MR-YOLO: An Improved YOLOv5 Network for Detecting Magnetic Ring Surface Defects

Area of Science:

• Materials Science
• Computer Vision
• Manufacturing Engineering

Background:

• Magnetic rings are crucial components in various industries, but manufacturing defects like cracks and adhesion are common.
• Existing defect detection methods, such as YOLOv5, suffer from high computational costs and large model sizes.
• Efficient and accurate defect identification is essential for ensuring the quality of magnetic ring production.

Purpose of the Study:

• To develop an enhanced, lightweight YOLOv5 (MR-YOLO) algorithm for identifying surface defects in magnetic rings.
• To address the computational and size limitations of conventional YOLOv5 in defect detection.
• To improve the efficiency and accuracy of magnetic ring defect identification in industrial settings.

Main Methods:

• Integrated Mobilenetv3 module into the YOLOv5 neck for reduced floating-point operations (FLOP) and enhanced feature expression.
• Applied Mosaic data augmentation to improve algorithm robustness.
• Incorporated SE attention module into the backbone and replaced SPPF module to focus on minor defects.
• Replaced CIoU loss with SIoU loss to boost network accuracy and training speed.

Main Results:

• Achieved a 59.4% reduction in FLOP and a 47.9% decrease in model parameters.
• Increased reasoning speed by 16.6% and reduced model size by 48.1%.
• Experienced a minimal drop of 0.3% in mean Average Precision (mAP), demonstrating high accuracy retention.

Conclusions:

• The proposed MR-YOLO significantly optimizes YOLOv5 for magnetic ring defect detection, offering a superior balance between efficiency and accuracy.
• This lightweight approach effectively reduces computational load and model size without compromising defect identification performance.
• MR-YOLO presents a viable and effective solution for real-time, high-volume defect inspection in magnetic ring manufacturing.