Computer Vision-Based Railway Viaduct Component Inspection Using Multi-task Multi-scale Deep Neural Network

Component inspection is the first step for the health assessment of railway viaduct. In current practical engineering, information from multiple sources is always taken into consideration simultaneously before assessments are implemented. Visual inspection of railway infrastructures by resorting to task-specific deep learning solutions has been investigated and leveraged actively in both research and practical engineering communities. However, in the context of railway infrastructure inspection, only few efforts were reported upon multi-task deep learning. In the present study, a multi-task deep learning method for component inspection of railway viaducts was proposed based on the state-of-the-art multi-scale deep neural network. Two tasks were referred for the component inspection: component segmentation and depth estimation. The multi-task neural network architecture was implemented by three modules that make full use of commonalities and correlations among tasks at different scales. Intersection over Union was reported to be more than 88% for component segmentation and root mean square error to be less than 1.6 m for depth estimation, with significant improvements. The multi-task model reported better performance with respect to both evaluation metrics and inference speed, presenting the effectiveness of the proposed method compared with conventional single-task deep neural networks.