Weakly Supervised Anomaly Detection by Utilizing Incomplete Anomaly Information

Anomaly detection aims to distinguish abnormal samples from normal ones. In this paper, only a limited number and types of anomaly samples are assumed to be available in the training set while most types of anomaly samples cannot be obtained. To effectively utilize this incomplete anomaly information, we introduce GMMAD. In order to get a good representation of samples, we first extract two different scale feature maps from the pre-trained ResNet, then use the fusion module to fuse the maps. Among the fusion module, the projector aims to project the larger maps into the same size as the smaller ones, then the maps are concatenated and fed into an integrator to produce final representations. The Gaussian Mixture Model is employed to characterize the distribution of normal features, whose parameters are updated according to the negative log-likelihood, and then it is used to guide the fusion module to output distinct features to detect anomalies. Through iterative training of the GMM and the fusion module, our approach effectively utilizes incomplete anomaly information to detect anomalies. Experiments on three medical datasets have shown that with only a limited number and types of anomalies, GMMAD outperforms existing methods in detecting anomalies.