A note on detection of distress regions in subway tunnels by using U-net based network

摘要

This paper presents an automated distress region detection method using subway tunnel images. We previously proposed a method for realizing distress detection in subway tunnels by using several kinds of fully convolutional networks, namely, FCN, U-net, Seg-net, Residual U-net, and Deeplab v3+. In our previous investigation, we found that the U-net got the highest performance in the subway tunnel distress detection task. However, this original U-net approach had several limitations in its network architecture for the task of distress region detection. In this paper, we attempt to improve the detection performance of U-net by using the ASPP (Atrous Spatial Pyramid Pooling) module from Deeplab v3+ network and remain the VGG-16 backbone rather than using ResNet backbone. By introducing this new architecture, we achieve higher performance than conventional methods. We verify the effectiveness of our method through experiments.