Automated Fabric Defect Detection Using Unsupervised Feature Learning with Deep Autoencoders

Abstract

Ensuring consistent quality in textile manufacturing requires reliable defect detection systems. This paper presents an unsupervised method using convolutional autoencoders to identify surface anomalies in woven fabrics. The model is trained purely on normal fabric samples and learns to reconstruct typical textures. Any significant reconstruction error indicates the possible presence of a defect. Three common defect types; holes, spots, and bars are used to evaluate the model's efficiency. The images were converted to grayscale to simplify processing and accelerate training. Experimental tests show that the model achieves high detection accuracy, especially for structurally pronounced defects such as holes. While the results are promising, some limitations remain, particularly for the detection of low-contrast or subtle defects such as spots. The proposed system can be useful in industrial inspection settings where annotated data is scarce. In future work, we plan to investigate more advanced structures and evaluate the impact of using color channels. The goal is to develop a lightweight and flexible solution for real-time deployment in small and medium-sized manufacturing environments.