Surface small defects are often missed and incorrectly detected due to their small quantity and unapparent visual features. A method named CSYOLOv3, which is based on CutMix and YOLOv3, is proposed to solve such a problem. First, a four-image CutMix method is used to increase the small-defect quantity, and the process is dynamically adjusted based on the beta distribution. Then, the classic YOLOv3 is improved to detect small defects accurately. The shallow and large feature maps are split, and several of them are merged with the feature maps of the predicted branch to preserve the shallow features. The loss function of YOLOv3 is optimized and weighted to improve the attention to small defects. Finally, this method is used to detect 512 × 512 pixel images under RTX 2060Ti GPU, which can reach the speed of 14.09 frame/s, and the mAP is 71.80%, which is 5%-10% higher than that of other methods. For small defects below 64 × 64 pixels, the mAP of the method reaches 64.15%, which is 14% higher than that of YOLOv3-GIoU. The surface defects of the workpiece can be effectively detected by the proposed method, and the performance in detecting small defects is significantly improved. © 2021, Editorial Department of Journal of Southeast University. All right reserved.
