Early detection of forest fires is critical but remains challenging due to small multi-scale targets in complex backgrounds and tight computing constraints on edge devices. This paper presents SFGI-YOLO, a lightweight real-time detection system optimized for Raspberry Pi 5. Built on YOLO11, it introduces a P2 detection head for fine-grained features, a Feature Enhancement Module (FEM) for low-contrast targets, GhostConv for light weight, and a C3k2_IDC module for multi-scale extraction. A LoG operator is embedded in preprocessing to suppress noise and enhance weak edges. The model is converted to OpenVINO and INT8-quantized for efficient edge deployment. Experiments on Raspberry Pi 5 show that the system achieves 93.5
Unmanned aerial vehicles (UAVs) offer numerous advantages and are widely utilized across various domains. However, illegal UAV activities pose significant security threats to public safety and infrastructure. Due to their robust anti-interference capabilities, infrared images play a crucial role in UAV detection. Nevertheless, existing methods often suffer from a mismatch between multi-scale feature representation and the receptive field. To address this issue, this paper proposes a novel infrared UAV detection model named RFMS-RTDETR. By integrating a large receptive field network with a multi-scale feature extraction mechanism, the proposed model effectively captures global contextual information and multi-scale features while mitigating the loss of fine-grained details typically caused by large receptive fields. Specifically, we design the MS-RFAM module to effectively expand the receptive field and extract multi-scale features. Additionally, a DHFE module is introduced to preserve the fine features of small targets. Comprehensive experiments conducted on our custom-built MSIR-UAV dataset demonstrate that the RFMS-RTDETR model achieves superior performance in infrared UAV detection tasks, providing an effective solution to the mismatch between scale feature representation and the receptive field. The project page: https://github.com/wukaixuan121/RFMS-RTDETR
Infrared object detection remains challenging due to low thermal contrast, weak structural cues, background interference, and the limited spatial extent of small or distant targets. These factors often degrade feature discriminability and lead to unstable target localization. To address these challenges, this paper proposes TheraRegNet, a thermal feature regulation network for infrared object detection. Specifically, a Channel-aware Local Importance Attention (CLI) module is introduced to enhance discriminative thermal responses by jointly modeling channel dependencies and local spatial importance. A Gated Selective Convolution Bottleneck (GSCB) is designed to suppress background-dominated feature propagation through adaptive spatial gating. Furthermore, a Multi-scale Thermal Contrast Attention (MTCA) module improves the perception of weak targets by exploiting multi-scale thermal saliency and contrast information. These components are integrated into a unified Thermal Feature Regulation (TFR) mechanism to progressively refine infrared feature representations. Experiments on the DroneVehicle, KAIST, and TTG datasets demonstrate that TheraRegNet consistently improves detection performance under low-contrast and cluttered conditions, achieving mAP@50 scores of 85.2
Dense aphid detection in complex agricultural scenes is challenging due to small object sizes, severe occlusions, and background clutter. This paper proposes RT-DETR-ASH, an enhanced real-time detection framework based on RT-DETR. Specifically, an Aggregated Attention (AA) module is incorporated into the backbone to improve feature extraction by integrating local and global contextual information. An Adaptive Sparse Self-Attention (ASSA) mechanism is further introduced into the Transformer encoder to emphasize informative regions while reducing computational redundancy. In addition, a hierarchical feature pyramid network with Context Anchor Attention (CAA-HSFPN) is designed to strengthen multi-scale feature aggregation. Experimental results on a public aphid dataset show that RT-DETR-ASH consistently outperforms the baseline RT-DETR in terms of mAP@0.5 and mAP@0.5:0.95, while maintaining competitive computational efficiency. These results demonstrate the effectiveness of the proposed framework for dense small-object detection in complex environments.
