Search Results (6,259)

Object detection in UAV aerial images remains fundamentally constrained by extremely small object scales, strong background interference, and progressive structural information degradation along the feature extraction pipeline. Current small-object detection methods suffer from two fundamental deficiencies rooted in their convolutional feature extraction pipelines: the smoothing effect of strided convolutions in early layers, which attenuates fine-grained details before backbone processing, and the feature overwriting phenomenon, where sequential transformations progressively erase structural information from earlier layers. We propose PFE-Det (Progressive Feature Evolution Detector), built upon the DEIM framework and constructing a continuous optimization pathway across three stages. A Feature Adaptive Enhancement Network (FAENet) is adopted as a front-end preprocessor to decouple high- and low-frequency components via a Laplacian pyramid at the input stage, mitigating early-layer smoothing at the input. A Multi-Receptive-Field Adaptive Fusion Module (MFAM) is designed to reorganize single-path features into structure-retaining and progressive enhancement paths and is further coupled with hierarchical receptive-field modeling, suppressing feature overwriting through multi-scale context modeling. A Multi-Path Gated State Space Modeling Block (MG-SSM Block) couples HSM-SSD-based long-range dependency extraction with Convolutional Gated Linear Units (CGLU) for adaptive feature selection in the encoder. Experiments on VisDrone2019 demonstrate an AP of 0.225 and an ${\mathrm{AP}}_s$ of 0.134, yielding a 13.5% relative improvement in small-object precision over the baseline. Cross-dataset evaluations on DIOR and UAVVaste, each under independent training and testing, further support the effectiveness of progressive feature evolution for UAV small-object detection. Full article

The catenary support component (CSC) is a key part of the electrified railway system, and its operational status directly affects railway operational safety. These components’ images are collected using inspection equipment and detected using computer vision techniques. However, catenary network inspection faces the following issues: (1) due to limitations in the equipment’s shooting angle and changes in viewing distance, the collected images contain multi-scale and multi-class problems, and (2) the railway environment is highly variable, and adverse weather conditions such as fog, rain, and low light affect the imaging devices, leading to degraded image quality. To address these issues, this paper proposes a novel detection framework, FF-DEIM, for detecting catenary support components. First, a dual-channel fusion network (DCFNet) is introduced, which significantly improves image quality by removing foreground interferences such as fog, raindrops, and dynamic blur. Second, a pretraining framework based on contrastive learning, mask image modeling with contrastive learning (MIMCL), is designed to enhance the model’s focus on key regions of the catenary network components, optimizing feature extraction capabilities and improving model convergence speed. Then, a feature-focusing pyramid network (FFPN) is proposed, which uses the focus feature module to fuse cross-level contextual features, enhancing the ability to capture local details and improving the model’s small object detection performance. Finally, a drone-based catenary network image dataset, including scenes with fog, rain, and low light, is constructed, and experiments validate the effectiveness of the proposed method. Full article

With the continuous penetration of renewable energy and power electronic equipment into modern power systems, the occurrence frequency of composite power quality disturbances has increased significantly. The accurate classification of various composite disturbances under strong noise remains a critical technical challenge. The existing single time–frequency transformation methods cannot simultaneously capture transient time-domain details and fine frequency-domain features of steady-state harmonics, while mainstream deep learning classification networks contain redundant parameters and introduce excessive computational overhead, failing to meet the real-time deployment requirements of power edge terminals. To address these limitations, a lightweight Coordinate Attention ResNet network named ResNet–LCA is proposed based on the dual-modal time–frequency fusion of the Continuous Wavelet Transform and Short-Time Fourier Transform. First, the two transforms are implemented separately to generate two groups of complementary time–frequency maps, which are concatenated along the channel dimension to fully extract the coupling features between the steady-state harmonics and the transient impulses. Second, a Haar wavelet subband mean aggregation module is designed for dimensionality reduction with negligible information loss. This module eliminates the channel redundancy introduced by the multimodal fusion and reduces the overall computational overhead at the input stage. Finally, a lightweight residual network integrated with Coordinate Attention is constructed, with Grouped Half-Convolution adopted to compress the model parameters. CA offsets the feature attenuation induced by the lightweight structural design and further improves the model’s noise immunity. A simulation verification was carried out on a simulated dataset covering 25 types of single and superimposed composite disturbances. At a signal-to-noise ratio of 20 dB, the proposed method achieved an average classification accuracy of 97.92%, with only 5.32 M total parameters and a single-sample GPU inference latency of 0.33 ms. Compared with standard ResNet-18 under 20 dB noisy conditions, the total parameter volume was reduced by 52.7%, the inference latency was shortened by 0.13 ms, and the classification accuracy was improved by 0.60 percentage points. The proposed method achieves coordinated optimization of classification accuracy, noise immunity and inference efficiency, and it can provide lightweight technical support for online intelligent power quality monitoring at the edge nodes of microgrids and islanded power systems. Full article

