Dynamic Dual-Branch Encoder and Deformable Spatial Focusing for Accurate Pavement Crack Segmentation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Pavement crack segmentation is vital for road safety, maintenance, and lifespan. Automated detection using computer vision offers efficiency gains but faces challenges due to complex crack features.
• Existing methods struggle with irregular crack distributions, shapes, and textures, impacting segmentation accuracy.
• Smart city initiatives require robust infrastructure monitoring, highlighting the need for advanced pavement analysis techniques.

Purpose of the Study:

• To develop an accurate and efficient pavement crack segmentation network (PCSNet).
• To address the challenges posed by complex visual features and irregular distributions in pavement crack images.
• To improve the generalisation and localisation capabilities of pavement crack detection models.

Main Methods:

• Proposed a novel pavement crack segmentation network (PCSNet) incorporating a dynamic dual-branch encoder and a deformable spatial focusing module.
• The dual-branch encoder utilizes pre-trained and self-trained branches for general and specific feature extraction, respectively.
• Dynamic feature fusion and deformable spatial focusing modules were employed to enhance feature extraction and refine crack morphology.

Main Results:

• PCSNet achieved high performance metrics on the DeepCrack dataset, including 85.34% precision, 86.16% recall, 85.75% F1 score, and 75.23% Mean Intersection over Union.
• The proposed network outperformed all comparative methods in pavement crack segmentation tasks.
• The dynamic dual-branch encoder and deformable spatial focusing effectively improved feature extraction and model generalisation.

Conclusions:

• PCSNet demonstrates superior performance in pavement crack segmentation compared to existing methods.
• The developed network offers a promising solution for automated road inspection, contributing to enhanced traffic safety and infrastructure management.
• The study validates the effectiveness of dynamic feature fusion and deformable spatial focusing for improving computer vision-based pavement analysis.