Abstract: The detection and quantification of surface defects in seamless steel pipes are crucial for quality judgment. However, existing detection methods mainly rely on grayscale image analysis and lack comprehensive judgment of defect depth, resulting in one-sided and inaccurate detection results. To solve this problem, 3D cameras are used to collect data on the surface of steel pipes, which can synchronously obtain grayscale images and point cloud data with the same size. By processing the point cloud data, the depth of defects relative to the reference surface can be calculated, and further quantified to obtain pseudo color depth images, which can intuitively display the depth information of defects. In order to improve the defect detection capability, a bilateral network structure is added to the Yolov5 model, where grayscale images and pseudo color depth images are input into the detail branch and semantic branch respectively to extract features. The data from the two branches are fused to obtain new intermediate features for object detection. The experimental results show that the pseudo color depth image generated by the relative depth measurement method can effectively eliminate interference such as jitter and torsion, and the depth measurement error is less than 0.1 mm. In addition, compared with the traditional grayscale image detection mode, the Yolov5 model with the addition of a bilateral network structure improved the mAP index by 4.7% and met the real-time detection requirements at a speed of 108 fps. On the basis of qualitative analysis of defects, quantitative analysis of defects was achieved by adding depth dimension, which not only improved the comprehensiveness of detection but also significantly improved the accuracy of detection.

Key words: defect detection, contour fitting, pseudo-color depth images, Yolov5, bilateral network