Dynamic multi-scale fusion network for road damage detection in complex street views.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Road surface deterioration due to traffic and climate change necessitates effective road damage detection (RDD).
• Existing RDD methods struggle with multi-scale features and accurate localization of irregular road damages like cracks.

Purpose of the Study:

• To propose a novel RDD method, the dynamic multi-scale fusion network (DMSFNet), to address limitations in feature extraction and localization.
• To enhance the accuracy and efficiency of road damage detection in complex street scenes.

Main Methods:

• Developed DMSFNet based on YOLOv11, incorporating a group fusion block (GFB) for scale-specific feature modeling.
• Introduced a multi-branch feature fusion (MBFF) module with dynamic snake convolution for geometric feature extraction.
• Integrated lightweight efficient channel attention (ECA) for multi-branch feature fusion.

Main Results:

• DMSFNet demonstrated effectiveness on SVRDD, RDD2020, and USRDD datasets.
• The method achieved a favorable trade-off between detection accuracy and computational efficiency.
• GFB enhanced accuracy for damages across all sizes, while MBFF improved localization of irregular patterns.

Conclusions:

• DMSFNet offers a significant advancement in road damage detection, particularly for varied scales and irregular shapes.
• The proposed method contributes to improved traffic safety and pavement management through accurate and efficient RDD.