Surface defect segmentation of steel pipes based on superpixel prior knowledge

Abstract Seamless stainless steel pipes (SSSP) are widely used in aerospace, petrochemical, and other fields, making the quality assessment of their surfaces particularly important. However, defect detection on the inner surface of SSSP is challenging, such as the scarcity of samples and the blurring of defect boundaries. Therefore, this paper proposes a method for surface defect segmentation of steel pipes based on superpixel prior knowledge and builds a classification model to achieve the segmentation task. Firstly, the images are superpixel clustered by Simple Linear Iterative Clustering (SLIC) with Edge Constraints and each superpixel is created as a description vector incorporating edge information. Then, the local description vectors are input into a support vector machine (SVM) for classification, and the foreground-background graph is constructed using confidence as priori knowledge. Finally, the graph model is constructed by co-constructing the terminal weights from the Gaussian Mixture Model (GMM) estimates and the SVM regression confidence, while to improve the edge segmentation accuracy, the neighborhood difference matrix is co-constructed with the superpixel mean and Canny edge image as the neighborhood weights, and the defects are segmented by using the minimum graph cut algorithm. Experimental results show that this method still demonstrates excellent accuracy and robustness even with a limited number of samples, achieving an intersection over union (IoU) of 87.94% on the self-made dataset, which outperforms other methods.