Abstract
Crack detection is essential for structural safety inspection but remains challenging due to noise, illumination variations, and complex backgrounds. In this paper, we propose CrackNet, a segmentation network specifically designed for concrete crack detection. CrackNet integrates three key modules: a lightweight multi-scale convolution enhancement block (LightMSCBlock) in the encoder to capture both local details and global context, a SAF attention module embedded in skip connections for scale-aware feature fusion and edge refinement, and a multi-scale feature fusion (MSFF) module in the decoder to enhance feature integration while reducing information loss. Extensive experiments on three public datasets-CFD, Crack500, and DeepCrack-demonstrate that CrackNet consistently outperforms state-of-the-art methods. Specifically, on CFD, F1 and IoU improve by 6.37% and 7.1% over SegFormer; on Crack500, F1 increases by 3.86% compared with MobileNetV3-UNet; and on DeepCrack, F1 and IoU gains reach 5.7% and 2.5%, respectively. Ablation studies further confirm the complementary effectiveness of LightMSCBlock, SAF, and MSFF. Overall, CrackNet achieves superior accuracy and robustness, showing strong potential for real-world engineering applications. The code is available at the following link: https://github.com/xzz-ya/CrackNet.git.