Cross-layer context boundary guided network for crack segmentation

Abstract

Recently, Convolutional Neural Networks (CNNs) and Transformers have been extensively investigated for concrete crack segmentation, achieving remarkable performance. However, most CNN-Transformer-based crack segmentation methods overlook the exploration of contextual relationships between adjacent layers, which are critical for enhancing crack perception. Moreover, current algorithms fail to fully exploit the physical characteristics of cracks (e.g., geometric shape and boundary correlations), leading to reduced segmentation performance on low-contrast boundaries. To address these issues, we propose a Cross-Layer Context Boundary Guided Network (CCBG-Net) for Crack Segmentation. Specifically, a Bidirectional Cross-layer Context-aware (BCCA) module is introduced, which extracts multi-scale features from adjacent layers and performs bidirectional feature fusion with the current layer to obtain the contextual relationships of adjacent layers for enhanced crack feature representation, especially thin cracks. Furthermore, a Boundary-Object-Guided Interaction (BOGI) module is developed to decouple boundary information and guide crack features through global-channel interaction to optimize boundary contours, providing discrimination ability for crack boundaries. Experimental results on several challenging benchmark datasets demonstrate that our CCBG-Net outperforms state-of-the-art crack segmentation methods. The code is available at https://github.com/zty-acc/CCBG-Net.