DSCE-CrackNet: On combining collaborative feature fusion and edge refinement for automatic crack segmentation

Cracks are widely found on the surface of structures. Accurate crack detection is important for reducing structural risks. However, cracks have complex morphological features and thin edges, making the detection results incomplete. Moreover, noises such as textures and scratches on the structure's surface interfere with model detection and lead to false alarms. To address such issues, the Dual-Stream Collaborative and Enhancement Crack Network (DSCE-CrackNet) is proposed to extract the global morphology and local detail features of cracks by combining the advantages of transformer and convolution. Specifically, a Global–Local Collaborative Fusion Module (GLCFM) is designed to eliminate redundant information and adaptively fuse long-range global context with thin local structural features, enabling the model to better distinguish true cracks from background textures and surface scratches that resemble cracks. In addition, a Multi-Scale Feature Enhancement-Based Edge Refinement Module (MERM) is developed to produce multi-scale crack feature maps and enhance edge local information, which sharpens thin crack boundaries and improves the continuity and completeness of segmented crack structures. Extensive experiments involving eleven state-of-the-art methods on five public crack datasets demonstrate the effectiveness and generalization capability of DSCE-CrackNet. On the Crack500 and DeepCrack datasets, DSCE-CrackNet achieves the best overall performance with Dice scores of 70.6% and 83.5%, respectively, yielding improvements of +4.0% and +4.9% over CrackFormer on the same datasets. Competitive and outstanding results are also attained on Rissbilder, CFD, and GAPs384. These quantitative gains confirm that the proposed GLCFM and MERM modules effectively strengthen crack feature representation and suppress background noise in complex structural scenes.