Improved YOLOv7-Based Method for Bird Detection in Substation Scenarios

The study presents an improved YOLOv7-based bird detection method for real-time bird activity monitoring during substation patrols. The proposed approach optimizes the network's backbone by incorporating PConv from the lightweight network FasterNet and coordinate attention (CA) mechanism, achieving a balance between recognition speed and precision in bird detection tasks. Additionally, the task-specific context decoupling (TSCODE) mechanism is adopted to enhance overall detection performance, and the Wise-IoU bounding box regression loss function is introduced, which utilizes a dynamic gradient gain allocation strategy to obtain optimal results. The performance of the improved network is compared to that of the original YOLOv7 network on a self-built bird dataset, demonstrating a mean average precision increase from 84.2% to 89.6% and a detection speed improvement of approximately 28.5%. These results indicate the proposed algorithm's advancement in both detection efficiency and accuracy.