Enhanced YOLOv8 for efficient road damage detection with spatial-channel reconstruction and multi-scale attention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Engineering

Background:

• Road damage detection is crucial for safety and intelligent transportation systems.
• Current deep learning methods face challenges with computational cost and feature extraction in complex road conditions.

Purpose of the Study:

• To develop an enhanced object detection network for road damage detection using the YOLOv8 architecture.
• To improve feature extraction and multi-scale feature fusion for more accurate and efficient road damage identification.

Main Methods:

• Integration of Spatial and Channel Reconstruction Convolution (SCConv) module into the YOLOv8 backbone to refine feature representation.
• Incorporation of Efficient Multi-Scale Attention (EMA) module into the YOLOv8 neck for adaptive spatial-channel attention.
• Utilizing YOLOv8l as a baseline and evaluating SCConv and EMA module integration strategies.

Main Results:

• The enhanced YOLOv8 model demonstrated superior detection accuracy compared to the baseline.
• The proposed network maintained computational efficiency, addressing limitations of existing methods.
• Optimized integration of SCConv and EMA modules led to improved road damage detection performance.

Conclusions:

• The enhanced YOLOv8 network offers a promising solution for accurate and efficient road damage detection.
• The integration of SCConv and EMA modules effectively addresses computational and feature extraction challenges.
• This approach contributes to advancements in intelligent transportation systems and road maintenance.