Depth-aware RGB-D concrete crack segmentation and quantification using progressive cross-modal attention

• Propose cross-modal feature fusion network for crack segmentation using RGB-D input. • Cross Attention Fusion module for bidirectional fusion of RGB and depth features. • A depth-assisted quantification method for crack length, width and depth estimation. • RGB-D cross fusion architectures are proposed for CNN and Transformer backbones. Cracks in infrastructures, such as concrete structures and pavements, pose significant risks to structural safety and durability. The development of crack geometry provides critical information of structural reliability hence there is need, recommended by standards, to precisely quantify the crack details, such as crack length, width or depth. Although deep learning has inspired automated crack detection, its capacities in profiling the crack geometry is still in doubt since most methods that rely solely on RGB images face challenges in field conditions with low contrast, surface contamination, and complex textures. Such conditions often result in blurred boundaries and unreliable geometric measurements, limiting their applicability in practice. To address these challenges, this study proposes a progressive Cross-Modal Fusion Transformer (CMF-Former) that integrates RGB and depth modalities through hierarchical representation and adaptive feature interaction. The network separately models RGB and depth representations to retain modality-specific features, and introduces a progressive cross-modal attention mechanism to adaptively fuse complementary information across semantic stages. A multi-scale decoder is used to further facilitate accurate crack localization and restoration. Additionally, a depth-assisted quantification method is developed by leveraging depth information to automatically estimate distance and spatial scale, enabling direct measurement of crack geometric features. Experimental results show that CMF-Former achieves a highest mIoU of 86.51%, outperforming other RGB-based and RGB-D based models. In addition to segmentation performance, the proposed RGB-D framework notably enhances geometric quantification. For crack width estimation, the proposed method achieved an average Root Mean Square Error (RMSE) of 1.167, representing a substantial improvement compared to other RGB-based methods. Moreover, the relative error rates for crack length and depth estimation are 2.19% and 6.188%, respectively, demonstrating improved accuracy in capturing crack morphology.