Han stone reliefs contain rich historical and artistic information, but their images often exhibit severe weathering, rubbing noise, and complex carving textures, making fine-grained classification challenging. To address the strong coupling between global semantic structures and local material textures, this paper proposes Frequency-Guided Collaborative ConvNeXt (FGC-ConvNeXt), a dual-branch spatial–frequency classification framework. The spatial branch extracts multi-scale semantic features using ConvNeXt-Tiny, while the frequency branch applies Fast Fourier Transform to model complementary structural and texture-sensitive information. A fixed frequency threshold of τ = 0.22 is used for frequency decomposition, and both branches are projected into a 512-dimensional feature space. The Spatial–Frequency Gated Fusion module dynamically adjusts their contributions at the sample level, while frequency-domain auxiliary supervision with a loss weight of λ = 0.4 prevents branch degradation during joint optimization. Experiments were conducted on a self-constructed four-class dataset containing 6368 Han stone relief images with an input resolution of 224 × 224 pixels. Over three independent runs, FGC-ConvNeXt achieved an accuracy of 97.6 ± 0.2%, a precision of 97.8 ± 0.2%, a recall of 97.4 ± 0.3%, and an F1-score of 97.6 ± 0.2%, outperforming the ConvNeXt-Tiny baseline by 4.1 percentage points in accuracy. The proposed model contains 29.8 M parameters. Visualization analyses provide qualitative support for more separated feature distributions and more concentrated activation patterns, while the robustness experiments indicate improved classification stability under the evaluated conditions. Full article

Background and Objectives: Osteoporotic vertebral compression fractures affect approximately one in four postmenopausal women and carry substantial morbidity, yet established clinical tools such as dual-energy X-ray absorptiometry (DXA) provide only patient-level risk and do not identify which specific vertebra is most likely to fail. Computed tomography (CT) acquired for unrelated indications is the most widely available three-dimensional substrate for opportunistic screening, but published machine learning models for vertebral fracture risk almost universally operate at the patient level. The present study aimed to develop and rigorously validate a per-vertebra prediction pipeline applicable to both routine clinical lumbar-spine CT and opportunistic abdominal CT, both acquired for indications unrelated to osteoporosis screening. Materials and Methods: Two independent retrospective cohorts were assembled from a single academic centre: a routine clinical lumbar-spine CT cohort of 106 patients yielding 478 evaluable vertebrae, and a routine abdominal CT cohort of 126 patients yielding 589 evaluable vertebrae. Vertebral bodies were segmented automatically with TotalSegmentator v2 and the trabecular core isolated by morphological erosion. A panel of 505 quantitative imaging biomarkers compliant with Image Biomarker Standardisation Initiative recommendations was extracted, covering trabecular density, vertebral morphometry, classical texture, trabecular network architecture, sub-endplate vulnerability, low-density topology, radial heterogeneity and adjacent muscle quality. Within-patient feature engineering expanded the input pool to 1293 contextual descriptors. Three model families were evaluated under fully nested leave-one-patient-out cross-validation: ElasticNet logistic regression, a softmax-ranking approximation of conditional logistic regression, and a Two-Stage model combining a patient-level fragility score with a within-patient vertebral outlier score. Patient-level bootstrap resampling (2000 iterations) was used to obtain 95% confidence intervals. Results: On routine clinical lumbar-spine CT the Two-Stage model achieved a per-vertebra AUC of 0.750 (95% CI 0.704 to 0.795), an F1 of 0.549, a within-patient concordance index of 0.693, an expected calibration error of 0.044, and Hit@3 of 0.934. It was the only model evaluated that returned calibrated probabilities; the softmax-ranking and ElasticNet baselines gave expected calibration errors of 0.232 and 0.218 respectively. On opportunistic abdominal CT, the softmax-ranking model gave AUC 0.672 (95% CI 0.615 to 0.727). Selected biomarkers were dominated by regional trabecular density and trabecular network architecture; a stable core of lumbar features entered the model in 100% of cross-validation folds, indicating high reproducibility. The closest prior per-vertebra CT-based predictor in primary, non-surgical patients (Muehlematter and colleagues, 58-patient cohort) reported a per-vertebra AUC of 0.64, which is one of several reference points for the present results. Ten methodological variants and sensitivity analyses, including rank fusion, internal tissue normalisation and additional biomechanical features, did not provide statistically significant gains, indicating that the binding constraint at this sample size is data volume rather than methodology. Conclusions: A two-stage decomposition that separates systemic skeletal fragility from within-patient vertebral outlier status produces well-calibrated per-vertebra fracture-risk estimates from routine clinical lumbar spine CT and was the only model evaluated to do so, which is what permits a per-vertebra output to be reported as an absolute risk rather than as an ordering alone; a within-patient ranking model is preferable for opportunistic abdominal CT. The discrimination advantage of the decomposition over that baseline is numerical and consistent but not statistically established at this sample size, and the work is presented as a transparent and reproducible single-centre benchmark for the still under-developed per-vertebra prediction task. Its clearest near-term value is opportunistic, namely flagging elevated per-vertebra fracture risk on CTs already acquired for unrelated indications without additional radiation, cost or a dedicated densitometric study. External multi-centre validation is the necessary next step. Full article

