Frequency perturbation analysis for anomaly detection using Fourier heat map

Anomaly detection is an essential task within an industry domain, and sophisticated approaches have been proposed. PaDiM has a promising direction, utilizing ImageNet-pretrained convolutional neural networks without expensive training costs. However, the cues and biases utilized by PaDiM, i.e., shape-vs-texture bias in an anomaly detection process, are unclear. To reveal the bias, we proposed to apply frequency analysis to PaDiM. For frequency analysis, we use a Fourier Heat Map that investigates the sensitivity of the anomaly detection model to input noise in the frequency domain. As a result, we found that PaDiM utilizes texture information as a cue for anomaly detection, similar to the classification models. Based on this preliminary experiment, we propose a shape-aware Stylized PaDiM. Our model is a PaDiM that uses pre-trained weights learned on Stylized ImageNet instead of ImageNet. In the experiments, we confirmed that Stylized PaDiM improves the robustness of high-frequency perturbations. Stylized PaDiM also achieved higher performance than PaDiM for anomaly detection in clean images of MVTecAD.