method called RetinaNet. In addition, we leverage a feature pyramid network (FPN) and focal loss optimization to solve the small object detection problem and to deal with imbalanced background-foreground samples issue, respectively. Experimental results show that the proposed single-stage pipeline can achieve high accuracy and faster speed in steel surface defect detection.

The issue of image semantic segmentation is renowned within computer vision and artificial intelligence. The ground truth in image segmentation is hard to produce and is time- and resource-intensive. Recent research on real-time image semantic segmentation based on deep learning has reduced image resolution through pooling operations, resulting in detail loss in the scene. In order to generate high-quality annotated data, in this paper, we propose a joint coarse-and-fine (JCF) architecture that can repair fragment defects based on a coarse module, and also produce fine details based on a fine module. The experiments show promising results compared to state-of-the-art methods.

This paper proposes a method that utilizes two stages of deep learning networks for segmentation and recognition of spinal vertebrae. The first stage involves a y-shaped model that learns the basic shape of vertebrae to segment the spine. The second stage determines the order of each vertebra in the spine. This two-stage method provides more reliable spine segmentation results.

In this paper, we propose a novel method for extracting brain region from T1-weighted magnetic resonance imaging. Our proposed method is based on finding a seed point which is located on brain tissue and then perform region growing. In particular, we firstly analyze intensity distributions of different cerebral tissues (including brain, cerebrospinal fluid, scalp and marrow) at middle height. We also measure the distance from image center to brain edge at the same height of brain scan. In the experimental validation, we have conducted experiment on the OASIS dataset and perform comparison with the state-of-the-art methods. Our propose method achieves the best performance on the OASIS dataset.

This paper proposes a speckle image recognition method using data mining techniques to ensure speckle identification system feasible for authentication. This is an interdisciplinary method that integrates the researches of optics, data mining, and image processing. Because objects have unique but imperfect surfaces, their laser speckle is capable of providing suitable identifiable features for authentication. In our method, matching points among speckle images acquired from one plastic card are extracted by scale-invariant feature transform (SIFT). The spatial relations among the matching points are then transformed to 9 direction lower triangular (9DLT) representations. Then, the Apriori algorithm mines frequent patterns so a useful association rule is obtained as the feature to identify the similarity between each of the speckle images for the purpose of authenticity verification. The proposed method is especially robust in the cases of card displacement and luminance change resulted from laser attenuation. Experimental results show that the proposed method has promising results and outperforms existing methods in identification accuracy.

This paper aims to improve the intra coding performance on H.264 and high efficient video coding (HEVC). A new intra prediction approach is proposed based on synthesizing two neighboring predictors. These two predictors are selected from different prediction directions and then the predicted block is generated by combining the two predictors using different weights. The weights for the predictors do not need to be saved, so the bit rate can be greatly reduced. The main contributions of this paper are proposing the following: 1) a highly efficient way and high compact compression to select the proper predictors; 2) a weight estimation method for the predictors; and 3) a reversible weight restoration method for the predictors to save the bit rate in the decoding phase. Experimental results show that the proposed method outperforms H.264/AVC and HEVC in intra prediction by 11.79% and 3.6%, respectively, in bitrate reduction.

This paper presents a detection-based method to subtract abandoned object from a surveillance scene. Unlike tracking-based approaches that are commonly complicated and unreliable on a crowded scene, the proposed method employs background (BG) modelling and focus only on immobile objects. The main contribution of our work is to build abandoned object detection system which is robust and can resist interference (shadow, illumination changes and occlusion). In addition, we introduce the MRF model and shadow removal to our system. MRF is a promising way to model neighbours' information when labeling the pixel that is either set to background or abandoned object. It represents the correlation and dependency in a pixel and its neighbours. By incorporating the MRF model, as shown in the experimental part, our method can efficiently reduce the false alarm. To evaluate the system's robustness, several dataset including CAVIAR datasets and outdoor test cases are both tested in our experiments.

Nowadays, Authentication mechanism is widely applied to digital images. In this paper, we propose a self authentication mechanism with recovery ability for digital images, which is employed to protect the image during transmission over the internet. In the scheme, the image is to authenticate whether the area is modified by comparing the generated authentication code and hidden authentication code together. In our experimental result, we show the positive result for the feasibility of the proposed scheme.

Compressive sensing (CS) has recently attracted much attention due to its unique feature of directly and simultaneously acquiring compressed and encrypted data based on their sparse or compressible properties. To securely transmit compressively sensed multimedia data over networks, it is required to support transcoder to securely convert compressed multimedia into several different types for diverse receivers. In this paper, a secure transcoding scheme for compressive multimedia sensing is proposed. We focus on securely converting compressively sensed multimedia data (not data compressed via standard codec) with a certain number of measurements into other different numbers of measurements without resorting to reconstruct the original data. We show that the security can be achieved via transforming multimedia re-sensing process into another secure domain at the transcoder. We also show that the computational security can be achieved while transmitting compressively sensed data between the sender (or each receiver) and the transcoder over networks.

