Crack detection by unsupervised anomaly detection method using SQ-VAE

Abstract

Though the research which detects the crack from the image using deep learning is actively carried out, many of them are techniques based on the supervised learning, and it is a problem that the the annotation work is large. Therefore, this study attempts to perform crack detection on concrete surfaces by unsupervised learning. Unsupervised learning is widely used in anomaly detection, and it is possible for a model that learns only normal data to identify anomalies. This study aims to verify the crack detection accuracy in an unsupervised anomaly detection method. SQ-VAE represents image features by a finite number of embedding vectors called codebooks. It is assumed that the codebook which learns only the normal data is far from the features of the anomalies, and the crack is detected based on the similarity degree of the features. Experimental results show that the proposed method can detect cracks with high accuracy.