Multi-scale feature fusion network for pixel-level pavement distress detection

作     者：Zhong, Jingtao Zhu, Junqing Huyan, Ju Ma, Tao Zhang, Weiguang 

作者机构：Southeast Univ Sch Transportat Nanjing 211189 Peoples R China 

出 版 物：《AUTOMATION IN CONSTRUCTION》 (建造自动化)

年 卷 期：2022年第141卷第0期

核心收录：

学科分类：08[工学] 0813[工学-建筑学] 0814[工学-土木工程] 

基　　金：National Key Research and Development Project of China [2020YFB1600102] Natural Science Foundation of Jiangsu Province [BK20180149] 

主　　题：Deep learning Encoder-decoder architecture Pavement distress Feature fusion Semantic segmentation Unmanned aerial vehicle (UAV) 

摘      要：Automatic pavement distress detection is essential to monitoring and maintaining pavement condition. Currently, many deep learning-based methods have been utilized in pavement distress detection. However, distress segmentation remains as a challenge under complex pavement conditions. In this paper, a novel deep neural network architecture, W-segnet, based on multi-scale feature fusions, is proposed for pixel-wise distress segmentation. The proposed W-segnet concatenates distress location information with distress classification features in two symmetric encoder-decoder structures. Three major types of distresses: crack, pothole, and patch are segmented and the results were discussed. Experimental results show that the proposed W-segnet is robust in various scenarios, achieving a mean pixel accuracy (MPA) of 87.52% and a mean intersection over union (MIoU) of 75.88%. The results demonstrate that W-segnet outperforms other state-of-the-art semantic segmentation models of U-net, SegNet, and PSPNet. Comparison of cost of model training and inference indicates that W-segnet has the largest number of parameters, which needs a slightly longer training time while it does not increase the inference cost. Four public datasets were used to test the generalization ability of the proposed model and the results demonstrate that the W-segnet possesses well segmentation performance.