A high precision YOLO model for surface defect detection based on PyConv and CISBA.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Industrial production relies on defect detection for quality control.
• Existing methods struggle with diverse defects, complex backgrounds, and multi-scale small targets, reducing performance.
• Accurate and efficient defect detection is crucial for manufacturing.

Purpose of the Study:

• To propose the EPSC-YOLO algorithm for improved surface defect detection efficiency and accuracy.
• To address challenges posed by multi-scale small targets and complex backgrounds.
• To enhance the detection of diverse defect types in industrial products.

Main Methods:

• Introduced multi-scale attention modules and novel pyramid convolutions in the backbone network.
• Replaced traditional Non-Maximum Suppression (NMS) with Soft-NMS to minimize information loss and improve overlapping box detection.
• Designed a new Convolutional Attention module (CISBA) to boost small target detection in challenging environments.

Main Results:

• EPSC-YOLO demonstrated improved performance over YOLOv9c, with increases in precision and recall.
• Achieved superior accuracy and significant advantages in real-time detection compared to YOLOv10 and MSFT-YOLO.
• Validation on NEU-DET and GC10-DET datasets confirmed the algorithm's effectiveness.

Conclusions:

• EPSC-YOLO effectively enhances multi-scale defect identification and small target detection.
• The algorithm offers a robust solution for real-time surface defect detection in industrial settings.
• EPSC-YOLO presents a significant advancement in automated quality inspection systems.