Defect Classification Using Synthetic Image Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In semiconductor wafer manufacturing, existing defect classification methods require extensive image data and time to achieve desired accuracy due to the infrequent occurrence of defects, limiting the rapid improvement of classification performance.
Innovation Solution
A defect classification apparatus and method that processes captured images using an image storage unit, defect class storage unit, image processing unit, and classifier learning unit to generate and learn from multiple defect images, even with a small amount of data, thereby increasing image data variation and improving classification performance quickly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image data is captured from actual defects in the manufacturing line, then the classification performance can be improved with real defect characteristics, but it takes a long time to accumulate sufficient image data because defects occur infrequently
Solution Approach 1:
The system performs preliminary actions by generating synthetic defect images through image processing transformations (rotation, scaling, flipping, cropping) before actual classification training. This allows the classifier to be pre-trained with varied defect images without waiting for actual defects to occur in the manufacturing line, thus resolving the contradiction between achieving high classification performance and the time required to accumulate sufficient real defect image data.
2Adaptability or versatility
If multiple types of image-capturing apparatuses are used to capture defect images, then the variation of image data increases, but the device complexity and cost increase
Solution Approach 1:
Instead of using multiple physical image-capturing apparatuses to generate varied defect images, the system creates copies and transformations of existing defect images through digital image processing operations. This includes rotating, scaling, flipping, and cropping images to generate multiple variations from a single captured image, thereby achieving high adaptability and image data variation without increasing device complexity or cost.
3Ease of operation
If design information is deformed and compared with images to be classified, then classification can be performed, but the method does not effectively learn the classifier with varied defect characteristics
Solution Approach 1:
The system performs preliminary transformation of defect images into various forms (rotated, scaled, flipped, cropped versions) before training the classifier. This allows the classifier to learn from diverse defect representations in advance, improving its ability to accurately classify new defect images while maintaining ease of operation through automated processing pipelines.
Data Source
AI summary
Provided is a defect classification apparatus classifying images of defects of a sample included in images obtained by capturing the sample, the apparatus including an image storage unit for storing the images of the sample acquired by an external image acquisition unit, a defect class storage unit for storing types of defects included in the images of the sample, an image processing unit for extracting images of defects from the images from the sample, processing the extracted images of defects and generating a plurality of defect images, a classifier learning unit for learning a defect classifier using the images of defects of the sample extracted by the image processing unit and data of the plurality of generated defect images, and a defect classification unit for processing the images of the sample by using the classifier learned by the classifier learning unit, to classify the images of defects of the sample.