Defect Detection of MEMS Based on Data Augmentation, WGAN-DIV-DC, and a YOLOv5 Model.

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Surface defect detection is critical for micro-electromechanical system (MEMS) acoustic thin film quality control.
• Deep learning model performance relies heavily on training data, which is often insufficient for defect detection in production settings.

Purpose of the Study:

• To develop an improved YOLOv5 model for real-time MEMS defect detection.
• To address the challenge of limited defect samples in training datasets.

Main Methods:

• An enhanced YOLOv5 model incorporating Mosaic augmentation and an additional prediction head was utilized.
• Wasserstein divergence for generative adversarial networks with deep convolutional structure (WGAN-DIV-DC) was employed for defect sample expansion and diversification.

Main Results:

• The optimal detection model achieved a mean Average Precision (mAP) of 0.901 and an F1 score of 0.856.
• The model demonstrated real-time detection speed at 75.1 Frames Per Second (FPS).
• Compared to the baseline model, mAP and F1 score increased by 8.16% and 6.73%, respectively, using the augmented dataset.

Conclusions:

• The proposed defect detection model significantly improves accuracy and efficiency for MEMS acoustic thin film inspection.
• The integration of data augmentation techniques enhances the robustness and performance of deep learning models in defect detection.
• This approach offers substantial convenience for MEMS production quality control.