For the problem of transmission line defect detection, this paper designs the improved YOLOv4-Tiny-EL model and YOLOv4-Tiny-EM model based on YOLOv4-Tiny. Their mean Average Precision (mAP) has increased by 2.31% and 3.53% respectively compared with the original model. To balance detection accuracy and speed, the YOLOv4-Tiny-EL model is selected for deployment in the edge device ZYNQ7020. In the model inference optimization stage, the Batch Normalization (BN) layer fusion method is adopted to reduce the computational load and storage overhead, ...
