Detection of road crack defects based on improved YOLOv5s model

Abstract

In order to meet the requirements of road crack defect detection accuracy and detection speed, a road crack defect detection algorithm based on the improved YOLOv5s network model is proposed. First, the Backbone area of the YOLOv5s model is improved, and the self-correcting convolution SCConv module is embedded in the Backbone area as a feature extraction network, which can better integrate multi-scale feature information; secondly, the Neck area of the YOLOv5s model is improved. SE-Net attention The module is added to the Neck area to improve the retention of defect feature information and strengthen the learning of feature information; finally, the Prediction area of the YOLOv5s model is improved to delete the 80×80 feature map branch suitable for detecting objects of smaller size, thereby reducing the complexity of the model and improving the Detection speed, the improved YOLOv5s network model is obtained. Using the improved YOLOv5s model to train and test on the road crack defect dataset, the experiments show that the improved YOLOv5s network models precision (P) is increased by 4.91%, the recall rate (R) is increased by 4.79%, and the average precision rate (mAP) is increased by 5.42%, the model size is reduced by 1.04MB, indicating that the improved YOLOv5s model has improved detection accuracy and detection speed.