Concrete Surface Crack Recognition Based on Coordinate Attention Neural Networks

In highway transportation infrastructure such as highways and tunnels, the proportion of concrete consumption is the highest, and concrete cracks are common concrete problems. Concrete cracks will greatly affect the bearing capacity and safety of the structure, easily leading to the interruption of transportation lines, causing great economic losses, and endangering personnel safety. Therefore, the effective identification and timely reporting of concrete cracks is of great significance for the maintenance of infrastructure such as roads and tunnels. In this paper, the CaNet, a deep learning network for identifying concrete cracks, is proposed, which takes ResNet50 as the backbone network. In order to capture the area with a small proportion of cracks, we added coordinate attention to the residual unit of ResNet50 to capture the cross-channel information, direction-aware information, and position-sensitive information from many vertical and horizontal directions so that the network can more accurately locate the narrow crack area. In experiments 3.2 and 3.3, the CaNet has an accuracy rate of 89.6%, which is higher than that of the compared network. In addition, the recall, F1 score, and precision of the CaNet network are 86%, 85%, and 87% , respectively. Therefore, the CaNet model is effective for identifying concrete cracks.