A Deep Supervised Pavement Crack Detection Network with Multiscale Feature Fusion and Feature Learning

Cracks are one of the most serious road diseases. Accurate detection of road cracks is an important part of road maintenance. With the development of modern society, infrastructure is increasing day by day. Traditional manual detection is not suitable for modern large-scale crack detection because it is time consuming and depends on people’s subjective consciousness. Due to the complex background and the complexity of the cracks, high-precision autonomous crack detection is still a challenging task. To solve the above problems, this paper takes the attention U-Net as the baseline network, and proposed a deep supervised pavement crack detection network for end-to-end crack detection. To realize effective feature fusion, an attention fusion block (AFB) is proposed for multi-scale feature aggregation. Meanwhile, based on the advantages of feature learning scheme, a progressive fusion block (PCF) is also proposed to capture more global context information. Faced with the problem of limited receptive fields, a multiple receptive field fusion (MRFF) block is proposed for multi-scale feature representation. Experiments on two public datasets show that the proposed network has obtained a better segmentation performance than the mainstream segmentation models.