Lightweight Bird's Nest Location Recognition Method Based on YOLOv4-tiny

Abstract

Activity around transmission lines after bird nesting poses a threat to the safe operation of the power grid. Identifying bird nests in transmission lines and towers is the key to preventing wading bird failures. With the development and change of deep learning and patrol mode, the accuracy and speed of bird's nest recognition algorithm are challenged. To address this challenge, this paper presents a bird's nest location recognition method based on YOLOv4-tiny. YOLOv4-tiny, a lightweight version of YOLOv4, consists of a backbone network, FPN, and YOLO head with 13×13 and 26×262 prediction scales. The YOLOv4-tiny algorithm extracts the features of the input bird's nest image data to realize the positioning recognition of the bird's nest. This paper chooses the open source Darknet framework as a tool, which combines Loss and AP curves in the visualization training process. The results show that the model accuracy of YOLOv4-tiny algorithm can reach 77.20%, and the inference speed on the edge computing platform Jetson Xavier NX can reach up to 50.39 FPS. The method proposed in this paper can meet the requirements of real-time target detection in both speed and accuracy, and has a good application prospect.