Subway tunnels often suffer from surface pathologies such as cracks, corrosion, fractures, peeling, water and sand infiltration, and sudden hazards caused by foreign object intrusions. Installing a mobile visual pathology sensing system at the front end of operating trains is a critical measure to ensure subway safety. Taking leakage as the typical pathology, a tunnel pathology automatic visual detection method based on Deeplabv3+ (ASTPDS) was proposed to achieve automatic and high-precision detection and pixel-level morphology extraction of pathologies. Compared with similar methods, this approach showed significant advantages and achieved a detection accuracy of 93.12%, surpassing FCN and U-Net. Moreover, it also exceeded the recall rates for detecting leaks of FCN and U-Net by 8.33% and 8.19%, respectively.