Proposal of Rotation-Robust Anomaly Detection Based on Inverse Projection Error Metric from Normal Feature Representation Space

This study proposes a rotation-robust anomaly detection method based on inverse projection error from a normal feature representation space. Conventional methods often have trouble detecting anomalies in objects that rotate, as their anomaly metrics, like Mahalanobis and Euclidean distances. Our method addresses this by creating normal representation space using only normal samples, capturing normal feature variations through principal component analysis (PCA) in a low-dimensional subspace. We then project test features into this space and inverse-project them back, using the difference between the original and reconstructed features as an anomaly score. This score helps distinguish between normal and abnormal features even when objects are rotated. Unlike deep learning methods, which often require more complex processing, this approach is simpler. Tests on the MVTec AD dataset show that our method performs well, achieving an AUROC of 96.8% for the Screw class, which is higher than PaDiM's 94.9% and PatchCore's 95.6%. This shows our method's strong ability to detect anomalies in rotated objects, making it effective for manufacturing needs and adaptable to various inspection scenarios.