MSFT-YOLO: Improved YOLOv5 Based on Transformer for Detecting Defects of Steel Surface

Area of Science:

• Computer Vision
• Artificial Intelligence
• Industrial Automation

Background:

• Intelligent inspection systems are crucial for industrial automation.
• Accurate and real-time object detection is a key challenge in computer vision for industry.
• Steel surface defect detection significantly impacts productivity and product quality.

Purpose of the Study:

• To address challenges in industrial object detection, including background interference, similar defect categories, scale variations, and poor small defect detection.
• To introduce an improved object detection model, MSFT-YOLO, for industrial defect detection scenarios.
• To enhance the accuracy and real-time capabilities of intelligent detection systems for industrial applications.

Main Methods:

• Developed the MSFT-YOLO model, an enhanced one-stage detector.
• Integrated a Transformer (TRANS) module into the backbone and detection heads for global feature integration.
• Implemented multi-scale feature fusion and employed data augmentation and multi-step training strategies.

Main Results:

• MSFT-YOLO achieves real-time detection performance.
• Demonstrated an average detection accuracy of 75.2% on the NEU-DET dataset.
• Achieved a 7% improvement over YOLOv5 and an 18% improvement over Faster R-CNN.

Conclusions:

• The MSFT-YOLO model offers significant improvements for industrial steel surface defect detection.
• The integration of Transformer modules and multi-scale feature fusion enhances detection capabilities.
• The proposed model provides an advantageous and inspiring solution for intelligent industrial inspection systems.