Semi-supervised semantic segmentation of SEM images considering multi-scale structural consistency loss in...

Area of Science:

• Semiconductor device fabrication
• Computer vision
• Machine learning

Background:

• Scanning Electron Microscopes (SEM) are crucial for inspecting semiconductor circuit patterns.
• Increasing complexity and decreasing scale of patterns challenge traditional rule-based image processing.
• Existing semi-supervised methods struggle with accurate pseudo-labeling and recognition consistency.

Purpose of the Study:

• To develop an advanced semi-supervised learning method for accurate pixel-level pattern segmentation in semiconductor inspection.
• To address limitations in current methods regarding pseudo-label accuracy and recognition of complex layouts.
• To improve the yield of semiconductor device fabrication through enhanced pattern recognition.

Main Methods:

• Proposed a novel semi-supervised learning approach for image segmentation.
• Introduced new loss functions to evaluate pattern structure consistency across multiple scales.
• Focused on improving the accuracy of pseudo-labels and handling large unmarked areas.

Main Results:

• Achieved a dramatic increase in accuracy relative to visual identification, from 10-12% to 100%.
• Significantly improved mean Intersection-over-Union (mIoU) at the pixel level, from 0.45-0.65 to over 0.94.
• Demonstrated superior performance compared to established methods like Unimatch and CAC.

Conclusions:

• The proposed semi-supervised learning method offers highly accurate segmentation for semiconductor patterns.
• The novel loss functions are key to achieving consistency and precision in complex pattern recognition.
• This advancement holds significant potential for improving yield and reliability in semiconductor manufacturing.