Abstract: When the general segmentation model is applied to the apparent cracks in the dam face concrete, the network suffers the problem of depth increasing that leads to excessive model parameters and certain loss of effective crack features. To reduce network memory occupation and feature loss, this paper develops a lightweight crack segmentation method based on a convolutional neural network. The network adopts an encoding-decoding structure, and uses a depth-separable convolution module and a lightweight feature extraction module to construct a cascade encoder; it is equipped with a decoder to fuse cross-scale information in the second stage of the encoder and to reconstruct the pixel-level geometric information lost in feature extraction to improve the accuracy of network segmentation. The experimental results show the model size of the network trained on the crack dataset of dam face concrete is 10.8 MB or a size reduction of 90.8% from U-Net, with its PA of 73.3% and IoU of 85.4%. The results verify the network is feasible in dam face crack segmentation and useful for improving the efficiency of dam face detection and maintenance.

Key words: deep learning, convolutional neural networks, crack segmentation, network lightweighting, deep separable convolution