Scan Path Optimization and YOLO-Based Detection for Defect Inspection of Curved and Glossy Surfaces.

Area of Science:

• Industrial Engineering
• Computer Vision
• Machine Learning

Background:

• Manual inspection of defects on glossy/curved surfaces is inconsistent and time-consuming.
• Existing automated methods struggle with the complexities of reflective and non-flat product surfaces.
• Need for a robust and efficient automated defect detection solution in mass production.

Purpose of the Study:

• To develop a novel defect detection framework for curved and reflective surfaces.
• To optimize laser scanning path generation for structured image acquisition.
• To evaluate the performance of deep learning models for automated defect identification.

Main Methods:

• Utilized a KEYENCE displacement laser sensor for data acquisition.
• Integrated Dijkstra's algorithm, Nearest Neighbor Algorithm, and Genetic Algorithm for scanning path optimization.
• Trained and compared YOLOv8, YOLOv9, YOLOv10, and YOLOv11 deep learning models on generated datasets.

Main Results:

• YOLOv11 demonstrated superior performance in defect detection, achieving an mAP50 score of 0.844.
• The proposed framework successfully generated structured image data for training deep learning models.
• YOLOv11 exhibited lower computational complexity and faster inference times compared to other YOLO architectures.

Conclusions:

• The developed framework provides a robust solution for automated defect inspection on challenging surfaces.
• YOLOv11 is highly effective for real-time defect detection in industrial applications.
• The integration of optimized scanning paths and deep learning enhances inspection accuracy and efficiency.