Efficient CrackUNet: hierarchical spatial-channel attention with multi-scale fusion for pavement and bridge crack segmentation

Abstract Automated crack segmentation serves as a cornerstone for proactive infrastructure maintenance, yet prevailing techniques struggle to achieve the necessary balance between real-time processing demands and sub-pixel segmentation precision. This study introduces Efficient CrackUNet, an encoder-decoder architecture featuring three key innovations. The model incorporates a hierarchical progressive convolution module for enhanced multi-scale feature extraction, complemented by a spatial-channel collaborative attention mechanism that improves crack feature discriminability, further performance gains are achieved through an optimized multi-scale semantic feature fusion module that precisely integrates encoder-decoder features. Extensive experimental evaluations on multiple low-resolution benchmarks demonstrate that the model achieves 79.77% mIoU and 76.62% F1-score while shows particular robustness. On the high-resolution ACPCrack-400 dataset, it reaches 85.37% mIoU and 57.06 FPS, supporting real-time applications. With only 6.681 million parameters and 50.72 GFLOPs computational complexity, representing an 80.6% parameter reduction compared to baseline UNet, offering a practical solution for accurate and efficient crack segmentation.