Cascade refinement extraction network with active boundary loss for segmentation of concrete cracks from high-resolution images
Lu Deng, Huaqing Yuan, Lizhi Long, Pang‐jo Chun, Weiwei Chen, Hong-Hu Chu
Abstract
Accurate extraction of cracks is important yet challenging in bridge inspection, particularly that of tiny cracks captured from high-resolution (HR) images. This paper presents a crack-boundary refinement framework (CBRF) for meticulous segmentation of HR crack images. First, a triple-scale feature extraction module is designed to enhance the representation of miniscule-crack pixels. Then, a cascade operation involving global and local steps is adopted to conduct the refinement. In addition, an active boundary loss is introduced into the training process to solve the semantic inconsistency of crack boundary areas. The first HR crack image dataset is established to thoroughly evaluate the CBRF. Finally, an unmanned aerial vehicle (UAV)-based case study is conducted on the Yinpenling Bridge, which further confirms the practicality of the CBRF in improving the safety and efficiency of UAV-based bridge detection while ensuring the accuracy.