Lightweight semantic segmentation of complex structural damage recognition for actual bridges

• Published in: Structural health monitoring 2023-09, Vol.22 (5), p.3250-3269
• Main Authors: Xu, Yang, Fan, Yunlei, Li, Hui
• Format: Article
• Language: English
• ISSN: 1475-9217; 1741-3168
• DOI: 10.1177/14759217221147015

Although structural damage recognition has been extensively investigated using deep learning and computer vision (CV) techniques, the following limitations exist for real-world applications: (1) the accuracy heavily relies on a large volume of network parameters; (2) the sensitivity to tiny cracks is limited due to low contrast between microcrack and background pixels; (3) the robustness on complex cracks with various morphological features and surface disturbances is inadequate. To address these issues, this study proposes a lightweight, accurate, and robust semantic segmentation method of complex structural damage recognition for actual bridges. Firstly, a modified DeepLabv3+ model is established using the lightweight MobileNetV2 backbone and transposed convolutions to reduce parameter volume and enhance the recognition capability of local minor damages. Secondly, the depthwise separable convolution is utilized instead of the standard convolution to decouple the spatial and channel interactions of feature maps. Thirdly, a refined atrous spatial pyramid pooling (ASPP) module is constructed at the backbone end using multilevel dilated convolutions to expand the receptive fields. Finally, a piecewise synthetical loss function based on focal and dice losses is designed for different training stages. A total of 3226 actual crack images in different scales, resolutions, and scenes are utilized to verify the proposed method. The results show that the mean intersection-over-union for complex cracks in various real-world scenarios reaches 0.776 with significant reductions of 91.5% in parameter volume and 38.9% in recognition time. Comparative studies demonstrate the superiority of the proposed method over existing lightweight crack segmentation models based on SegNet and DenseNet. In addition, ablation experiments demonstrate the necessity and effectiveness of the MobileNetV2 backbone, refined ASPP module, and piecewise synthetical loss function. Moreover, the robustness and expandability of the proposed method on new structural damage categories (including concrete spalling, rebar exposure, and cable corrosion) are also verified.