Search Results (33,255)

Visible-infrared (RGB-IR) object detection leverages multimodal information to ensure reliable perception in complex environments. However, dynamic scenes pose significant challenges due to the frequent inconsistency between scene-level modality contributions and local spatial reliability. Furthermore, standard feature extraction progressively attenuates boundary-sensitive structural cues, and unified fusion strategies often fail to capture spatially varying cross-modal complementarity. To overcome these limitations, we propose a Boundary-Guided Dual-Perspective Cross-Modal Fusion Network (BDPNet) to explicitly preserve shallow geometric structures and decouple deep semantic fusion into macroscopic and microscopic perspectives. Specifically, a Geometric Boundary Enhancement Module (GBEM) embeds Sobel-based high-frequency priors into shallow dual-modal features via residual spatial modulation, preventing the loss of crucial localization cues during downsampling. In the deep semantic space, a Hybrid Dual-Perspective Adaptive Fusion Module (HDAM) employs an illumination-aware branch for global modality weighting and a spatial confidence-driven branch for local cross-modal rectification. A spatial gating mechanism then dynamically reconciles these macro-environmental and micro-signal features. Extensive experiments on M3FD, LLVIP, and DroneVehicle demonstrate the effectiveness of BDPNet. Compared with state-of-the-art methods, BDPNet improves mAP${}_{50-95}$ by 0.8% and 1.0% on M3FD and LLVIP, respectively, and improves mAP${}_{50}$ by 0.6% on DroneVehicle, while using substantially fewer parameters and lower computational cost. Full article

With the rapid development of smart devices and mobile Internet, Location-based Consumer Electronics (LCE) increasingly rely on accurate destination prediction to support location-aware services. However, existing destination prediction methods often rely on recurrent architectures or incorporate unfiltered map information, which can limit both prediction efficiency and accuracy. To address these issues, we propose a Map-Aware Full Attention Model (MFAM) tailored for LCE devices, featuring an encoder-only full attention network optimized for real-time trajectory prediction. The main novelty of MFAM lies in a trajectory-conditioned Map-Aware Attention mechanism, which uses the observed trajectory to selectively identify and aggregate spatial map regions that are relevant to the destination prediction task, rather than directly incorporating the entire map. In addition, we introduce a sufficient grid-size criterion for AOI-based map representation to reduce the risk of losing small but informative geographic regions. These designs enable MFAM to exploit map information while maintaining a compact model structure and efficient inference. Experimental results on real-world trajectory datasets show that MFAM achieves a maximum Top-5 accuracy of 84.3% and 78.9% on two datasets, surpassing state-of-the-art methods. We further evaluate MFAM in real-world consumer electronics scenarios, demonstrating its robustness across different urban environments and transportation modes. Full article

Today, with the emergence of the Internet of Things (IoT) in smart home technology, new security issues have arisen, including device diversity, data privacy concerns, and susceptibility to advanced, complex attacks. Traditional Intrusion Detection Systems (IDS) face critical challenges, including single points of failure, high communication overhead, and inherent privacy violations, which the proposed decentralised approach addresses. In this paper, a novel Federated Anomaly Detection for Heterogeneous Smart Home Environment-based Zero-Trust Micro-Segmentation (CloudShield-IoT) framework is introduced. The proposed architecture introduces three key innovations: (i) the dynamic micro-segmentation engine to enforce zero-trust policies by periodically authenticating devices and monitoring their behaviour, (ii) the privacy-preserving federated learning module for collaborative intrusion classification across distributed smart home nodes without exchanging raw data. Note that the core detection module performs supervised intrusion classification using labelled attack categories; the term “anomaly detection” in this paper refers to the broader system-level behavioural-deviation monitoring achieved through the hierarchical trust-scoring mechanism. Moreover, (iii) the hierarchical trust-scoring module adaptively isolates compromised devices in real time. The framework is extensively tested on publicly available benchmark datasets, including N-BaIoT and IoT-23, and compared against 10 state-of-the-art baselines. CloudShield-IoT achieves 98.8% detection accuracy, an F1-score of 0.982, and an AUC of 0.993 without differential privacy, with an inference latency of only 12.4 ms. With practically meaningful (ε = 5.0, δ = 1 × 10⁻⁵) differential privacy via a moments accountant, the framework obtains an accuracy of 97.6%, an F1-score of 0.975, and an AUC of 0.989. The framework demonstrates strong resilience against Byzantine adversarial attacks. It maintains excellent scalability for deployments of up to 500 nodes, supporting heterogeneous device configurations ranging from 10 identical-type devices to 100 mixed-type devices per node. Full article

