Innovative method for pavement multiple damages segmentation and measurement by the Road-Seg-CapsNet of feature fusion. (21st March 2022)

• Record Type: Journal Article
• Authors: Dong, Jiaxiu; Wang, Niannian; Fang, Hongyuan; Hu, Qunfang; Zhang, Chao; Ma, Baosong; Ma, Duo; Hu, Haobang
• Is Part Of: Construction & building materials. Volume 324(2022)
• Publication Date: 2022-03-21
• Subjects: Multiple damage -- Road-Seg-CapsNet -- StyleGAN -- Feature fusion -- Intelligent segmentation; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2022.126719
• Languages: English
• ISSNs: 0950-0618

Abstract

Graphical abstract: Highlights: A novel method is developed for multiple pavement damages segmentation of asphalt images. The model architecture and training scheme are designed to achieve high accuracy. Using padded convolution and feature fusion to optimize the model, the segmentation accuracy of damage was improved. More efficiency and higher accuracy than other deep learning-based pavement damages segmentation methods. Abstract: Deep learning algorithms have achieved promising results in the field of pavement damage segmentation. Since the texture features of map crack damage are complex, existing segmentation methods may experience the problem of edge information loss for the damaged area. In this study, we therefore propose Road-Seg-CapsNet, a feature fusion model combined with StyleGAN, to segment multiple complex forms of damage to asphalt pavements. Firstly, a data enhancement algorithm based on StyleGAN is developed to amplify the dataset. Next, padding convolution is used in the convolution layer of Road-Seg-CapsNet to retain more image edge information, which effectively solves the problem of information loss at the edges of areas of damage with map cracks. The feature map in the upsampling layer is fused with the features of the convolutional layer to improve the segmentation accuracy of the model. An optimised dynamic routing algorithm is then applied, which reduces the model parameters and improves the efficiency of model training. Our experimental results show that the segmentation mAP of the proposed model can reach 0.942. To verify the superiority of the proposed method, a comparative experiment with FCN, Mask R-CNN and optimised Mask R-CNN is carried out. Based on the segmentation results, a quantification and measurement experiment is carried out, and the minimum accuracy for the damaged areas is found to be 0.903.