Accurate estimation of the State of Health (SOH) of lithium-ion batteries is critical to ensuring the safety and stability of energy storage and power systems. However, existing SOH evaluation methods fall short in adaptive denoising, multi-scale feature extraction, and effective fusion of global and local information. To overcome these limitations, this paper proposes a global–local feature fusion network with an adaptive wavelet filter (AWGLFuser) for end-to-end SOH estimation. The proposed model consists of three modules: an adaptive wavelet filter (AWF) module to suppress high-frequency noise and highlight key information in the frequency domain; a global–local feature extraction (GLFE) module to capture both global and local temporal dependencies at multiple scales; and a bidirectional cross-attention fusion (BCAF) module to enable deep interaction between global and local features, thereby facilitating their effective fusion. Comparison experiments on the NASA and XJTU datasets demonstrate that the proposed model yields lower estimation errors than the eight competing models. The average reductions in MAE, MAPE, and RMSE are 44.158%, 44.209%, and 38.014%, respectively, and this improvement is statistically significant against every comparison model. Furthermore, ablation studies clarify what each component contributes, with the modules proving mutually reinforcing when combined. AWGLFuser also attains a compact parameter count and storage footprint with competitive inference latency, despite comparatively higher FLOPs. Full article

Accurate mapping of paddy rice is essential for agricultural monitoring, yield estimation, and food security assessment. However, optical imagery is often affected by clouds and spectral confusion, while SAR imagery suffers from speckle noise and weak spatial detail representation. Simple optical and SAR feature concatenation is therefore insufficient for complex agricultural landscapes. To address these limitations, this study proposes ACBE-CroFuseNet, an optical and SAR cross-fusion semantic segmentation network for paddy rice extraction using Sentinel-1 SAR and Sentinel-2 optical imagery in Yancheng, Jiangsu Province. ACBE-CroFuseNet introduces two task-oriented designs for paddy rice mapping. First, an attention cross-fusion module is developed to adaptively model modality contributions and spatial responses between optical spectral–textural features and SAR scattering–structural features. Second, a boundary enhancement module with boundary supervision is introduced to strengthen the delineation of fragmented paddy fields and field edges. Multimodal feature aggregation and multi-scale deep supervision are further used to improve feature utilization and segmentation stability. Compared with UNet++, Swin-Unet, CroFuseNet, and CMFFNet under five-fold cross-validation, ACBE-CroFuseNet achieves the best overall performance. The extracted paddy rice area in Yancheng in 2025 demonstrates the applicability of the proposed method for large-scale crop mapping. Full article

