Conveyor belt damage detection method based on improved YOLOv4

Abstract

In order to solve the problems of low detection precision, slow detection speed and lack of damage detection for small areas in existing conveyor belt damage detection methods, a conveyor belt damage detection method based on improved YOLOv4 is proposed. Based on YOLOv4, this method improves the PANet path fusion network part, increases the fusion with the shallow characteristic layer, increases the fusion of the original 3 scales of the characteristic layer to 4 scales, improves the characteristic extraction capability of the model for conveyor belt damage, and improves detection precision. The number of convolutions after fusion of each characteristic layer in the PANet part is reduced from 5 to 3 so as to reduce the amount of calculation and improve the detection speed. The conveyor belt damage images are labeled and input into the improved YOLOv4 model for training and testing. The experimental results show that the conveyor belt damage detection method based on improved YOLOv4 has a fast loss convergence speed and has a good model training effect. Based on improved YOLOv4 conveyor belt damage detection method, the average precision of the conveyor belt tear, surface wear and surface defect detection has reached 96.86%, and the detection speed has reached 20.66 frames/s. Compared with YOLOv4, YOLOv3 and Faster-RCNN, the average precision has increased by 1.4%, 6.35% and 2.16% respectively, and the detection speed has increased by 2.39, 2.34 and 15.25 frames/s respectively. Compared with YOLOv4, the conveyor belt damage detection method based on improved YOLOv4 has higher detection precision and better detection effect for small areas damages.

Keywords

• belt conveyor; conveyor belt damage detection; yolov4; deep learning; panet; characteristic layer fusion