Automated fabric inspection systems are essential for maintaining consistent textile quality in modern manufacturing, where high-speed production and complex surface textures 1c`challenge traditional visual inspection. Intelligent, data-efficient anomaly detection methods are therefore increasingly important. Most existing fabric defect detection approaches depend on supervised or weakly supervised learning, requiring extensive labeled defect samples that are costly to acquire and insufficiently representative of real industrial variability. Additionally, many methods fail to explicitly preserve fine-grained texture semantics, leading to false alarms or missed defects on complex woven and patterned fabrics. This paper introduces a Self-Supervised Texture-Decomposition and Defect Amplification Network (STDAN). This novel framework uniquely decomposes fabric images into structure-consistent texture bases and residual anomaly cues without requiring defect annotations. STDAN integrates a texture-consistency encoder with a generative defect amplification module that exaggerates latent irregularities through controlled perturbation in the feature space rather than pixel-level synthesis. A contrastive self-supervised objective aligns normal texture representations while separating amplified defect responses, enabling the detector to learn discriminative anomaly boundaries from normal data alone. Experimental evaluations on multiple public fabric inspection benchmarks show that STDAN achieves superior detection accuracy and localization precision compared with state-of-the-art unsupervised and GAN-based methods, particularly for small, low-contrast, and previously unseen defects. The technique demonstrates strong robustness to texture variation and illumination changes, confirming its practical applicability. By uniquely combining texture decomposition and feature-level defect amplification, STDAN offers a scalable, annotation-free solution for reliable fabric quality inspection in industrial environments.
Sujitha et al. (Tue,) studied this question.