High-resolution remote sensing imagery provides valuable data support for accurate binary urban green space extraction. However, due to complex urban backgrounds, existing methods still face challenges in accurately delineating green space boundaries and preserving fine-scale spatial details. To address these issues, this study proposes a novel network, termed Hierarchical Gated Residual Fusion and High-Frequency Guided Deformable Upsampler Network (HGRHDNet), for boundary-enhanced urban green space segmentation. The proposed framework adopts ConvNeXt-L as the encoder backbone and incorporates a Hierarchical Gated Residual Fusion Decoder (HGRFD) to adaptively fuse multi-scale features through dynamic weighting and residual feature propagation. In addition, a High-Frequency Guided Deformable Upsampler (HFGDU) is introduced to enhance high-frequency detail reconstruction and cross-resolution feature alignment, thereby improving boundary localization accuracy. The proposed method was evaluated on three public datasets with different spatial resolutions and spectral characteristics, including WHDLD, UGS-1m, and UBGG. Experimental results show that HGRHDNet achieves Boundary Intersection over Union (BIoU) values of 43.82%, 18.22%, and 64.55% on the three datasets, respectively, consistently outperforming state-of-the-art methods. Both quantitative and qualitative analyses demonstrate that HGRHDNet effectively preserves narrow gaps between adjacent green spaces, elongated vegetation structures, and fragmented green space patches while reducing boundary ambiguity in complex urban environments. These results indicate that HGRHDNet provides a robust and effective solution for high-resolution urban green space extraction and has considerable potential for applications in urban ecological assessment, green space inventory, and sustainable urban planning. Full article

Manual visual diagnosis of apple diseases suffers from low efficiency, strong subjectivity and poor scalability for large commercial orchards. Existing research mainly targets diseases on single plant organs, whereas full-growth-cycle detection has to cope with extreme multi-scale differences among lesions. For instance, Valsa canker on tree trunks leads to extensive cortical necrosis, while early-stage anthracnose on fruits appears as tiny spots spanning only a few pixels. These significant scale gaps necessitate robust spatial feature aggregation and anti-noise ability to resist complex background interference. Aiming to achieve rapid and precise detection of diseases on multiple apple organs including leaves, fruits, trunks and branches, this work presents an enhanced YOLOv11 model equipped with the Information Integration Attention (IIA) module. The IIA module is embedded into the key fusion layers of the backbone and neck networks. It strengthens the extraction of fine-grained lesion features, recovers spatial location information via a bidirectional attention mechanism, and suppresses noise induced by uneven lighting and intricate backgrounds. To guarantee stable convergence on low-resource computing devices, a tailored training scheme is designed. Experimental results on a seven-category dataset with 7406 images demonstrate that YOLOv11-IIA reaches a precision of 0.763, a recall of 0.819, mAP@50 of 0.857 and mAP@50-95 of 0.699, which achieves clear performance improvements over the original YOLOv11 (mAP@50 improved from 0.485 to 0.857) and other attention-augmented detectors. The model operates stably on an NVIDIA GTX 1050 4GB GPU with an inference speed of 16 FPS for 640 × 640 input images; comprehensive quantitative computational metrics including parameter count, FLOPs and memory consumption will be fully measured in subsequent UAV deployment experiments. The proposed method provides a reliable technical reference for intelligent apple disease monitoring in smart orchard systems. Full article

