A Surface Crack Detection Method of Spillway Tunnel Based on an Improved U-Net Network

Abstract

Aiming at the problem of small defects that are difficult to detect in the automatic detection of cracks on the surface of the concrete lining of flood tunnels, resulting in low accuracy, this paper proposes an improved U-Net network model, which introduces a spatial attention mechanism to realize end-to-end detection of cracks on concrete lining surfaces. The spatial attention mechanism (clique blocks) is integrated into the U-Net pipeline to solve the problem that it is difficult to obtain low-level fine-grained features. The Squeeze-and-Excitation (SE) block is utilized for the spatial weighting of feature map during the down and up sampling, which realizes the fusion operation between low-level and high-level features and boosts the perception of detection model on small defects. We build a spillway tunnel crack dataset, compare the improved U-Net with the original one, and conduct an out-of-sample generalization experiment on the CFD dataset. The pixel accuracy (PA), mean pixel accuracy (MPA) and mean intersection over union (MIOU) of the improved U-Net are improved to 92%, 86%, and 87%, respectively. The proposed model has a good generalization ability which can effectively meet the requirements of crack detection in spillway tunnels.