Crack segmentation on steel structures using boundary guidance model

Cracks are an essential indicator of infrastructure degradation, and achieving high-precision, pixel-level crack segmentation is a common goal for artificial intelligence (AI)-enabled data processing. This study examines the inherent characteristics of cracks to introduce boundary features of cracks into crack identification and then builds a boundary guidance crack segmentation model (BGCrack), which includes four stages: backbone, boundary feature modeling, global feature modeling, and optimization of joint features. Some lightweight but effective modules are specifically designed and embedded in BGCrack to help achieve this goal. In particular, a high-frequency information enhancement (HFIE) module is designed for the edge modeling stage to better extract boundary features of cracks. A global information perception (GIP) module that combines frequency domain modeling and a Transformer unit (time domain modeling) is developed to help model global contexts and long-term dependencies. The ablation studies prove the validity of the model designs including the gradient loss function and, in particular, the boundary feature modeling, as intended in this study. In the experiments, among different models, BGCrack uses the fewest parameters and relatively low computation powers, but it has the best performance in all assessment measures. Furthermore, this paper also open-sources the code and the steel crack dataset, aiming to promote unified and impartial benchmarking.