Recent advances have been achieved in intelligent fault diagnosis of rolling bearings. However, noise interference and data imbalance remain critical challenges for achieving accurate and reliable fault diagnosis in practical industrial applications. To address the limited robustness of conventional deep learning models under noisy conditions and their bias toward majority classes in imbalanced scenarios, this study proposes a robust bearing fault diagnosis method based on an adaptive weighted heterogeneous ensemble learning framework. The proposed method begins with continuous wavelet transform (CWT), which is employed to preprocess raw vibration signals and convert them into time–frequency images. Subsequently, a residual convolutional denoising autoencoder augmented by the convolutional block attention module is developed, namely CBAM-RCDAE. CBAM-RCDAE is capable of effectively reducing and eliminating noise interference in two-dimensional image data, thus enhancing fault diagnosis accuracy. Furthermore, a heterogeneous ensemble learning framework consisting of three base learners, including Swin Transformer, a multi-scale convolutional neural network, and BiLSTM, is developed to enhance generalization capability. An adaptive weight selection (AWS) strategy is introduced to adjust the weights and aggregate the outputs of the three base learners for final fault classification. The proposed method is extensively evaluated on the PU and CWRU bearing datasets. Experimental results demonstrate that, under the most challenging imbalanced conditions, the proposed method improves the G-mean metric by 5.89% and 4.95% compared with the state-of-the-art methods on the PU and CWRU datasets, respectively. In addition, the proposed method exhibits superior noise robustness, enabling reliable fault diagnosis performance across various noise levels. Full article

This state-of-the-art review critically evaluates shotcrete for concrete rehabilitation using an application-based framework covering buildings, bridges, tunnels and underground works, hydraulic and marine structures, and industrial facilities. The review synthesizes deterioration mechanisms, substrate and interface conditions, material and process variables, mechanical and durability performance, quality control, and alternative repair systems. Recent studies are integrated with established guidance and case histories, and quantitative evidence tables are used to facilitate cross-study comparison. Across structural applications, the shotcrete–substrate interface is identified as a critical factor governing rehabilitation performance, with surface condition, moisture state, shrinkage, curing, and environmental exposure affecting bond and durability. The synthesis further demonstrates that performance requirements vary by application, including section restoration, confinement, bond, and durability in buildings and bridges; early-age support, toughness, and residual capacity in tunnels; and low permeability and abrasion/erosion resistance in hydraulic and marine structures. Five research questions are proposed to address key uncertainties in interface behavior, durability, field performance, material development, and quality assurance. The resulting framework provides a systematic basis for evaluating shotcrete rehabilitation strategies and identifying priorities for future research. Full article

Gene expression classification remains a challenging task due to the high dimensionality and heterogeneity of available datasets. In this study, we present a comprehensive empirical analysis combining neural networks and Support Vector Machines (SVMs) for gene expression classification. We evaluate a wide range of architectures, including convolutional neural networks (CNNs) and the recently introduced Kolmogorov–Arnold Networks (KANs), across more than ten publicly available datasets. Furthermore, we explore ensemble strategies and show that ensemble models achieve statistically significant improvements over recently proposed state of the art approaches. All developed code, along with the exact data splits used in our 10-fold cross-validation experiments, is publicly available to ensure full reproducibility of our results. Full article