UAV remote sensing imagery plays a critical role in applications such as traffic monitoring and urban management. However, small object detection in this domain remains highly challenging due to low resolution, severe occlusion, and cluttered backgrounds. Existing attention-based methods primarily emphasize spatial information while neglecting channel interdependencies, limiting their ability to reconstruct discriminative features from sparse pixels. To address this issue, this study proposes an Attention-based Appearance Reconstruction Module (AARM) that employs channel attention to dynamically capture global dependencies and reconstruct rich, distinctive appearance representations. Based on this core module, several architectural variants are developed, including C2AARM for deep feature fusion, C2AARM-Lite for computational efficiency, C2AARM-Adaptive for dynamic feature balancing, and AARMSPPF for multi-scale contextual encoding. Extensive experiments on UAV benchmark datasets AI-TOD, VisDrone, and UAVDT demonstrate that the proposed framework achieves stable improvements in detection accuracy while maintaining high inference speed (173.2 FPS). In AI-TOD, the best performing model obtains 55.7 https://github.com/wtc0214/AARM
Food image classification supports dietary monitoring, personalised nutrition and intelligent food-related devices, but the rapid evolution of convolutional neural network (CNN) architectures, transfer learning strategies and preprocessing pipelines makes it difficult to select models that balance predictive performance and deployment efficiency. This paper presents a structured literature review of CNN-based food image classification studies published between 2020 and 2025. Following a predefined protocol inspired by Kitchenham and Brereton, 40 peer-reviewed studies were selected from Google Scholar, IEEE Xplore and Scopus, and analysed in terms of datasets, architectures, transfer learning, preprocessing, imbalance handling, evaluation metrics, computational efficiency, explainability, reproducibility and reported limitations. ResNet variants remain the most frequently used backbone, whereas EfficientNet and MobileNet families are repeatedly reported as promising candidates for resource-constrained deployment. However, cross-study comparison is limited by heterogeneous datasets, train/test splits, input resolutions, augmentation policies, hardware, metrics and incomplete efficiency reporting. Food-101 dominates the benchmark landscape, creating a benchmark-dependent evidence base that may not generalise to regional cuisines or real-world acquisition conditions. We therefore interpret architecture recommendations as Pareto-type trade-offs rather than as evidence of a single optimal model. The review identifies persistent gaps in statistical validation, class imbalance treatment, explainability, reproducibility, public code availability and deployment-oriented evaluation. These findings provide practical guidance for embedded food recognition systems while clarifying the methodological limits of the current evidence base.
Real-time object detection requires an effective balance between detection accuracy and computational efficiency, particularly for deployment on resource-constrained devices. Although the YOLO family of detectors has achieved remarkable success in high-speed object detection, its backbone architectures often remain parameter-intensive, limiting their applicability in embedded and real-time systems. In this paper, we propose Extension Enabled Network: Exen-net, a lightweight YOLO backbone architecture that significantly reduces model complexity while preserving competitive detection performance. The proposed design decreases feature channel dimensions and enhances feature extraction through bidirectional top-down and bottom-up information flow, enabling efficient feature reuse with minimal architectural modifications. The proposed backbone has been validated on two representative YOLO generations (YOLOv5 and YOLOv8), demonstrating that the proposed design is applicable across multiple YOLO architectures with seamless integration into different versions and model scales. To evaluate its effectiveness, Exen-net was integrated into YOLOv5 and YOLOv8 and evaluated on the MS COCO dataset. The proposed backbone reduces backbone parameters by 36–52
Skin cancers, particularly melanoma, are common and potentially deadly diseases due to their increasing incidence and complexity of early diagnosis. Although dermoscopy, combined with deep learning approaches, has improved the detection of skin cancers, the precise segmentation and classification of skin lesions remain challenging due to their complex nature, characterized by variations in appearance and texture. This study proposes a new automated system for skin lesion segmentation and classification, integrating Efficient Channel Attention (ECA) mechanisms into a residual UNet architecture called ARCA-Net. This architecture improves feature extraction by replacing traditional residual blocks with attention residual blocks, allowing the model to focus on critical lesion features by exploiting inter-channel relationships while improving computational efficiency, thereby outperforming benchmark models. Furthermore, our method was strengthened by the integration of multiclass classification using multiple classification models. The model performance was evaluated on the ISIC 2016, ISIC 2017, and ISIC 2018 datasets, and further validated on the external PH2 dataset, where it demonstrated strong generalization and outperformed state-of-the-art methods. The proposed method demonstrated its ability to accurately detect lesion contours and preserve essential discriminative features for classification, thereby enhancing the overall performance of the automated medical diagnostic system.
