Automated masonry spalling severity segmentation in historic railway tunnels using deep learning and a block face plane fitting approach

Abstract

Masonry lined tunnel condition assessment is a predominantly manual process. It consists primarily of a visual inspection followed by a lengthy and subjective manual defect labelling process. There is therefore much potential for automation. Masonry spalling is a key indicator of a masonry tunnel’s condition. To obtain actionable detail about a tunnel’s condition, it is also necessary to determine the spalling severity, defined by the depth of spalling. This study presents an automated workflow to identify the depth of spalling from masonry tunnel 3D point cloud data obtained by lidar. Firstly, a tunnel point cloud is unrolled using a cylindrical projection and the points are rasterised into a 2D image taking pixel values of the offset of each point from the cylinder. Then, a 2D U-Net pretrained on both real and synthetic masonry lining data, is used to segment masonry joint locations to isolate individual blocks. A separate U-Net is used to segment areas of masonry damage and data obstructions, which are then masked out before a surface plane representing the theoretical undamaged surface location is fitted to each masonry block from the remaining points. This allows the depth of spalling to be measured directly. As a result, this method can automatically determine the depth of spalling despite the curved and often deformed nature of a masonry tunnel profile. Experiments show results competitive with those obtained by human assessors.