Comparison of Pre-Trained YOLO Models on Steel Surface Defects Detector Based on Transfer Learning with GPU-Based...

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Steel quality is critical in machinery; surface defects significantly impact it.
• Existing steel surface defect detection methods often lack real-time application focus and accessible datasets.
• There is a growing demand for efficient and accurate automated defect detection systems.

Purpose of the Study:

• To investigate the feasibility of state-of-the-art deep learning models, specifically YOLO variants, for real-time steel surface defect detection.
• To compare the performance of YOLOv5, YOLOX, and YOLOv7 on a small-scale dataset.
• To evaluate the real-time performance and accuracy trade-offs of optimized models on edge devices.

Main Methods:

• Trained YOLOv5, YOLOX, and YOLOv7 models on the NEU-DET dataset using a GPU RTX 2080.
• Evaluated model accuracy using mean Average Precision (mAP).
• Deployed trained models on Nvidia Jetson Nano and Xavier AGX, applying optimizations like TensorRT, precision reduction (FP16/INT8), and input size adjustment for real-time performance.

Main Results:

• YOLOX-s achieved the highest accuracy with 89.6% mAP on the NEU-DET dataset.
• Real-time deployment on Nvidia Jetson devices showed a trade-off between detection speed (fps) and accuracy (mAP) after optimization.
• Optimized models demonstrated reduced detection times suitable for real-time applications.

Conclusions:

• Deep learning models, particularly YOLOX, show significant promise for real-time steel surface defect detection.
• Model optimization techniques are crucial for balancing speed and accuracy in edge deployment scenarios.
• Further research can focus on larger datasets and more robust real-time detection strategies for industrial applications.