关键词: 缺陷检测, YOLOv8, O型密封圈, DySample

Abstract: Aiming at the problems such as small surface defect scale, high similarity between defect features and background, difficulty in feature extraction and the needs of industrial real-time detection, an improved YOLOv8n YOLO-Oring algorithm is proposed. On the basis of convolution calculation, a new CBLGhost module is designed by incorporating linear changes to reduce the computational resources required for calculating feature graphs. In view of the high similarity between defects and background, the DySample dynamic up-sampling module is introduced to make the sampling points concentrate in the target area and ignore the background part, so as to realize the effective identification of defects. In order to improve the detection efficiency, the C2f-OREPA module is designed to convert the complex structural reparameters into a single convolution layer, which can reduce a lot of training time while maintaining the feature expression ability. In order to improve the  ability of algorithm to identify small-scale defects, DyHead-DCNv3 detection head is designed to identify multi-scale targets, and make up for the shortcomings of traditional standard convolution in long-distance modeling ability and adaptive spatial aggregation ability, so as to better complete the detection task. Due to the lack of O-ring detection dataset, 1?734 datasets including 3 types of defects, such as scratches, dents and burrs, are established. Experimental results show that the mAP of YOLO-Oring algorithm reaches 94.1%, increasing by 1.3 percentage points, and FLOPs has decreased by 16%. Compared with mainstream target detection algorithms, the results show that YOLO-Oring algorithm has better detection performance for O-ring surface defects and is more conducive to industrial real-time detection.

Key words: defect detection, YOLOv8, O-ring, DySample