To address the problems of small fire-spot scale, blurred boundaries, complex backgrounds, and insufficient feature representation of weak targets in Unmanned Aerial Vehicle (UAV)-based early forest fire detection, a YOLOv8n-based forest fire detection model, termed TriRHC-YOLO, is proposed. The model first introduces Reparameterized VGG (RepVGG)Block into the backbone network to enhance the extraction capability of shallow local features. Subsequently, a Hierarchical Feature Attention (HFA) module is designed to collaboratively model fire-spot features from three levels, namely directional structures, local textures, and global semantics, thereby enhancing the network’s capability to discriminate fire targets and suppressing interference from complex forest backgrounds. Finally, a Cross Stage Partial Feature Fusion with Cascade Star Block (C2f-CStar) module is designed to improve the representation capability of the model for local structural information and weak salient fire-spot features under complex backgrounds through cascaded spatial feature reconstruction and a star-shaped multiplicative gating mechanism. In addition, a UAV-specific early forest fire detection dataset is constructed based on the FLAME and FLAME_VISION datasets, and experimental validation is conducted on this dataset. The experimental results show that the proposed TriRHC-YOLO outperforms several classical YOLO algorithms, including YOLO11n, YOLO12, and YOLO26, as well as six advanced YOLO-based improved models. The Recall, mean Average Precision (mAP)@0.5, and mAP@0.5:0.95 reach 0.769, 0.848, and 0.608, respectively. The results of the ablation experiments further verify the effectiveness of the three designed modules. Moreover, the proposed model contains only 3.181 M parameters and achieves 168.251 Frames Per Second (FPS), demonstrating favorable real-time detection capability. Overall, the proposed method can effectively improve the detection accuracy of early weak fire targets and the background suppression capability under complex forest backgrounds, making it suitable for real-time UAV-based forest fire inspection tasks. Full article

The increasing integration of renewable energy sources and smart grid (SG) technologies introduces significant challenges to power system stability due to the inherent variability and uncertainty of electricity generation and demand. While many machine learning and deep learning approaches have been proposed for stability prediction, various studies focus only on predictive performance and offer limited assessment of computational efficiency, statistical significance, and sustainability-related metrics. To address these gaps, this study suggests a hybrid fusion approach that combines Convolutional Neural Networks (CNNs), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF) classifiers through a soft voting strategy for SG stability prediction. The CNN component automatically extracts representative features, while XGBoost and RF contribute complementary classification capabilities, reducing the need for manual feature engineering. In addition to predictive evaluation, a Green AI-oriented benchmarking framework is introduced to evaluate model performance using predictive accuracy, computational runtime, memory consumption, and estimated computational CO₂ emissions associated with model training and inference. The proposed framework is assessed on the UCI SG Stability dataset using stratified cross-validation and statistical significance testing, including Friedman and Nemenyi post hoc analyses. Experimental results demonstrate that the fusion model achieves 97.68% classification accuracy and an AUC of 0.991, exceeding several individual machine learning, deep learning, and ensemble baselines. Statistical analysis shows significant improvements over recurrent deep learning models such as LSTM and GRU. However, the differences from strong tree-based methods, including XGBoost and RF, are not statistically significant. Furthermore, the proposed model reaches a high sustainability score of 0.802, indicating a favorable balance between predictive performance and computational resource requirements. The findings show that the proposed framework is effective and computationally efficient to predict the stability of the smart grid on the UCI benchmark dataset and also serves as a transparent green AI benchmarking methodology for comparative studies in the future. The validation of the approach on real-world smart grid data under noisy, not fully complete, and heterogeneous operating conditions is another interesting research direction for future work. Full article

Vibration monitoring of glass substrate transfer systems is crucial for ensuring the stable operation of Flat Panel Display (FPD) manufacturing equipment. Fabrication defects in the functional nanofiber membrane of flexible vibration sensors can significantly degrade sensing performance and lead to inaccurate monitoring results. To address the challenge of achieving an effective balance between detection accuracy and inference efficiency in such defect-dense scenarios characterized by large variations in defect scale, this paper proposes a novel defect detection model, termed MA-YOLO. The proposed model incorporates four key architectural enhancements: the Multi-level Bidirectional Feature Aggregation Network (MLBAN), the Multi-Receptive Field Adaptive Fusion Module (MRAF), the Morphology-Adaptive Feature Extraction Module (MA-C2f), and the Interactive Dynamic Decoupling Head (IDDH). These components collaboratively improve defect feature extraction, multi-scale feature fusion, and localization performance while maintaining a lightweight architecture and high inference speed. Experimental results on a self-constructed defect dataset demonstrate that MA-YOLO achieves a mean Average Precision (mAP@0.5) of 91.9%, which is a 3.1 percentage point improvement over the baseline model. Moreover, with only 9.15 million parameters and an inference speed of 119.05 FPS, the proposed model exhibits superior overall performance compared with several mainstream and state-of-the-art object detection methods. Full article

