Unsupervised Fabric Defect Detection via Autoencoder Reconstruction and One-Class SVM Analysis

Fabric defect detection is a critical quality control process in textile manufacturing, but remains challenging due to the complex texture and variability of fabric patterns. Traditional supervised deep learning methods require extensive annotated data, which is often impractical in industrial settings. To address this, we propose a novel unsupervised approach leveraging autoencoder-based architecture to detect fabric defects. Our method reconstructs input images and employs a one-class SVM model to analyze reconstruction loss, distinguishing defective from nondefective regions. The experiments were carried out on a binary text defect classification dataset, using a holdout validation strategy of 80–20. Our approach outperformed state-of-the-art models, including AlexNet, VGG-16, VGG-19, MobileNetV1, and ResNet-50, with an average improvement of 3.57% in accuracy and 2.98% in F1 score. Compared to ResNet-50, our method achieved 7.2% higher recall and 4.03% lower precision, reflecting its robustness in identifying rare defect patterns. These results demonstrate the effectiveness of unsupervised learning for fabric defect detection, offering a scalable and efficient alternative to traditional supervised methods. This work highlights the potential of reconstruction-based anomaly detection in industrial applications, providing a foundation for further exploration in defect detection for complex and heterogeneous surfaces.