基于改进YOLOv5s的高压柜目标检测算法OA

To address the issues of background interference in power equipment images and inaccurate detection of keyholes and indicator lights in robotic vision,this paper proposes an improved YOLOv5s-based object detection algo-rithm for high-voltage switchgear.By replacing standard convolutions with deformable linear convolutions,the algorithm enhances feature extraction for targets such as keyholes.Additionally,a context anchor attention module is introduced to capture long-range contextual information,thereby improving the model's feature representation.To further improve localization accuracy under uneven data quality,a bounding box loss function based on a dynamic non-monotonic focus-ing mechanism is adopted.Experimental results on a self-built substation high-voltage switchgear dataset show that the improved model outperforms the original YOLOv5s.The mAP@50 reaches 80.01%,an improvement of 1.2%,while the mAP@50:0.95 reaches 43.67%,an improvement of 12.4%.These results demonstrate that the proposed method offers better adaptability and higher detection accuracy for high-voltage switchgear detection tasks.