A crack detection network with multi-channel attention and enhanced information interaction

Roads frequently experience cracks. It adversely impact the safe passage of vehicles and pedestrians, and have the potential to alter the road’s structure. To address this issue, we propose a novel crack detection network. The network constructs multi-channel attention and enhanced information interaction mechanisms to capture more granular semantic information. In our network, each convolutional layer is followed by a convolution combining asymmetric convolutions and criss-cross attention to enhance the feature maps post-convolution. This is followed by spatial and channel reconstruction convolutions and shuffle attention to optimize the generated side-output features. By extensively mining features from the deep network and ingeniously integrating bottom-level and top-level features through a new feature fusion module. The network achieves precise crack prediction results. Extensive experiments on the general-purpose crack image datasets Crack500, CFD and DeepCrack demonstrate the model’s effectiveness. In these three datasets, F1-score values of 0.734, 0.635, and 0.881, MIoU values of 0.773, 0.726 and 0.888.