YOLO-DST: MEMS Small-Object Defect Detection Method Based on Dynamic Channel-Spatial Modeling and Multi-Attention...

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Vision

Background:

• Defect detection in Micro-Electro-Mechanical Systems (MEMS) faces challenges with metallographic images, including complex backgrounds, texture interference, and blurred edges.
• These issues hinder the effective identification of critical defects like bond wire breaks and internal cavity contaminants, impacting overall system reliability.

Purpose of the Study:

• To propose an advanced defect detection method for MEMS small objects, addressing limitations in current metallographic image analysis.
• To enhance the accuracy and stability of identifying bond wire breaks and internal cavity contaminants in MEMS devices.

Main Methods:

• Developed YOLO-DST (Dynamic Channel-Spatial Modeling and Triplet Attention-based YOLO), integrating dynamic channel-space blocks into a YOLOv8s framework.
• Incorporated multi-attention fusion in the network neck to suppress background interference and utilized a Transformer-based network for small-object perception enhancement.
• Constructed a dedicated MEMS small-object defect dataset for experimental validation.

Main Results:

• YOLO-DST demonstrated superior performance compared to existing mainstream detection models.
• The method achieved significant improvements in key performance indicators, including Precision and mAP@50% on the constructed dataset.

Conclusions:

• The proposed YOLO-DST method effectively overcomes the challenges of defect detection in MEMS metallographic images.
• The integration of dynamic channel-spatial blocks and multi-attention mechanisms significantly improves the detection of small defects, enhancing MEMS reliability.