GM-DETR: Research on a Defect Detection Method Based on Improved DETR.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Defect detection is crucial for industrial intelligence.
• The DETR (Detection Transformer) model enabled end-to-end defect detection.
• Challenges with DETR include handling complex backgrounds, low resolution, and slow convergence.

Purpose of the Study:

• To propose an improved DETR model, GM-DETR, for more effective defect detection.
• To enhance feature extraction, global interaction, and model efficiency.
• To improve accuracy and convergence speed for detecting defects in challenging conditions.

Main Methods:

• Integrated GAM global attention with CNN feature extraction in the DETR model.
• Implemented a layer pruning strategy to optimize the decoding layer and reduce parameters.
• Replaced L1 loss with MSE loss to improve sensitivity to small defect targets and accelerate convergence.

Main Results:

• The GM-DETR model demonstrated improved performance on a road pothole defect dataset.
• Achieved a 4.9% increase in mean average precision (mAP@0.5).
• Reduced the model's parameter count by 12.9%.

Conclusions:

• The GM-DETR model effectively addresses limitations of the original DETR for defect detection.
• The proposed optimizations enhance defect recognition in complex backgrounds and improve model efficiency.
• GM-DETR offers a promising solution for industrial defect detection tasks.