Reversible steganography becomes a popular hiding problem in recent years. A reversible steganographicmethod can reconstruct an original image without loss from the stego-image after extracting the embedded data. Unlike traditional reversible methods in which data is hidden in uncompressed images, we propose a reversible scheme for BTC (block truncation coding)-compressed images. The secret data embedded in the compressed image are more difficult to detect than in the uncompressed image. To achieve reversibility, the properties of side matching and BTC-compressed code are applied. The experimental results show that the proposed method is feasible for BTC-compressed images and can embed one more bit in each BTC-encoded block.

Assessment of image similarity is fundamentally important to numerous multimedia applications. The goal of similarity assessment is to automatically assess the similarities among images in a perceptually consistent manner. In this paper, we interpret the image similarity assessment problem as an information fidelity problem. More specifically, we propose a feature-based approach to quantify the information that is present in a reference image and how much of this information can be extracted from a test image to assess the similarity between the two images. Here, we extract the feature points and their descriptors from an image, followed by learning the dictionary/basis for the descriptors in order to interpret the information present in this image. Then, we formulate the problem of the image similarity assessment in terms of sparse representation. To evaluate the applicability of the proposed feature-based sparse representation for image similarity assessment (FSRISA) technique, we apply FSRISA to three popular applications, namely, image copy detection, retrieval, and recognition by properly formulating them to sparse representation problems. Promising results have been obtained through simulations conducted on several public datasets, including the Stirmark benchmark, Corel-1000, COIL-20, COIL-100, and Caltech-101 datasets.

The purpose of data hiding with reversibility property is to recover the original cover media after extracting the hidden data from the stegomedia. In this paper, we propose a reversible data-hiding scheme for embedding secret data in VQ-compressed codes based on the declustering strategy and the similarity property of adjacent areas in a natural image. Two declustering methods are proposed using the minimum-spanning-tree and the short-spanning-path algorithms, respectively. The proposed data-hiding method can achieve the benefits including easy implementation, completely recovering the original compressed codes, and high efficiency of embedding and extraction processes. The experimental results also show that the proposed method has more flexible and higher embedding capacity than other schemes.

Gel electrophoresis (GE) is an important tool in genomic analysis. GE results are presented using images. Each image contains several vertical lanes. Each lane consists of several horizontal bands. Two lanes are identical if the relative positions of the bands are the same. We present a computer method designed to compare the lanes and identify identical lanes. This method, developed using many image-processing techniques, is applied to segment the lanes and bands in GE images. The lanes are then converted into "position vectors" that describe the positions of the bands. Comparing lanes becomes equivalent to comparing the position vectors. This method can accurately identify identical lanes, helping biologists to identify the identical lanes from many lanes with much less effort.

The reversible steganographic method allows an original image to be completely reconstructed from the stegoimage after the extraction of the embedded data. The traditional reversible embedding schemes are not suitable for images compressed using vector quantization (VQ) and usually require the use of the location map for reversibility. In this paper, we propose a reversible embedding scheme for VQ-compressed images that is based on side matching and relocation. The new method achieves reversibility without using the location map. The experimental results show that the proposed method is practical for VQ-compressed images and provides high image quality and embedding capacity

Hashing schemes are widely used to improve the performance of data mining association rules, as in the DHP algorithm that utilizes the hash table in identifying the validity of candidate itemsets according to the number of the table's bucket accesses. However, since the hash table used in DHP is plagued by the collision problem, the process of generating large itemsets at each level requires two database scans, which leads to poor performance. In this paper we propose perfect hashing schemes to avoid collisions in the hash table. The main idea is to employ a refined encoding scheme, which transforms large itemsets into large 2-itemsets and thereby makes the application of perfect hashing feasible. Our experimental results demonstrate that the new method is also efficient (about three times faster than DHP), and scalable when the database size increases. We also propose another variant of the perfect hash scheme with reduced memory requirements. The properties and performances of several perfect hashing schemes are also investigated and compared.

The apical four-chamber view echocardiogram collected by a transthoracic transducer can be used to evaluate the left ventricle volume. In the diastole, the left ventricle and left atrium become one chamber. In this case, the left ventricle and left atrium need to be separated using a "mitral annular line" so the volume of the left ventricle can be estimated. In this paper, a nearly automatic method for identifying the mitral annular lines from two-dimensional (2-D) + one-dimensional (1-D) precordial four-chamber view echocardiogram is presented. This method employs the optical flow technique and graph-search approach. The mitral annular line sequence is found by finding the shortest path in a weighted directed graph. The vertices in the graph are candidates for the mitral annular lines. The weights on the directed edges are determined using the optical flow technique. The proposed method requires only a physician to provide a point that is always in the left ventricular chamber. Experimental results show that the average error for the left ventricle volume obtained based on the computed mitral annular lines is 3%.