High-speed e-beam metrology using a pattern recognition-based image enhancement method

In semiconductor manufacturing, Scanning Electron Microscopes (SEMs) are employed for critical dimension (CD) measurements, overlay assessments, and defect inspections to improve yield. However, SEM images frequently suffer from issues such as poor contrast and low signal-to-noise ratio (SNR), which lead to inaccurate metrology and incorrect defect inspections. As transistors and interconnects continue to shrink and become more complex, poor contrast and low SNR have become even more critical issues. Consequently, Semiconductor manufacturers and SEM device vendors are continuously seeking effective methods to enhance SNR and contrast. Although numerous image noise reduction methods have been proposed and show good performance in improving SNR, they cannot enhance image contrast simultaneously. As for the reported image contrast enhancement methods, we argue that they have two key drawbacks: (1) they typically apply image process functions uniformly across the entire image, without considering local variations, which fails to achieve a balanced appearance for each local region; and (2) these methods are generally used as pre-processing steps for specific tasks, which reduces the overall throughput of metrology or inspection systems. To address these challenges, we propose a novel metrology solution that integrates a pattern-recognition-based image contrast enhancement method with a convolutional neural network (CNN) noise reduction module. Experimental results demonstrate that the proposed solution achieves optimal visual quality and improves metrology throughput by more than three-fold.