Unsupervised Image Anomaly Detection and Localization in Industrial Textile Inspection

Abstract

Image anomaly localisation in texture patterns is a key challenge in industrial inspection, particularly in textile manufacturing scenarios. Recent unsupervised learning methods report strong performance on widely adopted benchmarks such as MVTec AD. However, it remains unclear to what extent these methods generalise across different texture distributions, anomaly types, and pattern structures commonly found in real-world industrial environments. This work investigates the current state of the art in unsupervised image anomaly detection and localisation, with a particular focus on texture-based inspection. It reviews existing model families, datasets, evaluation metrics, and development frameworks while providing a structured overview of the evolution of the field. The experimental component presents a comparative study of several unsupervised anomaly localisation methods applied to textile-oriented industrial scenarios. The evaluation compares these methods across multiple textile anomaly datasets, analysing their behaviour under different texture patterns and data distributions. The study also examines the influence of texture-related inductive biases and evaluates the suitability of the methods for deployment in real industrial production-line environments.