Surface Defect Detection based on Improved YOLOv3-Tiny Algorithm

Abstract

To take into account both accuracy and real-time performance in surface defect detection, we propose a new surface defect detection algorithm based on YOLOv3-Tiny. The algorithm first adds a YOLO layer that fuses shallow and deep features on the basis of YOLOv3-Tiny, to enhance the capabilities of microscopic defect detection through multi-scale features fusion. And the hybrid attention mechanism module, named SE-C, is employed before every YOLO layer. The SE-C module can decrease the weight of irrelevant background's features while improving the weight of the defect's features, it will improve the algorithm's robustness and accuracy. Finally, the algorithm re-clusters the anchor boxes based on K-means in each dataset. The experimental results reveal the improved algorithm has a good trade-off between the accuracy and the speed of defect detection, especially in easily confused background. More importantly, the algorithm can also be applied to other object detection on similar scenes.