Unsupervised Anomaly Detection Based on Data Augmentation and Mixing
Naoya Ishida, Yuki Nagatsu, Hideki Hashimoto · IECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society · 2020
When performing tasks using deep learning, the quantity and quality of data are important. However, in unsupervised anomaly detection, it is difficult to obtain a large amount of high-quality training data. Hence, data are typically extended via rotation, translation, and cutting. However, the typical data augmentation method may generate data far from the original image. To obtain various high-quality training data, we partially use AugMix, a data augmentation method that mixes the original image with images obtained via image processing, such as by rotating the original image. Using this data augmentation, new data can be generated while retaining the features of the original image as much as possible, and the diversity of training data can be enhanced, compared with performing the same image processing. Consequently, the effectiveness of the proposal method for improving the accuracy of unsupervised anomaly detection is confirmed.