Existing image segmentation algorithms face challenges related to low detection accuracy and a lack of specificity in crack detection. To address these challenges， this paper proposes an extended LG-Block module Extend-LG Block， which leverages a multi-scale feature fusion method. This module consists of multiple parallel dilated convolutions with different expansion rates. The number of branches and the expansion rate of dilated convolutions can be adjusted by parameters to change the size of its receptive field， and then extract and fuse crack features of different scales. By comparing the advantages and disadvantages of the network using a multi-scale feature fusion module in the deep layer and the network using a fixed scale structure for multi-scale feature fusion， a U-Net model with a variable scale structure named VS-UNet is proposed. The basic convolution Block in the UNet network is replaced by multiple Extend-LG blocks with different parameters. This structure performs multi-scale feature fusion in the shallow layer of the network， and the scale extracted by the multi-scale feature fusion module gradually decreases with the deepening of the network layer. This structure not only strengthens the detail feature extraction ability of the image while maintaining the original abstract feature extraction ability but also avoids the problem of increasing network parameters caused by the increase of convolution. Experiments are carried out on the DeepCrack dataset and CFD dataset. The results show that compared with the other two structures and methods， the proposed network with variable scale structure has higher detection accuracy and better segmentation effect for cracks of various sizes in visual experimental comparison. Finally， compared with other image segmentation algorithms， all indicators are improved to a certain extent compared with UNet， which proves the effectiveness of the improved network.