MCAMNet: A Novel Encoder-Decoder Network with Multi-directional Convolution and Attention Mechanism for Crack Detection

Accurate and timely detection and repair of pavement cracks is the key to road maintenance. Although the crack detection method based on image semantic segmentation have achieved good results till now, there are still some shortcomings, especially detection accuracy needing to be further improved. We propose a new network for crack segmentation based on encoder-decoder structure, called MCAMN et, which uses multi-directional convolution module instead of conventional convolution module for feature extraction, solving the problem of discontinuous and incomplete crack segmentation results. MCAMNet innovatively adds spatial attention block and efficient channel attention block after the encoder to reduce the interference of background and noise, solving the problem that the crack details and the segmentation results of small cracks are not fine. Additionally, we mix three loss functions of different levels by obtaining a mixed loss function to supervise the training process of MCAMNet, solving the problem of rough and fuzzy crack edges in the segmentation results. Experiments on four public datasets of pavement cracks prove that the F -score and mIo U of MCAMNet are better than existing methods.