Harris Hawks-tuned severity-aware YOLOv8 instance segmentation framework for vehicle damage assessment

Area of Science:

• Computer Vision
• Machine Learning
• Automotive Engineering

Background:

• Accurate vehicle damage severity assessment is vital for road safety, insurance, and inspections.
• Current computer vision methods struggle with subtle visual variations and irregular damage images, hindering precise severity discrimination.

Purpose of the Study:

• To develop a severity-sensitive YOLOv8-based instance segmentation system for automated vehicle damage severity assessment.
• To address challenges in discriminating precise damage levels due to visual variations and data disparity.

Main Methods:

• Implemented a YOLOv8 instance segmentation system incorporating curriculum learning for progressive training.
• Utilized Harris Hawks Optimization (HHO) for hyperparameter tuning with severity-sensitive objectives.
• Applied per-class confidence thresholding to optimize precision and recall for different severity levels.

Main Results:

• The proposed system demonstrated stable convergence and competitive performance with Box mAP50 of 0.271 and Mask mAP50 of 0.135.
• The lightweight YOLOv8s-Seg model outperformed a larger YOLOv8m-Seg baseline, indicating practical applicability.
• Strong detection of severe damages was observed, though subtle surface defects remain a challenge.

Conclusions:

• The developed framework provides an effective, understandable, and implementable algorithm for automated vehicle damage severity analysis.
• The severity-sensitive approach enhances the precision of automated vehicle inspection systems.
• Further research may focus on improving detection of subtle surface defects.