CrackNet: A Backbone Network for Pavement Crack Analysis

Currently, the methods for road crack recognition based on deep learning typically treat crack recognition as a general image detection or classification task without specific analysis of crack images. In this paper, we propose a novel deep learning crack recognition network called CrackNet, which is based on the Transformer and can serve as the backbone network for crack detection or classification. Compared to advanced Transformer models such as PVTv2, CrackNet has several advantages. Firstly, it allows the model to pay more attention to the features of cracks and their surrounding regions. This is achieved by CrackNet explicitly extracting crack features from the images through wavelet transformation, providing reliable mask information for gated convolutions, and then using gated convolutions to better extract features in the crack region. Additionally, by incorporating convolutions into self-attention blocks, CrackNet can compensate for the insufficient capability of Transformer models in extracting local features. To evaluate the effectiveness of our approach, we conducted crack classification experiments on a large-scale asphalt road crack dataset called CQU-BPMCC, and crack detection experiments on a dataset called RDD2022. The experimental results on both datasets demonstrate the superiority of CrackNet in crack recognition comnared to other general backbone network methods.