Vol. 45 · No. 5

Image-based virtual try-on technology aims to match in-store clothing with images of individuals wearing the garments accurately. However, current approaches often struggle to preserve the texture and pattern of the target clothing, particularly when dealing with complex poses. To address these limitations, we propose a novel framework, Style-VTON, for stylized image-based virtual try-on. This framework is specifically designed to maintain the fidelity of clothing textures and patterns, resulting in visually convincing try-on experiences. The Style-VTON framework comprises three specialized sub-networks, each performing a specific role. The first sub-network generates a highly detailed human parsing map, which accurately segments the person in the image. The second sub-network projects clothing patterns and textures onto the deformed body silhouette in real-time, ensuring spatial coherence and alignment. The third sub-network introduces user-driven customization, allowing for adjustments to garment texture, color, and style, thereby enhancing user experience through personalized virtual try-on results. Additionally, by leveraging the Pix22Dsurf method, the Style-VTON enhances its adaptability and achieves state-of-the-art performance across quantitative image quality benchmarks.

Cooperative perception enables connected and autonomous vehicles to overcome the inherent limitations of individual sensors, such as occlusion and a restricted field of view, by sharing perceived objects among vehicles. While this enhances situational awareness, it introduces significant redundancy in dense traffic, leading to unnecessary computation and communication overhead. This paper proposes a novel redundancy-aware perception framework that integrates explicit object-level redundancy assessment directly into the detection stage. We extend the YOLOv11x architecture to ingest lightweight occupancy and inverse-distance masks derived from shared data, enabling the model to perform simultaneously 2D object detection and binary redundancy classification in a single inference step. On V2X-Sim, the proposed model achieves a redundancy F1-score of 0.972 with a measured forward-pass latency increase of 7.70 % (0.687 ms) relative to the RGB detector. By identifying redundant objects early, the framework enables resource-saving strategies, such as context-aware downsampling and selective offloading.

The accurate identification of surface defects in industrial products is critical for ensuring consistent high-quality standards in modern manufacturing. Nevertheless, the prevalence of small-scale defects and pronounced cross-scale variability on industrial surfaces poses substantial challenges to achieving precise detection. To address these issues, this study introduces a high-precision neural network architecture, termed DAE-YOLO, which is developed based on the YOLOv11 framework and specifically tailored for industrial surface defect detection. To effectively mitigate the impact of cross-scale feature discrepancies, an efficient multi-scale attention mechanism (EMA) is embedded within the backbone network. This leads to the design of a novel feature extraction module, D-EMA (Deformable and Efficient Multi-scale Attention), which enhances the model's ability to capture both global contextual dependencies and local structural details. This module enhances the feature extraction capabilities of the convolutional neural network by fusing global context and local features, thereby enabling precise modeling of spatial information and pixel-level relationships. The design of D-EMA carefully considers the fusion of cross-spatial information, allowing the model to effectively process targets at varying scales while preserving accurate location information, significantly improving performance in defect detection tasks. Further, to address the challenge of detecting small targets, the latest deformable convolutional module was incorporated into the Neck component. A new bidirectional weighted feature pyramid network (BDFPN, bidirectional deformable feature pyramid network) was then designed, combining multi-scale features to enhance the representation of small defect features. Experiments conducted on two publicly available datasets (NEU-DET and PCB defect) demonstrated the effectiveness of the proposed network architecture. The results indicate that the proposed method outperforms YOLOv11 by 3.88 % in detection accuracy on the NEU-DET dataset, and by 2.8 % on the PCB defect dataset.

As digital platforms continue to reshape the way knowledge is acquired and shared, there is a critical need for effective community discovery algorithms that can operate in sparse data environments. Traditional methods often fail to capture the complex dynamics of online educational platforms, where explicit relationships among learners are typically underrepresented. Therefore, to address the challenges of sparse inter-learner connections within online learning communities, this paper proposes a novel framework that introduces the ``Cyber Learning Community'' model which reconceptualizes learner relationships by emphasizing collaborative and supportive interactions. Specifically, it combines a modified graph neural network (GNN) architecture and a bespoke community discovery algorithm to unearth and reinforce the latent networks within large datasets. The proposed methodology involves two primary tasks: predicting learners' interest labels from their online behavior, and identifying robust community structures through the innovative use of the Discrete Beetle Antennae Search algorithm (DBAS). The effectiveness of our approach is demonstrated through extensive experiments on real-world datasets, showing marked improvements in community coherence and learning engagement. The results indicate that our methods not only enhance the detection and formation of educational communities but also suggest scalable applications for broader social network analysis.

