A U-Net-like full convolutional pavement crack segmentation network based on multi-layer feature fusion

Cracks are an important indicator of pavement health, and it is difficult to achieve pixel-level segmentation of small and thin cracks. The existing network often experiences false segmentation and missed segmentation. Accordingly, a novel end-to-end U-Net-like full convolutional crack segmentation network is constructed. First, we propose a multi-layer feature fusion module to aggregate the texture and semantic features at each stage of encoder, so that the network can find smaller and thinner crack. Second, we design a novel residual structure with a pointwise convolution. Each stage of the encoder and decoder incorporates a residual structure to facilitate the fusion of feature maps with different spatial dimensions. It can also prevent the gradient vanish in the network training process. Finally, we utilise the maximum unpooling to restore spatial structure in up-sampling, which exploits the indices of maximum feature value in down-sampling. Therefore, high-frequency information is better preserved to help accurately restore the details of crack edges. To verify the proposed network performance, experiments are carried out on four open datasets, the proposed network can achieve better performance among five classical networks.