Abstract: Automatic segmentation of various types of pavement distresss is of great significance for improving the level of highway maintenance management. The lack of large-scale, diverse-scenario, and multi-type training data for pavement distresss has reduced the generalization ability of deep learning models in complex scenarios, and limited the development of engineering applications for pavement distress extraction algorithms. To address this issue, this paper collects and establishes a dataset called CPCD for the segmentation of multiple types of pavement distresss, aiming to accurately segment pavement distresss from CCD images. The new dataset covers 7 common types of pavement distresss, including negative samples with texture-similar noise, totaling 6967 images. Based on this dataset, a distress segmentation convolutional neural network called CBAM-HRNet, which combines attention mechanisms with high-resolution networks is proposed to improve the accuracy of fine-grained distress segmentation. Experimental results show that the F1 score and mIoU of the proposed model on the CPCD dataset are 91.30% and 84.64% respectively. Its performance is significantly better than other mainstream convolutional neural network models with sequential structures. The research findings will provide certain support for the research on highway automation detection and sustainable development in our country.

Key words: deep learning, complex scenarios, CPCD, semantic segmentation, pavement distresss