ABSTRACT In Bangladesh, where a burgeoning population strains limited road networks, wrong‐way driving (WWD) exacerbates traffic congestion and elevates accident risks, posing a critical public safety challenge. We present a pioneering real‐time WWD detection system that integrates cutting‐edge image processing and deep learning to transform traffic monitoring. Our methodology combines the Hough line transform for precise road boundary extraction with the state‐of‐the‐art YOLOv5 algorithm for high‐fidelity, real‐time vehicle detection. Vehicle direction is inferred from single‐frame orientation (distinguishing front/rear views) without multi‐frame tracking, ensuring lightweight operation and scalable deployment. Tested on a novel, region‐specific dataset from Chattogram, the system achieves an impressive 96% vehicle detection accuracy (mAP@0.5) during training and validation. Real‐world evaluations on 20 diverse test videos yield an overall 60.6% recall for wrong‐way vehicles, with a representative subset of 10 videos reaching 75% recall (and 41% precision). This research advances computer vision applications in traffic safety, providing an effective and deployable framework for reducing WWD incidents in high congestion urban regions and paving the way for enhanced intelligent transportation systems.

ABSTRACT Cardiovascular diseases (CVDs) remain a leading global health challenge, emphasising the need for advanced diagnostic systems that enable accurate detection and classification. This study presents a novel deep learning framework based on a dense and inverted bottleneck residual mechanism, integrating two custom convolutional neural network architectures: an 8‐block Dense model and a 9‐block Inverted Bottleneck Layered model. These models are designed to extract multiscale features from cardiac MRI scans for precise classification across five CVD categories. We enhance image quality using Contrast‐Limited Adaptive Histogram Equalisation (CLAHE) in the preprocessing pipeline and apply data augmentation to address class imbalance and promote model generalisation. Atom Search optimisation (ASO) is employed to reduce feature dimensionality while retaining critical information, and a serial feature fusion strategy is used to integrate the optimised feature vectors. Classification is performed using several neural network variants (Narrow NN, Medium NN, Wide NN, and Bi‐Layered NN). The proposed method is evaluated on both the Automated Cardiac Diagnosis Challenge (ACDC) and Sunnybrook Cardiac Data (SCD) datasets, achieving superior accuracies of 97.3% and 96.8%, respectively, compared to state‐of‐the‐art techniques, demonstrating its effectiveness for cardiovascular disease diagnosis.

ABSTRACT Depth images or point clouds offer true three‐dimensional insights into scene geometry, making depth perception essential for downstream tasks in computer vision. However, current commercial depth sensors often produce dense estimations with lower accuracy, especially on specular surfaces, leading to noisy and incomplete data. To address this challenge, we propose a novel framework based on latent diffusion models conditioned on RGBD images and semantic labels for depth completion and inpainting, effectively restoring depth values for both visible and occluded parts of specular objects. We enhance geometric guidance by designing various visual descriptors as conditions and introduce channel and spatial attention mechanisms in the conditional encoder to improve multi‐modal feature fusion. Using the MP6D dataset, we render complete and dense depth images for benchmarking, enabling a comprehensive evaluation of our method against existing approaches. Extensive experiments demonstrate that our model outperforms previous methods, significantly improving the performance of downstream tasks by incorporating the predicted depth maps restored by our model.

ABSTRACT Domain generalization poses significant challenges, particularly as models must generalize effectively to unseen target domains after training on multiple source domains. Traditional approaches typically aim to minimize domain discrepancies; however, they often fall short when handling complex data variations and class imbalance. In this paper, we propose an innovative model, the self‐supervised learning multi‐classifier ensemble (SSL‐MCE), to address these limitations. SSL‐MCE integrates self‐supervised learning within a dynamic multi‐classifier ensemble framework, leveraging ResNet as a shared feature extraction backbone. By combining four distinct classifiers, it captures diverse and complementary features, thereby enhancing adaptability to new domains. A self‐supervised rotation prediction task enables SSL‐MCE to focus on intrinsic data structures rather than domain‐specific details, learning robust domain‐invariant features. To mitigate class imbalance, we incorporate adaptive focal attention loss (AFAL), which dynamically emphasizes challenging and rare instances, ensuring improved accuracy on difficult samples. Furthermore, SSL‐MCE adopts a dynamic loss‐based weighting scheme to prioritize more reliable classifiers in the final prediction. Extensive experiments conducted on public benchmark datasets, including PACS and DomainNet, indicate that SSL‐MCE outperforms state‐of‐the‐art methods, achieving superior generalization and resource efficiency through its streamlined ensemble framework.

ABSTRACT This paper presents a face quality assessment approach that selects the highest‐quality face image using a two‐stage process from video streaming. In high‐traffic environments, traditional face recognition methods can cause crowd congestion, emphasizing the need for unconscious face recognition, which requires no active cooperation from individuals. Due to the nature of unconscious face recognition, it is necessary to capture high‐quality face images. In this paper, the FSA‐Net head pose estimation network is enhanced to FSA‐Shared_Nadam by replacing the Adam optimizer with Nadam and improving stage fusion. In the first stage, FSA‐Shared_Nadam estimates head pose angles, MediaPipe detects facial landmarks to calculate eye distance and aspect ratios, and sharpness is calculated using the Laplacian operator. Images are considered valid if they meet the criteria. A model trains a face quality scoring formula, learning how different head pose angles affect face recognition accuracy. In the second stage, face images are clustered, and the formula is applied to select the highest‐scoring face within each cluster. The approach was tested across multiple datasets, and a simulated security checkpoint scenario was created for practical testing. The results demonstrate the effectiveness of the FSA‐Shared_Nadam head pose estimation algorithm and the proposed face quality assessment approach.

