A NOVEL APPROACH TO INDUSTRIAL FABRIC DEFECT ANALYSIS USING ADAPTIVE ANOMALY REASONING AND HIERARCHICAL FEATURE LEARNING

Feyzanur Ozden¹ , Canan Tastimur²

One of the most significant problems businesses face in their production processes is the accurate and rapid detection of defective areas on the product surface. In this study, defect detection and classification were performed on our fabric defect dataset containing one normal and four defective fabric samples. Transformer-based unsupervised fault detection and CNN-transformer-based supervised fault classification are designed as an integrated architecture. During fault detection, images are separated into patch-based features, and then contextual relationships are learned via the transformer. In fault detection, our innovative module creates a memory bank with normal fabric images, and faulty regions in the fabric are detected thanks to the Mahalanobis distance and statistical distribution deviations. In the fault classification phase, Channel Attention, with the CNN-transformer hybrid structure, achieved a test accuracy of 97.78%. In addition, the model's attention mechanism, which correctly directs attention to faulty areas, is presented with visual results.

Keywords
Adaptive Feature Fusion, Anomaly Localization, Contextual Representation Learning, Explainable Artificial İntelligence, Industrial Defect Inspection,