Deepfake videos present a major challenge to the authenticity of digital media, creating a need for accurate and interpretable detection methods. This paper proposes DenseXplainNet, an explainability-guided local–global fusion framework for video deepfake detection that combines a convolutional neural network (CNN) for capturing local manipulation artifacts with a Vision Transformer (ViT) for modeling global contextual dependencies. Unlike conventional approaches that employ explainable artificial intelligence (XAI) solely for post-hoc interpretation, the proposed framework integrates Grad-CAM and Attention Rollout directly into the inference process. The generated explanation maps are transformed into quantitative clarity scores that guide frame selection, prediction weighting, and video-level aggregation. Experiments conducted using FaceForensics++ (C23) dataset that demonstrates DenseXplainNet outperforms standalone CNN, ViT, and conventional CNN–ViT fusion models, achieving an accuracy of 96.70
Semi-supervised medical image segmentation reduces dependence on dense annotations, yet existing CNN–Mamba designs rarely learn which branch should dominate at different spatial scales. We propose MCSD-Net, a heterogeneous dual-student framework that pairs a structure-oriented Mamba student with a detail-oriented ConvNeXt student. Its Cross-Scale Selective Fusion (CS2F) module uses input-dependent gates to combine aligned decoder features and branch logits. Cross-pseudo-supervision exploits unlabeled data, symmetric Kullback–Leibler regularization stabilizes predictions on labeled data, and boundary-aware supervision refines uncertain contours. The same design supports 2D and 3D segmentation with a fixed fusion coefficient and dataset-specific boundary constraints. Experiments on ACDC, Left Atrium, and BraTS19 under two low-label ratios show competitive Dice, Jaccard, ASD, and 95HD results. The gains are strongest on several overlap and boundary metrics, although MCSD-Net does not lead every surface-distance measure. Ablation and efficiency analyses confirm the contributions of heterogeneous pairing, CS2F, and the selected optimization components. Failure cases reveal remaining sensitivity to ambiguous boundaries and very small lesions.
Ciphertext cropping attacks are widely used to evaluate the robustness of chaos-based image encryption systems. Post-processing enhancement of degraded decrypted images after ciphertext loss is important for maintaining the visual usability of encrypted images under non-ideal transmission and incomplete storage conditions. Taking a permutation-diffusion encryption algorithm based on the generalized Fibonacci–Lucas transform (FLT) and bidirectional diffusion with bitwise exclusive-OR (XOR) operations as a concrete implementation, this paper investigates enhancement under ciphertext cropping. After standard decryption, ciphertext cropping produces spatially redistributed, structure-associated degradation rather than a local occlusion. The resulting degradation contains an impulse-like extreme-valued component and a large-error tail. Based on this observation, a two-stage cascaded post-processing method is proposed by combining an adaptive dynamically weighted median filter (ADWMF) and an adaptive switching trimmed vector median filter (ASTVMF). ADWMF suppresses dominant abnormal pixels, and ASTVMF refines the remaining degradation. On grayscale versions of eight 512× 512 standard test images under three cropping shapes and three cropping ratios, the proposed method achieves an average peak signal-to-noise ratio (PSNR) of 30.90 dB and an average structural similarity index (SSIM) of 0.8972, outperforming Adaptive Median, the strongest general-purpose baseline, by 6.70 dB and 0.2152. Using paired degraded and reference images for training, a task-specific denoising convolutional neural network (DnCNN) and a same-resolution model based on the very deep super-resolution (VDSR) architecture achieve higher scores. The proposed method requires no task-specific training. Gaussian and Poisson contamination, controlled-variant, and independent-algorithm tests indicate that achievable gains depend on post-decryption local statistics.
