A Contour-Based Instance Segmentation Method for Copper Tube

Image segmentation is an essential prerequisite task for the defect detection approach. To improve the performance of defect detection in copper tube welding areas, a contour-based instance segmentation method is presented in this study. The proposed method is a coarse-to-fine process and it can be divided in two stages. Firstly, the coarse contour of welding areas is generated by learnable initial contour proposal architecture. Different from existed approaches, the proposed architecture applied the inner center replace the bounding box center for the center point prediction to ensure the center point falls inside the contour. Moreover, a label sampling strategy based on Douglas Peucker (DP) algorithm is proposed to enhance the performance of corner points sampling and a normalized corner attention loss function is proposed to reduce the learning difficulty. Secondly, an coarse contour deformation framework based on the multi-scale vertex feature fusion and low-level feature enhancement Deep Snake modules is organized to optimize the coarse contour obtained by first stage. The experimental results show that our method has achieved certain improvements in both segmentation accuracy and segmentation speed compared to the baseline method on both SBD benchmark and the copper tube dataset.