YOLO Based Bridge Surface Defect Detection Using Decoupled Prediction

Authors:

Abstract
Bridge surface defect detection is an essential method for evaluating the bridge quality and subsequent repair. Convolutional neural network based intelligent object detectors have been powerful for defect detection in recent years. In this paper, we propose a model based on the YOLO detector for detection. We use the convolutional block attention module, decoupled prediction head, and focal loss function to improve performance. To verify our proposed model, we perform experiments on an open bridge surface defect dataset, and our model can obtain 90.3% mAP50 and 72.8% mAP75. The detection precision of our network is higher than the original YOLOv5.
Funding Information
  • Research and Development
  • National Natural Science Foundation of China