Surface defect detection of planar optical components based on OPT-YOLO

Abstract

Surface defects on optical components significantly impact imaging quality, durability, and reliability. Traditional optical and manual detection methods are inefficient and lack accuracy. Therefore, making machine vision is a crucial advancement for surface defect detection. In this article, an imaging system was constructed to generate a proprietary dataset of self-made optical component surface defects and an innovative OPT-YOLO algorithm was developed for detecting surface defects in planar optical components. The RepViTSEBlock and EMA modules were integrated with the C2f backbone network in OPT-YOLO, streamlining weights and enhancing feature extraction. To improve model expression, the MPCA attention mechanism was incorporated into the tail of backbone. Addressing background interference on defect features, ODConv was introduced into the Neck structure, replacing the Bottleneck configuration in C2f. The dysample module is also applied to refine the initial upsample structure to compensate for the problem of upsample easily losing low-level details and semantic information. For more efficient and accurate detection, SCDH detection head was adopted and CIoU was replaced with powerful-IoU, this lightweight structure accelerates convergence and improves anchor box accuracy. Finally, experimental results demonstrate that OPT-YOLO achieves mAP, Precision, and Recall scores of 0.970, 0.979, and 0.920, respectively, marking improvements of 3.2 %, 7.4 %, and 8.7 % over YOLOv8, while reducing FLOPs by 3.6 G, FPS has increased by 94. The heat map and feature map extractions confirm enhanced feature extraction capabilities, and ablation studies validate the performance gains from our enhanced modules. Overall, OPT-YOLO offers a more efficient and precise solution for planar optical component surface defect detection, highlighting its potential for industrial applications.