Unsupervised anomaly detection in the textile texture database

Anomaly detection in textile images poses significant challenges due to the scarcity of defective samples and the complex nature of textile textures. This study presents a novel image processing workflow that enhances the unsupervised Variational Autoencoder’s (VAE) ability to identify anomalies in textile images, addressing the limitation of insufficient defective samples in real-world manufacturing scenarios. The primary motivation behind this research is to develop a robust anomaly detection method that can be trained using only normal samples, overcoming the common imbalance between normal and defective samples in the textile industry. Our proposed method introduces domain-specific techniques to preprocess images, assess the adequacy of training samples, and employ intuitive visual methods to differentiate between normal and abnormal samples. A key strength of our approach lies in strategically cropping original images into smaller blocks, increasing training samples and computational efficiency. However, this cropping step introduces abrupt boundary issues that can hinder accurate anomaly detection. To mitigate this problem, we developed a refined image processing approach that effectively resolves boundary artifacts, enabling precise localization of abnormal regions. We trained, tested, and validated our VAE model using the TILDA textile texture database. The experimental results highlight the robustness of our method, achieving high identification rates of 74% for normal samples and 76.9% for abnormal samples, even when trained solely on normal samples. The insights gained from this study have significant implications for the textile industry, paving the way for more efficient and reliable quality control processes.