Temperature-Optimized Self-Supervised Contrastive Learning Enables Data-Efficient and Explainable Fabric Defect Detection

Fabric defect detection plays a pivotal role in the textile industry, as it ensures product quality and minimizes waste. Traditional machine learning models often struggle with the scarcity of labeled data, which is a major challenge in automated defect detection systems. This paper explores the potential of Self-Supervised Learning (SSL) techniques, such as SimCLR and BYOL, to address this issue by leveraging large volumes of unlabeled fabric data. By pretraining models using SSL, we propose a framework for fabric defect detection that minimizes the need for extensive labeled datasets while maintaining high classification accuracy. Furthermore, our approach integrates anomaly detection for localizing fabric defects, providing a robust solution for real-time monitoring in production environments. Experimental results demonstrate that fine-tuning SSL-based models on smaller labeled datasets significantly enhances defect detection capabilities across various fabric types, including plain, regularly printed, and irregularly printed fabrics. The SimCLR model with the ResNet50 backbone at a temperature of T = 0.1 achieved the best performance, yielding a fine-tuned accuracy of 91.4% and top-1 accuracy of 91.97%, and top-5 accuracy of 98.5%. This model also showed high precision and recall, particularly for the detection of contamination, stains, and cuts. Statistical validation with paired t-tests confirmed highly significant improvements (p < 0.05) in comparison to several baseline models, confirming the robustness of the proposed approach. Additionally, a quantitative explainability assessment revealed that LayerCAM achieved the highest intersection over union (IoU) overlap of 0.92 between attention maps and annotated regions, indicating superior interpretability and model transparency. This research showcases the promise of SSL in automating fabric defect detection while reducing reliance on labor-intensive manual inspections and enabling scalable deployment in industrial settings.