Improvement of Anomaly Detection Performance of PaDiM by Fast Fourier Convolution with Total Variation Regularization

PaDiM, an anomaly detection model using a pre-trained CNN on ImageNet, shows high performance. However, the pre-trained CNN model has a texture bias, resulting in poor performance for global anomalies. Therefore, we used Fast Fourier Convolution (FFC) to extract global features by extracting features from Fourier space in addition to feature extraction using 3 × 3 convolutional filter in the pre-trained model. Total Variation regularization was applied to the feature map of the FFC block during pre-training. This improved anomaly detection performance for global anomalies and robustness to perturbations in the frequency band.