Fuzzy-UNet: semantic-fuzzy integration for fabric defect identification

Abstract The detection of fabric surface defects is a critical component for quality assurance and operational efficiency within the textile industry. A significant challenge in this field is the effective incorporation of domain expertise to enhance the precision of defect recognition. This paper introduces Fuzzy-UNet, an innovative fabric defect detection system that integrates a semantic segmentation network with a fuzzy decision model, addressing the challenge of domain knowledge integration. By combining a cascaded data model with a knowledge-based model, Fuzzy-UNet harnesses the power of deep learning and artificial experience to refine the identification process. The data-driven model of our system is advanced UNet architecture, specifically tailored for detecting subtle defects in fabrics with non-standard aspect ratios. The system’s novel fuzzy decision model utilizes spatiotemporal data from multiple cameras, which is essential for enhancing the accuracy of the detection process. Our comprehensive experiments demonstrate the robustness of Fuzzy-UNet, with a significant increase in accuracy to 96.78% and a marked reduction in the False Positive Rate (FPR). The superior performance of Fuzzy-UNet over existing methods makes it a leading solution for industrial fabric defect detection.