Electroencephalography (EEG)-based emotion recognition is an important biosensing technique for affective brain-computer interfaces (BCIs), mental-state assessment, and physiological monitoring. Existing methods often rely on a single spectral descriptor or regular two-dimensional brain maps, which makes it difficult to jointly model local spatial representations, global spatial relations, and temporal dynamics. This paper proposes GLSTNet, a global-local spatial relations and temporal dynamics network for EEG emotion recognition. EEG trials are divided into short windows, from which multi-band spectral features are extracted and arranged into compact spatial maps. The local spatial encoder (LSE) learns local spatial and spatial–spectral representations from these compact multi-band spatial maps. The global spatial-relation encoder (GSRE) models long-range spatial relations between non-adjacent electrodes using a Pearson correlation prior and a learnable residual adjacency matrix. After local and global representations are integrated through gated fusion, the temporal dynamics encoder (TDE) models consecutive EEG windows using a gated recurrent unit with temporal attention. Comprehensive validation is conducted on two public EEG emotion datasets, the Database for Emotion Analysis using Physiological Signals (DEAP) and the SJTU Emotion EEG Dataset (SEED). In the subject-dependence setting, GLSTNet achieves 93.50 ± 3.22% accuracy for valence and 93.79 ± 3.64% accuracy for arousal on DEAP, and 92.48 ± 3.30% accuracy on SEED. In the subject-independence setting with target-subject calibration, GLSTNet obtains 75.61 ± 6.19% and 79.57 ± 5.99% accuracy for DEAP valence and arousal, respectively, and 88.22 ± 4.70% accuracy on SEED. These results indicate that integrating global-local spatial relations with temporal dynamics provides an effective representation strategy for EEG-based emotion recognition. Full article

Advanced driver assistance systems (ADAS) require a comprehensive understanding of multiple tasks related to the physical and mental states of drivers and traffic situations. Existing ADAS studies perform driver emotion recognition (DER), driver behavior recognition (DBR), traffic context recognition (TCR), and vehicle behavior recognition (VBR) using models designed based on single-task learning, thereby failing to reflect the interactions among tasks in real driving environments. This paper proposes perception and recognition with inter-modal selective multi-task learning (PRISM-MTL), an integrated multimodal and multi-task learning framework that jointly recognizes DER, DBR, TCR, and VBR. The proposed PRISM-MTL consists of a hierarchical stage-wise attention network (HSA-Net)-based multimodal encoder that extracts spatial features from heterogeneous multimodal inputs and task-specific modality fusion (TSMF), which selectively learns effective modality information for each task. This design addresses negative transfer, a key challenge in multi-task learning. In the multimodal encoder, HSA-Net extracts visual modality tokens that emphasize global structural patterns and key spatial regions from multi-view images, while Token-SE generates joint modality tokens that reflect the spatial configuration of joint data. TSMF generates task-specific fusion features that selectively emphasize the modality cues for each task. The generated task-specific fusion features are summarized through temporal mean pooling, and final predictions of driver states and traffic situations are produced by each task head. Experimental results show that the proposed PRISM-MTL achieves state-of-the-art performance on the public AIDE database, with an mAcc of 86.25% ± 0.35 for multi-task recognition of driver states and traffic situations. Full article

Concealed fires within the insulation layers of buildings, such as cold storage facilities and cinemas, present a serious fire hazard because heat generated by electrical faults can accumulate behind protective panels before ignition and then spread rapidly once combustion begins. Conventional fire detection methods have limited capability to identify these hidden thermal abnormalities at the pre-ignition stage. To address this problem, this paper proposes a deep learning method, called the Multi-Scale Cross-Modal Fusion Network (MSCMFNet), that uses multispectral images to identify abnormal heat sources beneath insulation layers before visible combustion occurs. A standardized experimental platform was developed to accurately simulate subsurface heat sources within the pre-ignition temperature range of insulation materials. Instead of relying on fixed temperature thresholds, the proposed method learns the characteristic spectral patterns produced by hidden heating. It extracts information from different spectral bands, combines these complementary features, and verifies the persistence of detected heat sources over time to reduce false alarms caused by non-fire disturbances. Experimental results demonstrate that the proposed method can effectively detect concealed thermal anomalies before ignition, providing reliable early warning and offering a promising approach to improving fire safety in buildings that make extensive use of insulation materials. Full article