Research on Crack Image Segmentation Method Based on UNet and CBAM

Crack detection is crucial for structural health monitoring, but traditional methods often struggle with complex backgrounds and fine cracks. To address this issue, this paper proposes an improved UNet-based crack segmentation model by integrating the Convolutional Block Attention Module (CBAM) at the feature concatenation stage. Unlike conventional approaches that apply CBAM solely in the encoder or decoder, our method enhances multi-scale feature fusion by introducing channel and spatial attention during skip connections. This allows the network to focus on crack-related features while suppressing irrelevant background noise. The dual-attention mechanism dynamically recalibrates feature representations across both channel and spatial dimensions, significantly improving the model's sensitivity to fine crack patterns while maintaining structural coherence. Experiments on concrete crack datasets demonstrate that our UNet-CBAM model achieves higher accuracy and robustness compared to baseline methods, particularly in handling low-contrast and thin cracks. The proposed architecture also shows enhanced generalization capability across diverse crack types and surface textures. The results indicate that attention mechanisms at feature concatenation can effectively improve segmentation performance in challenging scenarios.