A Novel Defect Inspection System Using Convolutional Neural Network for MEMS Pressure Sensors

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence
• Computer Vision

Background:

• Quality control of microelectromechanical systems (MEMS) sensors is crucial in the semiconductor industry.
• Current imaging-processing techniques face challenges in high-precision defect classification and localization due to small defect image sizes and varying scales.
• Accurate defect detection is vital for ensuring the reliability of MEMS sensors.

Purpose of the Study:

• To propose a simple, flexible, and efficient convolutional neural network (CNN) for inspecting MEMS pressure-sensor-chip packaging.
• To enhance the performance of existing region-based CNN models for defect detection.
• To address the challenges of small image sizes and scale variations in defect identification.

Main Methods:

• Development of an Accurate-Detection CNN (ADCNN) based on the Faster R-CNN architecture.
• Integration of random-data augmentation techniques to improve network robustness.
• Inclusion of specific defect classifiers to enhance classification accuracy.

Main Results:

• The proposed ADCNN achieved a mean average precision (mAP) of 92.39% for defect detection.
• The integrated defect classifier demonstrated a mean accuracy of 97.2% for classifying identified defects.
• The ADCNN model proved effective in handling variations in defect scale and image size.

Conclusions:

• The ADCNN offers a significant advancement in automated defect inspection for MEMS sensor packaging.
• The method provides a flexible and efficient solution for improving quality control in semiconductor manufacturing.
• The enhanced CNN architecture successfully addresses key challenges in high-precision defect classification and localization.