Perovskite materials are considered highly promising because their unique crystal structure enables exceptional properties for a variety of applications. Therefore, they have been a very active area of research for decades. Barium titanate (BaTiO${}_3$), one of the most popular among these structures, has prompted various attempts to obtain this material, both on a micro- and nanoscale, due to its versatility in multiple disciplines and fields of technology. This article aims at providing an overview of the literature discussing synthesis and characterization of barium titanate materials. A detailed comparison shows the links between synthesis method and material properties. Moreover, we have condensed the literature into tables highlighting the main conclusions and allowing newcomers in the field to obtain a rather comprehensive picture of methods and approaches. Full article

Automated prostate cancer (PCa) lesion segmentation using deep learning remains constrained by limited tissue contrast in standard diffusion-based MRI sequences, with state-of-the-art methods reporting Dice scores of 32% or lower on large patient cohorts. Synthetic correlated diffusion imaging (${\mathrm{CDI}}^s$) offers a promising solution, providing enhanced tissue contrast derived entirely from existing diffusion-weighted imaging (DWI) acquisitions at no additional clinical cost. This study presents the first comprehensive evaluation of ${\mathrm{CDI}}^s$ integration across the full standard multiparametric MRI protocol, encompassing 15 modality configurations and six segmentation architectures spanning CNN and transformer families on a cohort of 200 patients. ${\mathrm{CDI}}^s$ reliably enhances or preserves segmentation performance in the evaluated configurations, with 19 statistically significant improvements and no significant degradations across 42 direct comparisons. ${\mathrm{CDI}}^s$ enhancement primarily operates as a recall-driven mechanism, improving lesion detection sensitivity while largely preserving precision. ${\mathrm{CDI}}^s$ + DWI + T2w emerged as the strongest clinically meaningful configuration, achieving significant Dice improvement in four of six architectures, with no instances of degradation. Grad-CAM-based explainability analysis further reveals that ${\mathrm{CDI}}^s$ focuses poorly localized CNN attention toward lesion boundaries, while transformer architectures exhibit more stable attention patterns that are less sensitive to ${\mathrm{CDI}}^s$ integration. These results establish validated ${\mathrm{CDI}}^s$ integration pathways and provide architecture-specific deployment guidance for clinical implementation. Full article

Identifying the associations between non-coding RNAs and drug resistance (RDRAs) is crucial for uncovering resistance mechanisms and screening effective biomarkers. However, existing graph-based methods typically rely on purely data-driven learning and commonly assume a simplified noise independence hypothesis, overlooking the complex dependencies among feature, structural, and label noise in biomedical networks. This leads to issues such as spurious associations, poor generalization, and lack of interpretability for noisy association prediction tasks. To address these challenges, we propose Prior-RDRGSE, a prior-knowledge-guided dependency-aware graph learning framework. This framework integrates both dependency-aware graph noise modeling and domain knowledge into graph representation learning. Specifically, we first construct a heterogeneous bipartite graph and employ a deep generative inference encoder to jointly infer the underlying clean graph structure and the association signals, thereby explicitly modeling and purifying the intertwined complex noise within the network. Next, we design a resistance-semantics-conditioned interaction module that injects disease-specific and mechanism-related semantic priors into attention queries, explicitly guiding subnetwork interactions in a biologically plausible manner. Furthermore, we introduce a resistance consistency constraint based on KL divergence, which regularizes model training by aligning the learned association distribution with prior distributions derived from clinical and literature data. Comprehensive experiments demonstrate that Prior-RDRGSE achieves state-of-the-art performance in RDRA prediction and significantly outperforms existing methods. Full article

