Crack Detection of Track Plate Based on YOLO

In order to solve the problem of track plate cracks detection, night low light environments and crack growth irregularity, this paper proposes an improved YOLO (you only look once). In view of the problem that the YOLO has poor detection effect on small targets and overlapping targets, this paper refers to the RPN-RCN, removes the fully connected layer based on the original YOLO network, Anchor Box and a new convolution layer added to it. The improved network is used to superimpose the lower characteristics of shallow convolutional structure in the original YOLO and the abstract semantic features extracted from the deep convolutional network. Experiments show that the improved YOLO effectively enhances the detection ability of small targets and overlapping targets, and improves the detection accuracy of track plate cracks. In addition, the real-time performance of the improved algorithm also meets the engineering requirements.