Optimizing surface defect detection with YOLOv9: the role of advanced backbone models

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Object detection models like YOLO are crucial for identifying surface defects.
• Optimizing these models is essential for efficient industrial inspection.

Purpose of the Study:

• To evaluate six distinct backbone networks integrated into the YOLOv9 framework for steel strip surface defect detection.
• To compare their performance in terms of accuracy, computational complexity, and efficiency.

Main Methods:

• Integrated ResNet50, GhostNet, MobileNetV4, FasterNet, StarNet, and RepViT backbones into the YOLOv9 framework.
• Conducted systematic evaluations on the NEU-DET and GC10-DET datasets.
• Compared backbones using metrics like mAP50, F1-score, parameter count, and GFLOPs.

Main Results:

• RepViT demonstrated superior overall performance with an mAP50 of 68.8% and a balanced precision-recall.
• GhostNet provided excellent computational efficiency with 41.2M parameters and 190.2 GFLOPs.
• Results were validated using YOLOv5-m, confirming consistency.

Conclusions:

• RepViT is recommended for high-accuracy surface defect detection.
• Lightweight architectures like GhostNet are suitable for real-time industrial applications.
• The study provides practical guidance for backbone selection in defect detection tasks.