Automatic steel girder inspection system for high-speed railway bridge using hybrid learning framework

The steel girder of high-speed railway bridges requires regular inspections to ensure bridge stability and provide a safe environment for railway operations. Unmanned aerial vehicle (UAV)-based inspection has great potential to become an efficient solution by offering superior aerial perspectives and mitigating safety concerns. Unfortunately, classic convolutional neural network (CNN) models suffer from limited detection accuracy or redundant model parameters, and existing CNN-based bridge inspection systems are only designed for a single visual task (e.g., bolt detection or rust parsing only). This paper develops a novel bi-task girder inspection network (i.e., BGInet) to recognize different types of surface defects on girder from UAV imagery. First, the network assembles an advanced detection branch that integrates the sparse attention module, extended efficient linear aggregation network, and RepConv to solve the small object with scarce samples and complete efficient bolt defect identification. Then, an innovative U-shape saliency parsing branch is integrated into this system to supplement the detection branch and parse the rust regions. Smoothly, a pixel-to-real-world mapping model utilizing critical UAV flight parameters is also developed and assembled to measure rust areas. Finally, extensive experiments conducted on the UAV-based bridge girder dataset show our method achieves better detection accuracy over the current advanced models yet remains a reasonably high inference speed. The superior performance illustrates the system can effectively turn UAV imagery into useful information.