In manufacturing, inconspicuous anomalies—small, low-contrast defects with subtle visual signatures—pose a major challenge for automated quality assurance. These defects often propagate through downstream processes, increasing material waste, energy consumption, and rework, while also compromising long-term product reliability. To address this problem, the ICAN (Inconspicuous Anomaly as Noise) training strategy is introduced as an unsupervised approach tailored to manufacturing inspection. The method models weak defects as structured perturbations added to nominal samples and trains a network to recover the original image. This encourages the extraction of representations that more clearly distinguish subtle anomalies from normal variation. ICAN is evaluated on three publicly available manufacturing datasets and compared with a reconstruction-based autoencoder and a standard CNN classifier. The results show that ICAN reduces false positives while preserving high true-positive rates, providing a more reliable inspection method.
Ghansiyal et al. (Thu,) studied this question.