Self-supervised distributional and contrastive learning model for image anomaly detection

Anomaly detection is a critical task in computer vision and machine learning, which aims to detect anomalies that are away from the distribution of normal data. Prevalent methods perform anomaly detection via pretraining the model or constructing pretext tasks based on instance-level contrastive learning. However, they ignore the underlying distribution of features for normal data, which is crucial for detecting anomalies. In this paper, we propose a self-supervised distributional and contrastive learning model for anomaly detection. For distributional modeling, based on the augmented dataset with given rotated images, we propose to model latent features of normal sample set under each rotation transformation via a Gaussian Mixture Model (GMM). By the proposed GMM clustering loss, we maximize the probability of data to the GMM model with rotation that is applied to the data, and minimize its probability to the GMMs of other rotations. For contrastive learning, we discriminate augmented features via rotation-based contrastive learning that constructs positive pairs augmented from the same instance under the same rotation and negative pairs augmented from different instances under rotations. Our approach jointly conducts distributional modeling of image features considering feature clustering properties, and discriminative learning of features in a self-supervised way. We also design detection scores from different modules to detect anomalies. Extensive experiments on image datasets (e.g., CIFAR-10, CIFAR-100 and ImageNet-30) demonstrate that our method achieves favorable performance compared with several state-of-the-art methods on unsupervised image anomaly detection tasks. Ablation studies demonstrate the effectiveness of the key modules in our model.