To address background interference and viewpoint tilt in complex industrial inspection, we propose an automatic pointer meter reading method integrating FSS-DeepLabV3+ with spatially constrained OCR. YOLO26s first localizes meter regions, followed by perspective correction to reduce viewpoint distortion. Within a lightweight DeepLabV3+ framework using MobileNetV4, a cross-layer interactive feature fusion (CLIFF) module aligns low- and high-level features to enhance shallow details. A slender-target perception and context enhancement (SPCE) module strengthens local context modeling and the perception of slender directional structures, including pointers and scale marks. At the output stage, a semantic difference modulation (SDM) module compensates for semantic discrepancies in initial segmentation and refines boundaries. Spatially constrained OCR is then combined with angle mapping to obtain interpretable readings. Experiments on a self-constructed pointer meter dataset show that FSS-DeepLabV3+ achieves high segmentation accuracy and inference efficiency. Compared with DeepLabV3+ using the same MobileNetV4 backbone, it improves mIoU, ACC, ACC-Class, and fwIoU by 8.36, 0.45, 0.04, and 0.71 percentage points, respectively. In the final reading stage, the mean absolute error decreases from 0.105 to 0.059, and the average fiducial error decreases from 0.756
To address feature obscuration, difficulty in detecting small targets, and limited real-time performance in nighttime driver-drowsiness detection, this paper proposes Low-Light Eye-State YOLO (LES-YOLO), a lightweight detector based on an improved YOLOv11-n architecture. CondConv2D is incorporated into the backbone to enhance the representation of ocular texture features under low illumination; GhostConv is used to reduce computational redundancy and improve inference efficiency; and a Swin Transformer module with shifted-window self-attention is introduced into the detection head to strengthen small-target feature fusion under occlusion. In addition, a dynamic drowsiness-assessment model based on the PERCLOS criterion is developed to quantify drowsiness through temporal analysis of eyelid-closure patterns. Experiments on a mixed nighttime dataset show that LES-YOLO achieves an mAP@0.5 of 92.5
In recent years, multi-modal fusion has emerged as a prominent research direction in gait recognition to overcome the inherent limitations of uni-modal methods. While existing approaches typically employ sparse skeleton joints as the auxiliary modality, optical flow captures richer dense motion information, making it a more promising choice. In this paper, we propose SSM-Gait, a novel multi-modal framework that effectively integrates the complementary advantages of optical flow and silhouette sequences for a robust gait representation. Unlike existing multi-modal fusion methods, our approach places greater emphasis on isolating modality-specific noise while preserving discriminative features. We propose the Modality-Aware Spatial Fusion Module (MA-SFM). It utilizes a dynamic selection mechanism based on selective state spaces and a shared-private dual-channel architecture to explicitly decouple features into cross-modal consensus and modality-specific information, effectively isolating noise while preserving distinctive features. Furthermore, to address gait cycle discrepancies caused by varying walking frequencies, we design the Modality-Aware Temporal Fusion Module (MA-TFM). By combining a temporal pyramid structure with a weight-sharing mechanism, this module performs multi-scale temporal fusion, thereby enhancing the model’s robustness. Extensive experiments on the CASIA-B and GREW datasets demonstrate average rank-1 accuracies of 96.7
Recent advancements in 3D Gaussian-based scene rendering have demonstrated significant potential for efficient neural scene representation. However, accurately capturing geometric boundaries and fine details in complex scenes during optimization remains challenging. Specifically, two key issues persist: (1) the blurring of edge details in high-contrast objects and (2) the loss of texture details in small, distant objects with sparse point cloud distributions. To address these challenges, we propose a method that introduces two key improvements: an anchor re-growing module that dynamically increases neural Gaussian density in high-gradient regions using edge-aware optimization, and the Segment Anything Model (SAM) for generating accurate object segmentation masks to guide segmentation-based loss computation. Our collaborative optimization strategy significantly enhances boundary clarity and texture fidelity in 3D Gaussian rendering, particularly in complex scenes. Evaluated on multiple datasets, our method demonstrates substantial improvements in rendering quality, achieving a 0.29 improvement in PSNR on the Tanks Temples dataset compared to the baseline method. Project page: https://github.com/Mazycity57/Edge-GS
