Multi-scale Defect Detection from Subway Tunnel Images with Spatial Attention Mechanism

Abstract

With the development of modern transport systems, larger-scale subway tunnels are constructed worldwide. After the construction, maintenance becomes an important issue. Due to the limitations of the check time and the complex environment, it is required to detect defects accurately and efficiently. In subway tunnels, the main defects are cracks (small defects) and water leakage (large defects), which are hardly to achieve high precision detection for both. In this paper, we propose a new multi-scale defect detection network based on HRNet to solve this problem. We applied an atrous spatial pyramid pooling module and a spatial attention module to improve the capacity of defect detection. Through the experiments, we verify the effectiveness of our method.