RER-YOLO: improved method for surface defect detection of aluminum ingot alloy based on YOLOv5.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Aluminum ingot alloy quality is crucial for industrial production and intelligent manufacturing.
• Traditional defect detection methods suffer from low precision and slow processing rates.
• Surface defect inspection of aluminum ingot alloys is vital for industrial engineering.

Purpose of the Study:

• To address limitations in traditional aluminum ingot alloy surface defect detection.
• To propose an improved YOLOv5-based model for enhanced detection accuracy and speed.
• To validate the effectiveness of the RER-YOLO model on an aluminum ingot alloy dataset.

Main Methods:

• Image preprocessing techniques including random rotation, translation, contrast, and brightness transformations were applied.
• A Res2Net multi-scale feature extraction block replaced the C3 block in YOLOv5.
• An over-parameterization-based re-parameterized convolutional block was integrated into the Res2Net residual and baseline models.

Main Results:

• The RER-YOLO model achieved a mean average precision of 75.1% on the aluminum ingot alloy dataset.
• This represents a 4.9% improvement over the conventional YOLOv5 without increased inference delay.
• Detection accuracy for burr defects, often difficult to extract, improved by 12.7%.

Conclusions:

• The RER-YOLO model offers significant improvements in detecting surface defects in aluminum ingot alloys.
• The proposed model enhances generalization capacity and fitting ability while maintaining inference speed.
• This study provides valuable insights for advancing automated surface defect inspection in industrial settings.