Low-Dose Computed Tomography (LDCT) reduces radiation exposure but introduces severe quantum noise and streak artifacts that degrade image quality. To address these challenges, we propose the Physics-Guided Sequential State Space Transformer ($P{S}^{3}T$), a projection-domain denoising framework that combines sequential state-space modeling with a photon-aware attention mechanism to capture long-range dependencies across projection angles with linear computational complexity. A differentiable Filtered Backprojection (FBP) layer further enforces reconstruction-domain consistency during training. The proposed framework was evaluated on the Mayo Clinic LDCT and Projection Dataset using patient-level dataset partitioning. Experimental results demonstrate that $P{S}^{3}T$ consistently outperforms state-of-the-art methods, including DRL, SADiff, and GEDFormer, across the abdomen, head, and chest datasets. On the abdomen dataset, $P{S}^{3}T$ reached a peak PSNR of 42.40 dB, an SSIM of 0.9020, and the lowest RMSE of 0.0076 across anatomical regions. Statistical analysis using 95% confidence intervals and paired Wilcoxon signed-rank tests confirmed that these improvements were significant ($p<0.05$). Furthermore, $P{S}^{3}T$ achieved the lowest reconstruction consistency loss (0.0128 at epoch 50), demonstrating stable convergence and the effectiveness of incorporating acquisition physics into projection-domain LDCT denoising. Full article

Remaining useful life (RUL) prediction of power transformers is essential for ensuring reliable operation and supporting condition-based maintenance. However, the complex nonlinear degradation characteristics and long-term temporal dependencies of transformer monitoring data pose significant challenges for accurate RUL prediction. Moreover, most existing deep learning-based prediction methods provide deterministic predictions without explicitly quantifying predictive uncertainty, while the consequences of RUL overestimation and underestimation in practical maintenance are inherently asymmetric. Therefore, this paper proposes an uncertainty-aware RUL prediction method for power transformers by integrating the Patch Time Series Transformer (PatchTST) with deep evidential regression (ER-PatchTST). PatchTST is used to transform long time-series inputs into patch-level representations, enabling the model to capture local temporal patterns and long-range dependencies through patch-wise tokenization and channel-independent modeling. In addition, the deep evidential regression module is designed by placing a Normal–Inverse-Gamma (NIG) prior over the parameters of the Gaussian likelihood, which can simultaneously predict RUL and quantify aleatoric and epistemic uncertainties in a single forward pass. Furthermore, a safety-oriented RUL indicator and hierarchical warning strategy are developed, and different maintenance actions are initiated when alarms at different levels are triggered. Experiments on the ETT dataset demonstrate that ER-PatchTST achieves competitive forecasting performance compared with state-of-the-art time-series forecasting methods while simultaneously providing predictive uncertainty estimates. On the DGA dataset, ER-PatchTST achieves the best RUL prediction performance among the compared methods and provides informative uncertainty quantification for condition-based maintenance decision support. Full article

Influence maximization (IM) in social networks aims to identify the optimal set of seed nodes that maximizes influence spread under a given diffusion model. The standard Grey Wolf Optimizer (GWO) suffers from two fundamental limitations when applied to this problem: an inflexible exploration–exploitation transition controlled by a linearly decreasing parameter; and a rigid three-level leadership hierarchy that suppresses individual diversity and promotes premature convergence. In this paper, we propose a Multi-Role Cooperative Grey Wolf Optimizer (Multiple-roles GWO) that addresses both limitations through two complementary mechanisms. First, a Q-learning-based adaptive phase transition mechanism monitors population diversity, fitness improvement rate, and iteration progress in real time, enabling the algorithm to dynamically shift between exploration and exploitation. Second, inspired by the principle of division of labor, the exploitation phase is restructured into a four-role cooperative framework comprising leaders, explorers, followers, and losers, each executing a distinct search strategy to improve local search coverage and maintain population diversity. Experiments on six real-world social networks under the Independent Cascade model show that Multiple-roles GWO achieves competitive or superior influence spread compared with state-of-the-art heuristic baselines, with comparable computational efficiency. Full article

