Strategic SA-UNet: Integrating self-attention blocks into U-Net for efficient crack segmentation

Accurate crack segmentation plays a crucial role in ensuring safety and mitigating disaster risks during road inspections and structural health monitoring. However, traditional image processing techniques often struggle with low detection accuracy and poor generalization performance due to the diverse morphology of cracks and the presence of background noise. To address these challenges, MixSegNet, a model that combines the strengths of convolutional neural networks (CNNs) and Transformers, has been proposed and demonstrated to achieve high segmentation performance. However, this enhanced precision comes at the cost of prolonged training cycles, which limits its applicability in operational environments such as infrastructure inspection, where new data must be acquired and processed continuously and rapidly. In this paper, to address this limitation, we propose Strategic SA-UNet (Strategically Integrated Self-Attention U-Net), a novel crack segmentation network. The model strategically integrates a computationally efficient U-Net based CNN with a Self-Attention Block between the encoder and decoder to effectively fuse local features with global context, thereby maintaining high segmentation accuracy while reducing training time and computational cost. Experimental evaluations on publicly available datasets demonstrate that Strategic SA-UNet achieves segmentation accuracy comparable to MixSegNet, while reducing training time by 83%, Floating Point Operations (FLOPs) by 63%, and Model Parameters by 96%. Furthermore, Strategic SA-UNet achieves a high mean Intersection over Union (mIoU) even with a small number of epochs, highlighting its superior training efficiency. These results suggest that Strategic SA-UNet is an efficient segmentation model, especially suitable for real-time infrastructure inspection and structural monitoring applications.