The preparation of expert-annotated computed tomography (CT) datasets is an important prerequisite for developing pulmonary nodule detection models. This study presents a transformer-assisted pulmonary nodule pre-annotation prototype that generates editable candidate bounding boxes for subsequent expert review. The prototype integrates CT preprocessing, a DETR-style detection component, annotation review and data export, while retaining expert control over the final annotations. The detection component was retrospectively evaluated on LUNA16 using the official ten-fold partition and compared with representative CNN- and DETR-based detectors under a common implementation framework. The effects of ImageNet and pulmonary-nodule-specific PN9 pretraining were also evaluated. The proposed detector achieved an overall candidate sensitivity of 97.2
Small-sample visible-light weld inspection is difficult because circular metal surfaces introduce specular reflection, low local contrast, weld-like background texture, and elongated defects that may occupy only a small region after image resizing. This paper proposes GED-YOLOv5, a task-adapted YOLOv5 detector for circular pipe weld defect detection. The model combines an improved Gather-and-Distribute (GD) neck for cross-scale feature propagation, a C3_DSC backbone block for elongated defect boundaries, and Efficient Channel Attention (ECA) for defect-related channel selection. Source-domain pre-training is further used to reduce overfitting under limited target-domain samples. On the source-domain weld dataset, GED-YOLOv5 increases mAP@0.5 from 88.6 53.15 ± 3.20 51.17 ± 2.97 19.07 ± 0.68
Intelligent video applications require compressed streams that support both human viewing and downstream machine vision. Existing neural video compression methods mainly optimize pixel-level reconstruction and may not preserve task-relevant semantics under high compression. To address this limitation, we propose a joint human–machine optimized neural video compression framework based on task-aware conditional priors. Specifically, semantic representations from downstream vision networks are exploited to enhance semantic preservation during training, while the proposed framework maintains a unified bitstream without additional semantic transmission. A feature similarity supervision mechanism is further introduced to maintain semantic consistency between original and reconstructed representations. Experimental results indicate that the proposed framework provides competitive rate–distortion performance and improves downstream detection and tracking performance in the evaluated settings. Compared with a human-vision-oriented baseline, mean average precision at an intersection-over-union threshold of 0.50 (mAP _50 ) is improved by up to 2.0 percentage points, and multi-object tracking performance gains up to 1.024 Higher Order Tracking Accuracy (HOTA) points at the evaluated operating settings.
High-resolution displays require substantial image data processing and frequent memory access, which increases power consumption. To address these issues, adaptive image data approximation methods based on local image characteristics and the human visual system are proposed. It approximates image data by considering pixel-background differences, spatial frequency, tailing zeros, and brightness sensitivity. It combines these HVS-based approaches to improve image data reduction while preserving perceptual image quality. Experimental results using the Kodak dataset show an average image data reduction of 43.77
Bird’s-eye-view (BEV)-based 3D object detection integrates complementary semantic cues from cameras and geometric information from LiDAR. However, differences among camera viewpoints and spatially varying modality reliability can introduce cross-view inconsistencies and cross-modal conflicts in BEV representations. To address these challenges, this paper proposes VQGate3D, a multimodal BEV-based 3D detection framework. In the camera branch, learnable view embeddings encode camera identities to alleviate view-dependent feature discrepancies and improve cross-view consistency. For multimodal fusion, distance and LiDAR point-density priors are constructed as BEV-space reliability cues, and a lightweight gating network predicts grid-wise camera–LiDAR fusion weights for spatially adaptive fusion. The fused features are further refined using channel attention and forwarded to the BEV encoder and detection head. On the nuScenes validation split, VQGate3D achieves 69.8
