[2N6-GS-10] AI application

[2N6-GS-10-03] Comparison study of continuous and discrete latent variable models for anomaly detection of river revetment using VAE

〇Yukino Tsuzuki¹, Ryuto Yoshida¹, Junichi Okubo¹, Junichiro Fujii¹ (1. Yachiyo Engineering Co., Ltd.)

Keywords:VAE, DIP-VAE, SQ-VAE, image generation, anomaly detection

By considering a single block as a uniform component, an anomaly detection method for river revetment using VAE is being studied. This previous study was conducted for a single block shape, and it is necessary to investigate a method that can be applied to blocks of various shapes for practical use in the future. However, it has been pointed out that the reconstruction performance of conventional VAE is low. Hence, there is a possibility that some blocks may be difficult to reconstruct when block shapes are varied. On the other hand, VQ-VAE has been proposed to generate high-resolution images by using discrete latent variables. Furthermore, SQ-VAE, which improves on VQ-VAE to enhance reconstruction performance, has been proposed. In this study, we investigate the characteristics of both the continuous latent variable model (DIP-VAE) used in the previous study and the discrete latent variable model (SQ-VAE) with high reconstruction performance, with the aim of performing anomaly detection on various types of blocks. We find that there exist block shapes that are difficult to reconstruct in the DIP-VAE, in which case the anomaly detection performance is significantly degraded. SQ-VAE can reconstruct blocks of any shape, but it can reconstruct not only normal regions but also abnormal regions. Finally, we suggested the possibility of an anomaly detection method using latent variables of SQ-VAE.