[2K1-ES-2] Machine learning: Image classification

[2K1-ES-2-05] Comparison Study of Deep Anomaly Detection Techniques in Images

〇Takuya Ono¹, Junichiro Mori¹ (1. Graduate School of Information Science and Technology, The University of Tokyo)

Keywords:Anomaly Deterction, GAN, Metric Learning

Anomaly detection is an important technique in the field of computer vision and other numerous areas. In recent years, deep learning has become mainstream and is known to have better performance than traditional methods. Due to the nature of the anomaly detection problem, the number of anomalous data is extremely small, and the data set is often extremely imbalanced. Therefore, it is difficult to apply supervised learning, hence development of unsupervised/semi-supervised learning is expected. In this paper, we outline the implementation of the state-of-the-art method using the semi-supervised learning (e.g. AnoGAN, convolutional auto-encoder, etc.) and metric learning using class labels (e.g. AdaCos, L2 softmax, etc.) for anomaly detection in images, and then evaluated its performance. As a result, we concluded that the performance with the current deep anomaly detection method is poor for practical applications, and there is considerable room for improvement. Finally, we describe the problems of these methods and the current research issues.