Weakly Supervised Fatigue Crack Detection in Steel Bridge Girders Using a Proposed Two-Stage Network Training with a Segmentation Refinement Module

Existing semantic segmentation methods for fatigue cracks in steel bridge girders are fully supervised and thus demand manual annotation of pixel-level labels, which is time-consuming. Recently, there have been remarkable developments in semantic segmentation under image-level tag supervision. However, these weakly supervised approaches are still inferior to the fully supervised manner in terms of accuracy. To mitigate this gap, this paper commits to improving the correlation between high-level semantics to low-level appearance. A two-stage training manner with a segmentation refinement module for progressively refining pseudolabels and training the segmentation network was proposed. First, an activation modulation and recalibration scheme was recommended, which leverages a spotlight branch and a compensation branch to locate both the discriminative and less-discriminative object regions. Then, the generated pseudolabels were used as supervision to train the segmentation network in the proposed two-stage manner. In the first stage, the network was pretrained to learn all essential information and provide a basic segmentation performance, aiming to facilitate network convergence in the following training. To develop the inference quality, in the second stage, the pretrained network was further trained recursively with the designed segmentation refinement module to improve the labels using two postprocessing algorithms between each iteration. Overall, our method achieves comparable inference results to fully supervised approaches while significantly reducing annotation workload, which improves the efficiency of routine bridge inspection.