Pixel-level segmentation method of concrete cracks based on MU-Net

In order to improve the segmentation accuracy of complex crack and lightweight network, this paper proposes a concrete crack segmentation network named MU-Net (Modified U-Net). In the encoder part, depthwise separable convolution is used to reduce network parameters, and inverted residual structure and attention mechanism are combined to fully extract features while highlighting target features. The decoder part also introduces the depthwise separable convolution and attention mechanism, and improves the segmentation accuracy by integrating the deep and light layers of information. In order to strictly evaluate the effectiveness of the network, this paper constructs a dataset of concrete cracks, which contains complex backgrounds and a variety of cracks, which is more in line with the engineering practice. Experimental results of 5-fold cross validation show that the proposed semantic segmentation network has superior performance compared with other advanced semantic segmentation networks.