Power Transmission Line Foreign Object Detection based on Improved YOLOv3 and Deployed to the Chip

The application of object detection is becoming more and more widely in various fields, including the power industry, of course. And YOLOv3 is one of the most popular algorithms in the field of object detection owing to its high performance and efficiency. However, the conventional YOLOv3 algorithm is still too heavy to deploy on mobile or embedded platforms. Consequently, this paper proposes a method to improve the YOLOv3 thus it can be easily deployed to embedded platforms without losing performance. First, substitutes the backbone of YOLOv3, i.e. Darknet53 for MobileNet, which has been proven to be a very efficiency framework for lightweight network. Second, there are numerous redundancies in the detection heads of YOLOv3 and will take a lot of time in the inference process, so we prune the detection heads to a dead-simple structure. Various experiments on our own Power Transmission Line datasets verify our method has state-of-the-art performance while can meet the requirements for deployment to the mobile platforms.