SEM Image Denoising via Virtual Noise Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image denoising methods for scanning electron microscope (SEM) images require a noise-free correct image for training, which is impractical and can damage semiconductor patterns during multiple photographing processes, and struggle with removing white noise that degrades image quality and pattern identification.
Innovation Solution
The method involves extracting a noise patch from non-pattern regions of SEM images, generating imitated virtual noise, and training a denoise deep learning model using pairs of noisy and noisier images to remove noise without a noise-free reference, utilizing a noise parameter estimation network (NPE-net) and an L2 loss function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple photographing processes are performed to obtain a noise-free correct image for training, then the denoising model can be trained effectively, but the semiconductor patterns may be damaged
Solution Approach 1:
The patent creates virtual noise data by copying and transforming existing noisy SEM images through Gaussian noise addition. Instead of requiring multiple physical photographing sessions to obtain noise-free reference images, the method generates synthetic clean images by subtracting estimated noise components from noisy images, thereby avoiding repeated electron beam exposure that could damage sensitive semiconductor patterns.
2Measurement precision
If a noise-free correct image is required for training the denoising model, then the denoising performance can be improved, but the training process becomes impractical and resource-intensive
Solution Approach 1:
The patent implements a self-service training approach where the denoising model learns to remove noise by processing noisy images and comparing outputs against synthetically generated clean versions. The system uses its own noisy image inputs to create training targets through noise estimation and subtraction, eliminating the need for external noise-free reference images that would require separate acquisition processes.
Solution Approach 2:
The method transforms the training data parameters by adding controlled Gaussian noise with varying standard deviations to create a spectrum of noise levels. This parameter transformation allows the model to learn across multiple noise conditions from a single set of noisy images, improving generalization without requiring additional physical imaging sessions for each noise level.
3Measurement precision
If white noise is present in the SEM image, then the image quality degrades and pattern identification becomes difficult, but obtaining a noise-free reference image requires additional photographing that increases processing time
Solution Approach 1:
The patent performs preliminary noise estimation and virtual clean image generation before the actual denoising training begins. By pre-processing the noisy images to create synthetic clean references through noise parameter estimation and Gaussian noise subtraction, the system prepares training data in advance, eliminating the need for time-consuming multiple photographing sessions during the training phase.
Data Source
AI summary
There is provided an image denoising method including extracting a noise patch from a noisy image, outputting a noise parameter by inputting the noise patch to a noise parameter estimation network (NPE-net), generating imitated virtual noise based on the output noise parameter, generating a noisier image by adding the imitated virtual noise to the noisy image, training a denoise deep learning model by inputting the noisy image and the noisier image as a pair to the denoise deep learning model, inputting the noisy image to the trained denoise deep learning model, and outputting a denoise image obtained by removing noise from the noisy image by the trained denoise deep learning model.