To address the problems of inefficiency and lack of robustness to the environment in the current traditional image processing algorithms for tread surface defect detection, this paper proposes an improved tread surface defect detection method based on the Faster RCNN. The improved network first uses Resnet50 as the feature extraction network, and adds a self-attention mechanism to the feature fusion output part of the Feature Pyramid Network to enhance the detection ability of the detection network for small defects, and finally uses the K-means++ clustering algorithm to cluster the anchor frames of the tread defect dataset, and uses the clustering results to customize anchor frames that are more suitable for wheel-to-tread defects.The experimental results show that the improved Faster RCNN network has an average detection speed of 68 ms, an average accuracy (mAP) of 97.3% and an accuracy of 39.3% for the detection of small target defects.