Anomaly Detection Combining Discriminative and Generative Models

Abstract

This paper proposes a method to accurately detect anomaly from an image by combining features extracted by discriminative and generative models. Automatic anomaly detection is a key factor for reducing operation costs of visual inspection in a wide range of domains. The proposed method consists of three sub-networks. The first sub-network is convolutional neural networks as a discriminative model for extracting features to distinguish between anomaly and normal. The second subnetwork is a variational autoencoder as a generative model to extract features representing normal. The third sub-network is neural networks to discriminate between anomaly and normal on the basis of features from the discriminative and generative models. Experiments were conducted using pseudo anomalous images generated by superimposing anomaly which was manually extracted from real images. Results of the experiments show that the proposed method improves the area under the curve by 0.08-0.33 points compared with that of a conventional method. With high accuracy, automatic visual inspection systems can be implemented for reducing operation costs.