EM-YOLO: high-precision electronic component detection via multi-scale attention and dynamic feature fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Electrical Engineering

Background:

• Automated optical inspection (AOI) faces challenges with missed detections and low accuracy in electronic component circuit board inspection.
• Existing algorithms struggle with detecting small or densely packed components.

Purpose of the Study:

• To develop an enhanced target detection algorithm, EM-YOLO, based on YOLOv11 for improved electronic component inspection.
• To address limitations in feature representation, multi-scale fusion, and localization accuracy.

Main Methods:

• Designed C3k2_EMA backbone with Efficient Multi-scale Attention (EMA) for enhanced tiny component feature extraction.
• Proposed BiSPD-FPN neck structure integrating Bidirectional Feature Pyramid Network (BiFPN) and Spatial Pyramid Depthwise Convolution (SPDConv) for optimized feature fusion.
• Introduced Focal-DIoU loss function in the detection head to improve bounding box regression and localization accuracy.

Main Results:

• EM-YOLO achieved a 0.5% and 3.9% increase in mean Average Precision (mAP@0.5) on self-made and public PCB datasets, respectively.
• Demonstrated a 1.3% and 4% reduction in false negative rate (FNR) compared to benchmark algorithms.
• Outperformed mainstream algorithms in detection accuracy and significantly alleviated missed detections.

Conclusions:

• The EM-YOLO algorithm offers superior detection accuracy and reduced false negatives for electronic components.
• The proposed optimizations effectively enhance feature representation, multi-scale fusion, and localization.
• EM-YOLO provides a robust and generalizable solution for automated circuit board inspection.