High-Speed Die Bond Quality Detection Using Lightweight Architecture DSGβSI-SECS-Yolov7-Tiny

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence in Quality Control
• Computer Vision for Industrial Inspection

Background:

• Die bonding is crucial for IC packaging quality and yield.
• Increased automation speeds necessitate faster, more accurate visual inspection.
• Existing inspection methods struggle with high-speed production lines, leading to misclassifications.

Purpose of the Study:

• To develop a high-speed intelligent vision inspection model for die bonding.
• To improve classification accuracy and adapt to next-generation automated machinery.
• To enable real-time process parameter adjustments for enhanced yield and reduced costs.

Main Methods:

• Developed DSGβSI-SECS-YOLOv7-tiny, an enhanced lightweight model.
• Integrated depthwise separable convolution, Ghost convolution, and a learnable Sigmoid activation.
• Incorporated SE layer, ECA-Net, Coordinate Attention, and Small Object Enhancer for improved performance.

Main Results:

• Achieved an inference speed of 294.1 FPS.
• Reached a precision rate of 99.1%.
• Demonstrated superior performance compared to the previous DSGβSI-YOLOv7-tiny model.

Conclusions:

• The DSGβSI-SECS-YOLOv7-tiny model offers a significant advancement in high-speed die bonding quality inspection.
• The model's efficiency and accuracy support real-time defect detection and process optimization.
• This technology promises to improve IC packaging yield and reduce manufacturing losses.