The proliferation of the Internet of Everything (IoE) has transformed smart home environments, necessitating intelligent, secure, and responsive systems that can make real-time decisions at the network edge. This paper presents HomeNetEdge, an edge-AI-enabled smart home automation framework that combines lightweight ensemble learning models with blockchain-based security to address the pressing challenges of privacy, identity theft, and autonomous device control. The system integrates logistic regression and random forest classifiers using principal component analysis (PCA) for feature selection, ensuring accurate and efficient predictions of device states (``ON''/``OFF'') with minimal computational overhead. HomeNetEdge operates on commercial off-the-shelf hardware, including Raspberry Pi servers, 5V relay modules, and sensor arrays, and leverages local edge computing to eliminate reliance on cloud infrastructure. A custom Android application interfaces with the system via HTTP and Apache servers for seamless user control. To ensure robust security and trustworthy communication between devices, blockchain-based authentication mechanisms are embedded. The framework is evaluated under both controlled and real-life conditions, demonstrating real-time performance, scalability, and resilience against cyber threats. This work contributes a practical, cost-effective, and scalable edge computing architecture that supports real-time AI-driven decision-making, personalized automation, and secure data handling – critical for the future of intelligent IoT ecosystems.

Aiming at the positioning problem of dynamically deployed temperature and smoke sensors in intelligent fire protection systems, an ultra-wideband (UWB) positioning method was applied. This led to the design of a dynamically deployed sensor that integrates temperature and smoke sensing, positioning, and wireless communication functions. The adaptive obstacle compensation trilateral measurement algorithm and nonlinear least squares compensation were used to achieve accurate positioning in harsh fire environments. Experiments show that in complex indoor environments, when the number of partition walls on a single coordinate axis does not exceed 3, the positioning error can be within 3 cm. The sensor can also obtain the on-site temperature and smoke distribution in real time.

Falls are a leading cause of severe injuries and fatalities among the elderly, particularly those living alone. Timely and accurate fall detection is critical to ensuring rapid medical intervention and mitigating associated health risks. Traditional fall detection methods, such as wearable sensors and conventional computer vision techniques, often suffer from high false alarm rates and limited adaptability to real-world environments. These limitations stem from challenges like occlusions, lighting variations, and the inability to reliably distinguish between actual falls and non-harmful postures, such as sitting or lying on a bed. To address these challenges, we propose a novel fall detection system that integrates fine-grained human posture segmentation with environmental context awareness. Our approach leverages an enhanced YOLOv11-Seg model to perform real-time semantic segmentation, enabling precise differentiation between actual falls and similar but non-critical scenarios. By analyzing the spatial relationship between the detected individual and surrounding furniture, our method significantly reduces false alarms while maintaining high detection accuracy. Experimental evaluations conducted across multiple benchmark datasets demonstrate the robustness of our solution, achieving superior performance in terms of accuracy, precision, recall, and F1-score. The proposed system not only outperforms existing state-of-the-art methods but also proves highly effective for real-world fall detection applications, offering a reliable and scalable solution to enhance elderly care and safety.

Federated learning (FL) promises to protect user data through collaborative training, yet gradient leakage attacks can accurately reconstruct sensitive information from shared updates. While homomorphic encryption (HE) offers a formidable defense, its high computational and communication costs often make it impractical for real-world deployment. In this work, we propose FL SelHE Federated Selective Homomorphic Encryption, a framework that strategically secures only the most vulnerable layers of a model. By using a one-time sensitivity profiling based on gradient magnitudes, we identify critical layers and apply the CKKS encryption scheme selectively. Experimental evaluation demonstrates that the framework achieves 94.79 % accuracy, comparable to unencrypted training, while reducing the aggregation time by over 40 % compared to fully encrypted models. Furthermore, under gradient inversion attacks, the reconstructed images degrade to near-random noise (PSNR 5 dB and SSIM 0), confirming strong resistance to data leakage. These findings prove that selective encryption is a scalable and efficient solution for robust privacy in sensitive domains like healthcare.

The uncertainties of wind power generation and load significantly impact on the economy of distribution network dispatching. Therefore, operating an optimal dispatching strategy can improve the economic performance of the power grid. In modern data-driven optimization approaches, the proper volume of data generates remarkably optimal results. However, due to the difficulties of data collection, transmission, and storage, wind power data is naturally scarce, which severely deteriorates the performance of the optimization. Additionally, the uncertainty existing on the load side leads to more complicated scenarios in optimization, which further prevents the optimization from achieving the optimal dispatching strategy. Therefore, to solve the aforementioned issues, this paper presents an economic dispatching optimization approach for a distribution network while considering wind power data augmentation and uncertainties of both wind power and load. Firstly, the synthetic minority over-sampling technique (SMOTE) is adopted to generate a proper number of wind power generation output data samples. And then, a variational autoencoder (VAE) is further employed to generate adequate data samples with a similar distribution to the original data samples. Secondly, the deep embedded clustering algorithm (DECA) deeply characterizes and clusters the augmented data samples to figure out typical wind power generation output scenarios overrepresented in the samples. Thirdly, for each typical scenario, an ambiguity set is constructed using an imprecise Dirichlet data-driven model (IDDM); And then, the details, including the objective functions and constraints of the economic dispatching for the distribution network are analyzed and modeled. Finally, a two-stage distributed robust optimization (DRO) model considering the uncertainties of both wind power generation output and load for achieving the optimal economic dispatching strategy for the distribution network is established and solved; The case studies demonstrate the effectiveness of the presented approach.