SEM Image Learning Model for Substrate Noise Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pattern inspection and measurement devices struggle with noise in captured images, leading to reduced accuracy in length measurement and defect inspection of substrates.
Innovation Solution
A learning model is generated using machine learning to receive an SEM image as input and output an image with noise removed, utilizing training data from multiple images captured under various conditions and correcting positional shifts to improve noise removal accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image processing is used for pattern inspection, then the device structure remains simple, but noise in captured images reduces measurement precision and inspection accuracy
Solution Approach 1:
The patent replaces conventional mechanical image processing methods with a machine learning-based system. A learning model trained on paired images (noisy input, clean output) automatically removes noise and enhances image quality, substituting complex manual processing algorithms with an AI-driven approach that achieves superior measurement precision.
Solution Approach 2:
The patent creates a digital copy of the noisy image and uses the learning model to generate a corrected version. The training data consists of paired original and corrected images, allowing the system to learn the transformation from noisy to clean images and apply this knowledge to new inputs, effectively copying the correction process across multiple images.
2Measurement precision
If conventional image processing is used for defect inspection, then the processing method remains simple, but noise reduces inspection accuracy
Solution Approach 1:
The patent replaces traditional defect detection algorithms with a machine learning model that has been trained to distinguish between noise and actual defects. The learning model processes images by learning from training data, enabling more accurate defect inspection while managing algorithmic complexity through automated learning rather than manual rule-setting.
Solution Approach 2:
The system uses training data consisting of paired images (noisy input, clean output with known defect annotations) to train the learning model. This feedback mechanism allows the model to continuously improve its ability to detect defects while filtering noise, with the training process providing feedback on performance metrics like inspection accuracy.
3Measurement precision
If multiple images are captured under various conditions for training, then noise removal accuracy improves, but data collection time and complexity increase
Solution Approach 1:
The patent performs preliminary actions by capturing multiple images under various conditions before the actual inspection work begins. These pre-captured images are used to train the learning model, allowing the system to learn optimal noise removal techniques in advance. This preliminary data collection enables faster and more accurate noise removal during subsequent inspection processes.
Solution Approach 2:
The system varies multiple parameters during image capture, including focal length, aperture, illumination conditions, and substrate orientation. By changing these parameters across different captured images, the training data becomes diverse and representative of various inspection scenarios, improving the learning model's ability to handle different conditions and enhancing overall noise removal accuracy.
Data Source
AI summary
In the method of generating the learning model according to this embodiment, an information processing device acquires a plurality of images of a target substrate captured in chronological order, generates training data in which one image as input among two images selected from the plurality of acquired images and the other image as output are associated with each other, and generates a learning model configured to receive an image obtained by capturing a target substrate as input and output an image from which noise of the image has been removed by machine learning using the training data. It is preferable that the plurality of images includes images under different conditions, and that the two images associated as input and output in the training data are images under the same conditions.