Against the backdrop of escalating global warming and increasingly frequent forest fires, traditional fire monitoring technologies face challenges of poor scene adaptability and a scarcity of data resources. To address these issues, this paper proposes a real-time uncrewed aerial vehicle (UAV) forest fire smoke detection method that combines generative diffusion models with lightweight edge computing. First, the method utilizes a pre-trained Stable Diffusion model fine-tuned via Low-Rank Adaptation (LoRA) to generate a forest fire smoke image dataset (DMFFSD) encompassing various scenarios and climatic conditions, thereby resolving the issues of data scarcity and insufficient sample diversity. Second, to improve real-time detection performance, we propose a lightweight cross-spatial feature enhancement network (CSFENet), which enhances the recognition capability for smoke and flame targets with blurred boundaries and scale variations. Additionally, an efficient cross-spatial path aggregation network (ECS-PAN) architecture is designed to achieve fast and efficient multi-scale feature fusion, ensuring the model remains lightweight. Extensive experimental results demonstrate that, on the same test set, mainstream detection models trained using DMFFSD achieve a 24
In industrial production, aluminum surface defect detection is crucial, and deep learning is the current mainstream approach. However, it often suffers from high missed detection rates, high computational cost, and excessive model parameters. To address these issues, we propose a lightweight real-time detection method. First, we preprocess the original dataset for data augmentation. Second, by integrating dilated convolutions, structural reparameterization, and ADown modules into an LMS-Backbone, the model reduces computational complexity and parameters while enabling efficient multi-scale feature extraction. Then, the Efficient Cascaded Ghost-Transformer Encoder Layer (ECG-TEL) focuses on defect regions via a multi-level attention mechanism. Subsequently, the Content-Guided Augmented Attention Fusion (CGAAFusion) module strengthens defect feature fusion and discrimination. Finally, the decoder and prediction head perform localization and classification. Experiments on our self-constructed aluminum profile dataset show that LW-RTDETR reduces FLOPs by 50.5
The precise and efficient sorting of coal and gangue is a crucial link in realizing green coal production. Addressing the challenge that existing algorithms struggle to balance accuracy and efficiency due to extreme underground illumination, physical occlusion, and background interference, this paper develops a lightweight detection model named LWD-YOLO based on the YOLOv11 framework. Its task-oriented design comprises: (1) an LSFE module that integrates Ghost-style feature generation, parameter-free SimAM calibration, and a conditional residual path to balance feature representation and computational cost; (2) the adoption of the existing WIoUv3 loss to provide dynamic non-monotonic focusing during bounding-box regression; and (3) a DyHead-inspired PTS-Head that retains task and spatial refinement while omitting the original cross-scale attention branch for the fixed-camera conveyor-belt setting. The contribution lies in this application-specific adaptation, coordinated integration, and controlled validation rather than in claiming that the constituent operators are individually new. Experiments on a self-built coal and gangue dataset demonstrate that LWD-YOLO achieves an mAP_50 of 97.2
To address the trade-off between detection accuracy and efficiency in steel surface defect detection, this paper focuses on the computational redundancy in RT-DETR and proposes an efficient lightweight detector, termed GER-DETR. First, a GELAN-SimAM-based feature representation structure is developed by integrating the existing GELAN architecture with the parameter-free SimAM attention mechanism to replace RepC3 in the neck of RT-DETR. Combined with the existing ADown module, it further optimizes the backbone network, thereby significantly reducing model parameters and computational cost, while maintaining effective feature representation. Second, the ESDI module is introduced to enhance the feature fusion network, strengthening the model’s ability to effectively fuse and exploit multi-scale features, thereby improving detection accuracy. Third, the RFD module is incorporated to improve shallow downsampling in the backbone and downsampling operations in neck feature fusion, enabling the preservation of richer information and enhancing feature robustness. Finally, Powerful-IoU v2 is adopted to improve bounding box regression accuracy and enhance defect localization performance. On the NEU-DET dataset, GER-DETR attains 78.65
Infrared ship detection is essential for maritime perception, but weak texture, blurred boundaries, scale variation, and complex thermal clutter make it difficult for lightweight detectors to balance accuracy and deployment efficiency. This paper proposes TSCDet, which introduces task-specific thermal–structural cue learning through the thermal–structural cue modulation (TSCM) block. TSCM couples fixed vertical finite differences of learned features with horizontal continuity scanning and residual response preservation, while asymmetric feature modulation provides stronger positive enhancement and restrained negative suppression for weak or discontinuous targets. Experiments on the Raytron infrared ship dataset show that TSCDet achieves 90.727
