An edge-available defect detection And Localization Flow Model

Unsupervised defect detection aims to identify and localize unpredictable defects in industrial manufacturing processes caused by uncontrollable factors. Flow-based unsupervised models have recently attracted considerable attention from the research community. However, existing methods generally suffer from limited sensitivity to fine edge structures in images, making it difficult to effectively capture boundary information of defective regions, as well as excessive redundancy in feature representations, which degrades both discriminative power and computational efficiency. To address these limitations, we propose an Edge-Aware Defect Detection and Localization Flow model (EADFlow). EADFlow integrates a Frequency Domain Edge-Aware Module to enhance the modeling of high-frequency edge information and introduces a Focused Local and Global Attention Module to reduce feature redundancy and strengthen feature representation capability. Experimental results show that EADFlow achieves state-of-the-art performance across multiple industrial defect detection benchmarks significantly outperforming existing advanced methods.