Enhanced Fabric Defect Detection With Feature Contrast Interference Suppression

Detecting fabric defects is a critical task in the textile industry, as it involves identifying and localizing imperfections in fabrics. However, many existing unsupervised defect detection methods are sensitive to variations and disturbances in the normal background of fabric images, leading to a high false alarm rate and unsatisfactory performance. To address this issue, we propose a novel defect detection method called feature contrast interference suppression (FCIS) that aims to enlarge the feature distance between foreground and background textures while suppressing the distribution of background features to obtain a more robust background representation. Our approach consists of several key components: First, we introduce background-foreground contrastive learning (BFCL) to enhance feature representation between background and foreground elements at the feature level. Second, we use multiscale denoising knowledge distillation to aid the model in learning more accurate representations of normal background features. Finally, we design an attention decoder that refines and segments the defect mask generated through feature comparison, resulting in improved detection accuracy. Our proposed FCIS has been validated using five mainstream datasets. Experimental results demonstrate that our proposed method outperforms existing approaches, achieving improvements of 2.38% in image-level AUROC and 3.73% in pixel-level AUPRO, along with a 9.12% enhancement in average defect localization precision.