Fabric defect detection plays a vital role in the quality control of the textile manufacturing industry. However, it remains challenging because of defect diversity, complexity, and environmental factors. Deep learning-based methods efficiently extract visual features, improving detection accuracy and inference speed. However, most deep learning methods fail to adequately capture tiny, irregular, and blurred-edge defect features due to diverse fabric texture backgrounds and complex defect traits. To address these issues, we propose CMNet, a novel hybrid CNN–Mamba detection network with a dual-branch, heterogeneous feature-extraction architecture. First, we propose an Adaptive Weighted Adjacent Context Coordination Module (AWA-CCM) to enhance features of tiny fabric defects and edge textures while effectively suppressing interference from complex backgrounds. Second, we propose a Multi-Scale Large-Kernel Attention Mamba (MLAMamba) module to improve the network’s global representation capability for defects with variable scales and irregular shapes. Finally, we construct a Bidirectional Fusion Module (BFM) to dynamically balance local detail information and global structural information for efficient, complementary feature fusion. Experimental results on our self-developed fabric datasets, obtained using six widely adopted metrics, show that CMNet significantly outperforms state-of-the-art methods. Full article

Microrobots have the power to transform healthcare and the medical sector by improving diagnosis and targeted drug delivery. Real-time detection of microrobots is critical to ensure reliable operation. Ultrasound (US) imaging is used for detection because it is non-ionizing, low-cost, and easy to set up. Using US images is challenging because certain microrobot shapes (classes) remain difficult to detect due to their small size, low contrast, and unstable appearance. This paper proposes Selective Class-Aware Refinement (SCAR), a lightweight post-detection method built on a baseline YOLO detector that boosts its real-time detection capabilities. SCAR improves weak-class detection for visually challenging objects that are small, low-contrast, or unstable, without sacrificing inference speed. To identify weak classes, SCAR is designed to analyze class-specific detection behavior offline. It does so by examining failure statistics and visual difficulty cues. During prediction, SCAR selectively focuses on difficult-to-predict classes using a compact specialist detector applied to adaptive local crops whose size scales with the detected object’s dimensions. SCAR limits the refinement to small local regions rather than reprocessing the entire US frame. SCAR preserves near-baseline frame-per-second (FPS) while improving detection. We conducted experiments on the open dataset USMicroMagSet and tested the method on two YOLO baselines and five other state-of-the-art detectors. The results show that SCAR improves weak-class performance with limited computational overhead: it raises weak-class detection accuracy from near-zero to a usable level, improves overall mAP, and maintains consistent performance. Full article

Remote-sensing small-object detection offers significant advantages and holds great importance in monitoring and measurement fields such as scene perception and environmental monitoring. Although deep neural networks have advanced the development of remote-sensing object detection, challenges remain, including large-scale variations and the difficulty of detecting dense, minute objects. To address these challenges, we propose a dynamic multiscale adaptive mechanism-driven remote-sensing object detection network (DMAM). First, to overcome the inherent limitations of traditional convolutional fixed sampling positions and uniform parameter distributions, we introduce adaptive kernel convolution (AKConv). By dynamically adjusting sampling positions and optimizing parameter distributions, AKConv enables more flexible and efficient feature extraction. Second, to effectively leverage prior spatial knowledge for expanding the receptive field, we propose a dynamic multiscale context adaptation (DMCA) module. This module implements a content-aware dynamic gating mechanism, which adaptively allocates weights between local details and global context by analyzing image content at each spatial location. By integrating local features with global information, it enhances scene comprehension, thereby improving detection accuracy and robustness. Finally, a shape-aware metric function (Shape-IoU) is introduced, which incorporates a shape-adaptive weighting mechanism to dynamically adjust the penalty weights for different geometric factors based on the target’s own shape characteristics, thereby achieving more precise bounding box regression. Results on three public datasets show that the proposed method demonstrates robust performance compared to state-of-the-art detection networks. Specifically, DMAM achieves an average accuracy of 97.7% on the RSOD dataset, 92.5% on the NWPU VHR-10 dataset, and 88.0% on the DIOR dataset. Full article