Texture-Adaptive Fabric Defect Detection via Dynamic Subspace Feature Extraction and Luminance Reconstruction

Defect detection in textile manufacturing is critically hampered by the inefficiency of manual inspection and the dual constraints of deep learning (DL) approaches. Specifically, DL models suffer from poor generalization, as the rapid iteration of fabric types makes acquiring sufficient training data impractical. Furthermore, their high computational costs impede real-time industrial deployment. To address these challenges, this paper proposes a texture-adaptive fabric defect detection method. Our approach begins with a Dynamic Subspace Feature Extraction (DSFE) technique to extract spatial luminance features of the fabric. Subsequently, a Light Field Offset-Aware Reconstruction Model (LFOA) is introduced to reconstruct the luminance distribution, effectively compensating for environmental lighting variations. Finally, we develop a texture-adaptive defect detection system to identify potential defective regions, alongside a probabilistic ‘OutlierIndex’ to quantify their likelihood of being true defects. This system is engineered to rapidly adapt to new fabric types with a small number of labeled samples, demonstrating strong generalization and suitability for dynamic industrial conditions. Experimental validation confirms that our method achieves 70.74% accuracy, decisively outperforming existing models by over 30%.