MRHCNet: Multi Resolution Hybrid Convolution based U-Net for Crack Detection with Mix Dilated Linear Pooling and Laplacian Pyramid Mix Module

Abstract

Deep learning based methods have achieved the state-of-art results in crack detection. However, these models have some drawbacks. Firstly, these models often use down-sampling operation to reduce the resolution of features, which would cause feature information loss. Secondly, square filtering windows are used to extract strip crack features, which would merge unrelated feature information. Thirdly, these models are limited to learn detailed feature information and edge features of cracks, which would cause detail and edge feature information loss. To solve these problems, a Multi Resolution Hybrid Convolutional Module (MRHCM) is designed to avoid feature information loss caused by down-sampling. Additionally, a Mix Dilated Linear Pooling (MDLP) is proposed to capture the strip feature information of cracks. Furthermore, a Laplacian Pyramid Mix Module (LPMM) is designed to learn detailed feature information of cracks. Also, a Sobel Mix Cross Pooling (SMCP) is used to enhance the network?s ability of learning edge features. Finally, a Multi Resolution Hybrid Convolution based U-Net (MRHCNet) with MDLP, LPMM and SMCP is proposed by us. Experimental results show that our MRHCNet could achieve the accuracy of 0.931, 0.825 and 0.989 on the Cracktree200, CRACK500 and CFD datasets respectively, which is higher than that of the traditional models.