Adaptive Continual Learning for Defect Detection in Dynamic Textile Manufacturing

In the textile industry, ensuring high product quality is challenging due to complex designs and the demand for customization. Traditional manual inspection is labor-intensive and inaccurate, making it unsuitable for modern production. This research proposes an advanced automated defect detection system that integrates a design specific auto-encoder and a common defect detection module. The auto-encoder learns and reconstructs textile patterns to detect subtle anomalies in new designs without requiring labeled or defective data. Concurrently, the common defect module recognizes common defects across various fabrics. The system uses active learning to optimize retraining and continual learning to adapt to new data without forgetting past knowledge. Experimental results show significant improvements in accuracy and robustness, reducing the need for frequent retraining and enhancing defect detection efficiency, thus supporting high quality textile production.