Self-Supervised Image Anomaly Detection Through Diverse Pseudo Anomaly Insertion

Industrial anomaly detection through deep learning-based vision systems is becoming a critical inspection tool for deploying efficient, defect-free manufacturing lines in smart factories. In fully automated manufacturing lines, deep self-supervised learning (SSL) based anomaly detectors are ideal inspection tools, as they require no data annotations during a periodic detector retraining. In this paper, we propose a novel anomaly insertion framework designed to seamlessly integrate with various state-of-the-art self-supervised anomaly detectors (ADs). This framework enhances anomaly detection and segmentation performance by incorporating diverse synthetic anomalies during training and produces labels and segmentation masks automatically. We introduce an artificial anomaly insertion tool that enables practitioners to manipulate normal and anomalous source images, creating training samples with varied characteristics and distributions. Our tool consolidates various pseudo anomaly insertion (PAI) techniques into a single platform and offers adaptive or selective anomaly insertion modes based on user preferences. We integrated our anomaly insertion tool into baseline self-supervised AD and evaluated its performance on well-known public datasets such as MVTec, VisA and KolektorSDD2. Our framework not only reduces dependency on labeled data but also improves robustness of AD across various real anomaly patterns, paving the way towards robust self-supervised anomaly segmentation paradigm. We proposed a novel Affine transformation-based pseudo-Anomaly insertion scheme (AA) that demonstrates an improvement in $AP_{pixel}$ of 5.2% for MVTec and 10.1% for VisA dataset with respect to the second best PAI technique. To use the PAI library, code is provided at https://github.com/RizwanAliQau/pail.git