Research on a large-scale non-planar metal surface defect detection method based on chocolate block array stitching approach and improved YOLOv8 algorithm

Abstract

Detecting defects on large-scale non-planar metal surfaces remains a major challenge that limits further improvements in detection performance. To reduce the time-consuming process and improve the stitched image, this paper proposes a new image stitching approach based on the Chocolate Block Array Stitching technique and the improved DEW-YOLOv8 (DWR-ECA-WIoU You Only Look Once v8) algorithm. The Chocolate Block Array Stitching approach involves locating the optimal camera pose for sub-sections, automatically cropping images according to predefined coordinates and applying perspective transformation, and finally performing grid-based arrangement and stitching according to spatial pose information, thereby achieving an efficient synthesized image for the large-scale non-planar metal surface. The DEW-YOLOv8 presents three improvements compared to the original YOLOv8: first, the original Bottleneck structure in the Faster Implementation of CSP Bottleneck with two convolutions (C2f) module of YOLOv8 is replaced with a Dilation-Wise Residual Module (DWR) to enhance the model’s feature extraction capability for multi-scale defects. Second, an Efficient Channel Attention (ECA) mechanism is introduced at the end of the backbone network to improve the model’s adaptability. Third, the original Complete Intersection over Union (CIoU) loss function is replaced with a Wise Intersection over Union (Wise-IoU) loss function based on a dynamic non-monotonic focusing mechanism. Experimental results demonstrate that the proposed DEW-YOLOv8 detection algorithm outperforms the original, with a 1.8% increase in precision, a 1.3% improvement in recall, and a 1.3% enhancement in mean average precision (mAP). Moreover, when combined with the Chocolate Block Array Stitching approach, it reduces the time cost by 65.5% compared to direct detection on the original full image.