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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Structural Engineering

Background:

• Accurate crack segmentation is vital for infrastructure safety and monitoring.
• Existing methods like MixSegNet achieve high accuracy but require extensive training.
• Limitations include poor generalization and high computational cost in operational settings.

Purpose of the Study:

• To develop an efficient crack segmentation network that reduces training time and computational cost.
• To maintain high segmentation accuracy comparable to state-of-the-art models.
• To enhance applicability in real-time infrastructure inspection.

Main Methods:

• Proposed Strategic SA-UNet, integrating a U-Net based CNN with a Self-Attention Block.
• The Self-Attention Block fuses local features with global context between encoder and decoder.
• Evaluated on publicly available crack datasets.

Main Results:

• Strategic SA-UNet achieved segmentation accuracy comparable to MixSegNet.
• Reduced training time by 83%, FLOPs by 63%, and model parameters by 96%.
• Demonstrated superior training efficiency with a high mean Intersection over Union (mIoU) in fewer epochs.

Conclusions:

• Strategic SA-UNet is an efficient and accurate crack segmentation model.
• The model is well-suited for real-time infrastructure inspection and structural monitoring.
• Offers a practical solution for continuous data processing in operational environments.