**1. Introduction**

With the growth of economies worldwide, various infrastructures such as tunnels, bridges, and viaducts have been constructed [1], which are indispensable for our daily life and are used by a large number of people in their daily lives [2]. However, infrastructures built more than five decades ago are experiencing aging problems, and the number of dilapidated infrastructures will significantly increase in the near future [1], and their maintenance cost will also increase exponentially [3,4]. Under these circumstances, a more efficient maintenance and management of infrastructures have become an urgent issue. Recently, much interest has been shown in smart maintenance and management technologies, including artificial intelligence (AI), Internet of Things (IoT), and big data analysis. These techniques have already been applied to real-world problems in various fields [5–9]. These techniques are required in the field of infrastructure to improve the efficiency and accuracy of infrastructure maintenance [10,11].

As an important infrastructure, urban railway systems have been mainly constructed during the high-speed economic growth period. In urban areas, the overground transportation network is already dense and its expansion potential is limited. On the other hand, the subway transportation environment, such as subway tunnels, is expected to expand further in the future. However, through the high frequency of use, tunnels that were built decades ago inevitably decay and suffer from a number of defects. Without repairs, these defects lead to significant economic losses and threaten safety.

In order to maintain a high level of security and economic growth, the daily maintenance and inspection of tunnels is necessary. Traditional inspection methods mainly rely on tunnel wall images taken by inspection vehicles or inspectors [12]. Inspectors look

**Citation:** Wang, A.; Togo, R.; Ogawa, T.; Haseyama, M. Defect Detection of Subway Tunnels Using Advanced U-Net Network. *Sensors* **2022**, *22*, 2330. https://doi.org/10.3390/ s22062330

Academic Editors: Carlos Morón Fernández and Daniel Ferrández Vega

Received: 7 February 2022 Accepted: 13 March 2022 Published: 17 March 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

for deterioration such as cracks and leaks when taking images and the deterioration of tunnel walls is evaluated and repaired according to their conditions. This process is performed manually, and it takes much time and labor. Technologies that enable the automatic detection of defects are required to facilitate this process [13,14].

The standard strategy for supporting the inspection of subway tunnels is to construct a detector for the estimation of defects from tunnel wall images. Among all kinds of defects, automated crack detection has been studied for a long time, and various methods based on image processing have been proposed [15–19]. Recently, in the field of computer vision, the performance of image recognition has been significantly improved with the emergence of deep learning, which has been useful for various tasks [20–24]. Therefore, it is expected that image recognition technology will enable the development of a detector that can automatically identify defects in infrastructures.

Deep learning-based methods have achieved higher performances in detecting defect in infrastructures than traditional methods that use handcrafted image features [25–27]. However, when applying deep learning methods to real-world problems, various characteristics and situations have to be considered. Since there are various kinds of defects in subway tunnels such as cracks, cold joints, and leakages, existing deep learning methods cannot be directly applied to this task. Specifically, the following problems need to be addressed to improve detection performance:
