Saliency detection for strip steel surface defects using multiple constraints and improved texture features

Abstract

Surface defect detection of strip steel is still a challenging task for its complex variances, e.g., intra-class defects exist large differences in appearance while inter-class defects contain similar parts. To address these issues, we regard the defect object as the salient part of the image and propose a novel, effective saliency propagation algorithm based on multiple constraints and improved texture features (MCITF). Firstly, we deliberately design 83-dim texture features that are used to generate label matrix (among which the label information viewed as the important basis of diffusion process) in the framework of multiple-instance learning. Then we resort to Laplacian regularization viewed as smoothness constraint for enlarging the gap between defect objects and background, and high-level prior (background, object, and mid-level feature) constraints for constraining the label information propagation process locally in order to uniformly highlight the complete defect objects while effectively suppress the non-salient background. Finally, we observe that the superpixel segmentation algorithm based on spectral clustering can adequately capture the edge information of defects, thus promoting to yield high-quality pixel-level saliency maps. Experimental results implemented on the real challenging strip steel benchmark database demonstrate that our MCITF model outperforms state-of-the-art methods with large margins and strong robustness in terms of eight evaluation metrics.