Discriminative Learning for Supervised Anomaly Detection

초록

Most unsupervised anomaly detection methods learn distributions from normal samples, which solve the class-imbalance problem with different amounts of samples. However, learning only normal samples can affect the model to have loose decision boundary and low discriminality. In this paper, we propose the novel supervised approach for detecting anomalies by exploiting known anomaly samples. Our method uses embeddings from ImageNet backbone model and transfer extracted features towards target domain using feature adaptor. Then, anomaly score is computed using GAN discriminator, which differentiates positive and negative samples to estimate normality of an image. Our approach achieves an anomaly detection AUROC of 99.2% on MVTec AD benchmark. © 2023 IEEE.