An improved EAE-DETR model for defect detection of server motherboard

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Server motherboard defect identification faces challenges with small targets, rotation deviations, and scale variations, leading to missed and false detections.
• Existing detection models struggle with the complexity and variability of defects in industrial environments.

Purpose of the Study:

• To develop an enhanced detection model, EAE-DETR, for high-precision and high-efficiency server motherboard defect identification.
• To address limitations in detecting small, rotated, and unevenly scaled defects.

Main Methods:

• Proposed the CSP-EfficientVIM-CGLU module for enhanced feature extraction and parameter reduction.
• Introduced the AIFI-ASSA module to mitigate background noise and improve sensitivity to minor defects.
• Developed the EUCB-SC upsampling module for efficient feature reconstruction.

Main Results:

• EAE-DETR achieved a mean Average Precision (mAP) of 78.5% (IoU=0.5) and 32.6% (IoU 0.5-0.95) on the PCBA-DET dataset, outperforming the baseline.
• Demonstrated a 21.7% reduction in parameter count and a 12.0% decrease in computational load.
• Achieved mAP50 of 96.1% and mAP50:95 of 65.1% on the PKU-Market-PCB dataset.

Conclusions:

• EAE-DETR offers a robust solution for high-precision and high-efficiency defect detection in server motherboards.
• The model effectively handles challenges posed by complex industrial environments.
• This advancement supports the intelligent manufacturing sector by improving quality control processes.