RCA: YOLOv8-Based Surface Defects Detection on the Inner Wall of Cylindrical High-Precision Parts

In this study, we propose a random cropping augmentation (RCA) based on an optimized YOLOv8 algorithm by introducing a tiny object detection layer to improve the accuracy of surface defect detection on the inner wall of micro-holes of cylindrical high-precision parts. The inner surfaces of cylindrical high-precision parts possess attributes like high smoothness, elongated channels, limited visibility, minute flaws, and challenges capturing image data. These characters impose significant difficulties in automating defect detection for this part type. Taking the automobile fuel injector valve seat as an example, industrial CCD cameras and lenses are combined with a circular light source to acquire magnified image data. The images are preprocessed using tile and adaptive histogram equalization algorithms. Then, the RCA algorithm expands and randomly combines the image data. We compare the proposed algorithm with the original faster RCNN, YOLOv5 and YOLOv8 algorithm for surface defect detection on the inner wall of the automotive fuel injector valve seat. The experimental results demonstrate that after extending the dataset with the RCA algorithm, there have been significant improvements in the mean average precision (mAP) for the faster RCNN, YOLOv5, and YOLOv8 algorithms. The mAP@0.5 for these algorithms has increased from 37.1, 48.7, and 62.2 to 60.9, 72.6, and 81.1, respectively. Furthermore, the improved YOLOv8 algorithm, which incorporates tiny object detection module, achieved an mAP@0.5 of 87 on the dataset extended by the RCA algorithm, with precision and recall reaching 89.7 and 80.6, respectively. The improved YOLOv8 + RCA algorithm performs better than mainstream algorithms in surface defect detection on the inner wall of the fuel injector valve seat while meeting the requirements of industrial production quality inspection for detection speed. © King Fahd University of Petroleum & Minerals 2024.