ABSTRACT In this paper, we propose to use weighted ‐norm for approximating the solution of general ‐norm regularization problem for recovering hyperspectral images (HSI) corrupted by a mixture of Gaussian‐impulse noise. As a special case of , we design an optimization framework to accommodate the combined effect of different noise sources. An initial impulse noise pre‐detection phase decouples the raw noisy HSI data into impulse and Gaussian corrupted pixels. Gaussian corrupted pixels are handled by data‐fidelity term in while impulse corrupted pixels possess more Laplacian like behavior; modeled using . Solutions of problems involving in data fidelity and regularization terms complicate the optimization process but are less sensitive to the outlier pixels. On the other hand, the least square solutions for the data misfit are computationally efficient but generates solutions which are quite sensitive to the outlier pixels; which is the characteristic of impulse corrupted pixels. Therefore, in this paper, we decouple the set of pixels into two distinct parts; handled using two separate data fidelity terms. Total variation (TV) is used on the Casorati matrix representation of the input data to exploit similarity along both spatial and spectral dimensions. The resulting optimization problem is reformulated as iteratively reweighted least square for the general ‐norm problem for for data fidelity terms and for the TV regularization term. Experiments conducted over synthetically corrupted HSI data and images obtained from real HSI sensors confirm the suitability of the proposed weighted norm optimization framework (WNOF) over a wide range of degradation scenarios.

Abstract The purpose of image steganalysis is to detect whether the transmitted images in network communication contain secret messages. Current image steganalysis networks still have some problems such as inappropriate feature selection and easy overfitting. Therefore, this paper proposed a new spatial image steganalysis method based on convolutional neural networks. To extract richer features while reducing useless parameters in the network, this paper introduced the Im SRM filtering kernel into the image preprocessing module. To extract effective steganography noise from images, this paper combined depthwise separable convolution and residual networks for the first time and introduces them into the steganography noise extraction module. In addition, to focus network attention on the image regions where steganography information exists, this paper integrated the coordinate attention mechanism. This module will make the network pay attention to the overall structure and local details of the image during network training, improving the network's recognition ability for steganography information. Finally, the extracted steganography features are classified through a classification module. This paper conducted a series of experiments on the BOSSBase 1.01 and BOWS2 datasets. The improvement in detection accuracy is between 1.2% and 18.2% compared to classic and recent steganalysis networks.

ABSTRACT Unsupervised anomaly localization aims to identify anomalous regions that deviate from normal sample patterns. Most recent methods perform feature matching or reconstruction for the target sample with pre‐trained deep neural networks. However, they still struggle to address challenging anomalies because the deep embeddings stored in the memory bank can be less powerful and informative. Specifically, prior methods often overly rely on the finite resources stored in the memory bank, which leads to low robustness to unseen targets. In this paper, we propose a novel subspace‐guided feature reconstruction framework to pursue adaptive feature approximation for anomaly localization. It first learns to construct low‐dimensional subspaces from the given nominal samples, and then learns to reconstruct the given deep target embedding by linearly combining the subspace basis vectors using the self‐expressive model. Our core is that, despite the limited resources in the memory bank, the out‐of‐bank features can be alternatively “mimicked” to adaptively model the target. Moreover, we propose a sampling method that leverages the sparsity of subspaces and allows the feature reconstruction to depend only on a small resource subset, contributing to less memory overhead. Extensive experiments on three benchmark datasets demonstrate that our approach generally achieves state‐of‐the‐art anomaly localization performance.

ABSTRACT The main challenge in visible‐infrared person re‐identification (VI‐ReID), which involves matching images of target persons across different modalities, is the significant modality gap between visible and infrared images. Current approaches generally fall into two categories: network architectures that map features from different modalities into a shared feature space, and methods that focus on modality generation and transformation. However, these methods often fail to fully account for contextual relationships, and the generated modalities may lack semantic interpretability. To address these limitations, this paper proposes a mixed modality generator that aligns the visible and infrared modalities as closely as possible within a unified feature space. To effectively leverage multi‐layered information, we introduce a hierarchical feature aggregation module, which establishes connections between features across different layers. Comprehensive experiments on the SYSU‐MM01, RegDB, and LLCM datasets demonstrate that our method significantly outperforms existing state‐of‐the‐art approaches, setting a new benchmark for VI‐ReID performance.

ABSTRACT The original generative adversarial network (GAN) model may struggle to adequately capture global information in images, particularly during complex decomposition tasks, leading to limitations in image clarity, detail retention and overall consistency. To address this challenge, we propose the large separable kernel attention generative adversarial network (LSKA‐GAN) model, building upon the blind image decomposition network (BIDeN). The LSKA module enhances BIDeN's global information capturing capability, thereby improving the quality and clarity of generated images. Experimental results demonstrate that LSKA‐GAN achieves obvious improvements in hybrid image decomposition. Compared to BIDeN, LSKA‐GAN exhibits an increase of 1.39 dB in peak signal‐to‐noise ratio (PSNR) and 0.04 in structural similarity index (SSIM). These improvements enable LSKA‐GAN to generate clearer images with more complete details, marking a notable advancement in image decomposition technology.