High precision anomaly detection based on pre-trained features enhanced by only large amount of normal samples

As improvement of superior method called PaDiM in anomaly detection, we propose a method based on pre-trained features enhanced by training with consolidating normal samples to its centroid in feature space. PaDiM pre-trains the model with only ImageNet and parameterizes the features of target normal images by normal distribution. However, this method pre-trains the model while ignoring normal images that follow a normal distribution, which leads to performance degradation. In contrast, our method centralizes the features of normal images during pre-training, and as a result, the mean Image/Pixel AUROC of the proposed method was higher than that of PaDiM (94.2/96.2 and 93.6/95.7, respectively) in experiments with the MVTec AD dataset.