A feature-enhanced transformer for point-like micro-defect detection on semiconductor wafer surfaces.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Semiconductor wafer defects impact chip yield and reliability.
• Existing visual inspection methods struggle with minute, low-contrast, noisy, and dense defects.

Purpose of the Study:

• To develop an advanced deep learning model for robust wafer defect detection.
• To improve the stability and accuracy of automated visual inspection in semiconductor manufacturing.

Main Methods:

• Proposed Wafer-DETR, a Transformer-based detector with a frequency-spatial collaborative enhancement backbone (FresNet).
• Integrated Fast Fourier Transform (FFT) spectral modeling and edge enhancement for detail reinforcement.
• Introduced a cross-scale adaptive fusion module (CSAF) for noise resilience and a content-aware upsampling operator (CARAFE) for boundary enhancement.

Main Results:

• Wafer-DETR achieved 79.7% average precision (AP) at IoU 0.50 and 25.2% AP averaged over IoU 0.50-0.95.
• Demonstrated absolute gains of 3.3% and 1.9% over RT-DETR, respectively.
• Showed more stable detection performance in challenging low-contrast, noisy, and densely defective scenarios.

Conclusions:

• Wafer-DETR offers effective technical support for automated visual quality inspection in wafer manufacturing.
• The proposed method significantly enhances the detection of critical wafer surface defects.