RSA-TransUNet: a robust structure-adaptive TransUNet for enhanced road crack segmentation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Deep learning advancements are critical for intelligent transportation safety, particularly in road crack segmentation.
• Existing methods struggle with fine crack textures, varied widths, blurred edges, and multi-class segmentation, limiting practical deployment due to high computational costs.

Purpose of the Study:

• To propose RSA-TransUNet, a novel model addressing limitations in current road crack segmentation techniques.
• To enhance the accuracy, robustness, and efficiency of road crack detection for intelligent transportation systems.

Main Methods:

• Introduced the Axial-shift MLP Attention (ASMA) mechanism for capturing long-range dependencies and multi-scale crack features.
• Developed the Adaptive Spline Linear Unit (ASLU) to improve adaptability to structural irregularities and microstructural variations.
• Implemented a Structure-aware Multi-stage Evolutionary Optimization (SMEO) strategy to enhance convergence speed and generalization performance.

Main Results:

• RSA-TransUNet demonstrated superior segmentation accuracy and robustness on Crack500, CFD, and DeepCrack datasets.
• The proposed ASMA and ASLU mechanisms effectively handled fine-grained textures, blurred edges, and width variations.
• The SMEO strategy significantly improved training efficiency and model generalization.

Conclusions:

• RSA-TransUNet offers a significant advancement in road crack segmentation, outperforming existing methods.
• The model's robustness and accuracy show strong potential for real-world intelligent transportation safety applications.
• This work provides a foundation for more reliable automated road inspection systems.