Crack segmentation on concrete, steel, and asphalt surfaces remains challenging due to irregular crack patterns, low contrast, and noise interference, particularly in complex environments. Although deep neural network–based methods show promise, they often struggle to balance fine-grained feature extraction with contextual understanding. Moreover, no unified model effectively detects cracks across concrete, steel bridges, and asphalt pavements on bridge decks, while most existing models are too large for edge deployment. This paper introduces CrackSeg-GWD, a lightweight encoder–decoder model integrating Group Normalization, Weight-Standardized Convolutions, DropBlock regularization, and a Symmetric Unified Focal Loss to enhance stability, reduce overfitting, and handle class imbalance. With only 0.414 M parameters and 0.849 GFLOPs, it achieves high accuracy with low computational cost. Evaluated on five public datasets, SteelCrack, YCD, Crack500, DeepCrack, and Ozgenel, CrackSeg-GWD outperforms ten state-of-the-art models, achieving consistent gains across five metrics and confirming its suitability for real-time structural monitoring and construction automation.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号