Defect Classification Using Synthetic Templates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional learning-type classification methods for defect classification in semiconductor substrates require a large number of defect images, leading to suboptimal classification at the start of production and over-learning issues with small image datasets, making it difficult to accurately classify defects.
Innovation Solution
A defect classification method that creates defect templates and models by combining teaching images without defects with defect templates, calculating feature amounts, and setting classification classes, allowing for proper classification even with limited or no defect images by using these models and stored relations in a storage unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of defect images are collected for learning-type classification, then defect classification accuracy is improved, but the system cannot properly classify defects at startup when defect images are unavailable
Solution Approach 1:
The system performs preliminary actions by creating defect templates and combining them with teaching images without defects to generate synthetic defect images before actual defect data is available. This preliminary model creation enables the classification system to function at startup without requiring collected defect images, thereby resolving the contradiction between needing accurate classification and having no defect images at startup
2Productivity
If a small number of defect images are used for classification, then the system can operate with limited data, but over-learning occurs and classification accuracy deteriorates
Solution Approach 1:
The system creates copies of defect patterns through templates and combines them with teaching images to generate multiple synthetic defect images. This copying approach allows the system to operate with limited actual defect data while maintaining classification accuracy, as the synthetic images provide additional training samples without requiring collection of numerous real defect images
3Ease of operation
If conventional learning-type classification is used, then defect classification can be performed, but it requires collecting a large number of defect images which is time-consuming and impractical
Solution Approach 1:
The system introduces an intermediary approach by using defect templates as a bridge between teaching images without defects and the final defect classification. Instead of directly requiring numerous real defect images, the templates serve as intermediaries that can be combined with available teaching images to create synthetic defect data, thereby enabling classification functionality without the need to collect large quantities of actual defect images
Data Source
AI summary
A defect classification apparatus of the present invention includes a design unit and a diagnosis unit. In the design unit, a model creation unit combines a defect template in a template storage unit with a teaching image to create a defect model, and a classification class setting unit calculates feature amounts of a defect in the defect model and sets a classification class of the defect. The relation between the feature amounts of the defect and the classification class is stored in a storage unit. In the diagnosis unit, a feature amount calculation unit calculates feature amounts of defects from a captured inspection object image of the substrate, and a classification unit classifies the defects of the substrate into classification classes from the relations between the feature amounts of the defects and the classification classes in the storage unit based on the calculated feature amounts of the defects.