关键词: UNet++, 裂缝检测, 非对称卷积, 多尺度特征融合, 注意力机制

Abstract: To ensure road quality and safety, automated crack detection is crucial for the maintenance of concrete pavements. To address the issue of pixel information loss caused by excessive down-sampling in existing deep learning-based methods, a concrete crack detection network based on progressive context interaction and attention mechanisms was proposed. First, an optimized UNet++ was used as the backbone, where asymmetric convolution blocks were applied to enhance feature extraction. Second, Progressive Context Interaction Mechanism (PCIM) was introduced to efficiently capture and fuse multi-scale features across adjacent stages. Third, in the feature enhancement phase, the Attention Combination (AC) approach was used to improve feature representation capability. Finally, in the feature fusion phase, a Multi-Semantic Attention Dynamic Fusion Module (MADFM) was utilized to enhance detail recovery and retention. The testing results on three public datasets (DeepCrack, Crack500, and CFD) show that, compared to DeepCrack, CrackFormer, and PAF-Net (Progressive and Adaptive Feature Fusion Network), the proposed method achieves superior performance. The F-score on the DeepCrack test set has been improved by 1.33, 5.07, and 3.93 percentage points, respectively; on the Crack500 test set, it has been enhanced by 3.04, 4.35, and 0.82 percentage points, respectively; and on the CFD test set, it has been increased by 3.03, 6.00, and 4.73 percentage points, respectively. These results fully prove that the proposed method has achieved enhanced accuracy in crack detection and exhibits excellent robust performance across different test sets.

Key words: UNet++, crack detection, asymmetric convolution, multi-scale feature fusion, attention mechanism