An Efficient Lightweight Method for Steel Surface Defect Detection.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Surface defects are common in steel production, posing challenges for traditional detection methods.
• Accurate and efficient defect detection is crucial for quality control in the steel industry.

Purpose of the Study:

• To develop an advanced deep learning model for precise steel surface defect detection.
• To improve upon existing YOLOv11 capabilities for detecting complex defects with greater efficiency.

Main Methods:

• Proposed LCED-YOLO model based on YOLOv11 architecture.
• Incorporated C3K2-MSE module for enhanced edge information extraction.
• Introduced LDConv and a lightweight decoupling head for model optimization.
• Utilized a learnable attention factor with CIoU loss for improved localization of difficult samples.

Main Results:

• LCED-YOLO achieved improved mAP50 scores of 79.8% on NEU-DET and 70.3% on GC10-DET datasets, outperforming YOLOv11.
• Reduced model parameters by 19% and floating-point operations by 23%.
• Demonstrated enhanced detection capability, particularly for challenging defect samples.

Conclusions:

• LCED-YOLO offers a lightweight yet highly accurate solution for steel surface defect detection.
• The proposed model effectively addresses the limitations of traditional methods and meets industrial demands for precision and efficiency.