Synthetic Defect Image Generation for Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating learning data for defect identification in semiconductor manufacturing are time-consuming and costly, especially when dealing with the scarcity of defect data compared to normal cases.
Innovation Solution
A learning data generation device that cuts out parts of normal image data to create pseudo-defect images, combines these with two-dimensional graphics representing defects, and assigns labels to the resulting images, thereby augmenting the learning data set.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of defect learning data is increased to improve defect identification accuracy, then the accuracy of defect classification is improved, but the time and cost required for data preparation increases enormously
Solution Approach 1:
The patent creates synthetic defect images by copying and combining normal image regions with defect patterns. Instead of manually collecting real defect images, the system generates artificial defect data by overlaying defect patterns onto normal images, thereby creating a large dataset without the time-consuming process of manual data collection and annotation.
Solution Approach 2:
The patent performs preliminary preparation of normal image data by pre-processing and storing it in a database. This preliminary action allows the system to quickly generate defect images on-demand by combining pre-processed normal images with defect patterns, eliminating the need for time-consuming real-time processing when creating the learning dataset.
2Measurement precision
If the number of defect learning data is increased to improve defect identification accuracy, then the accuracy of defect classification is improved, but the cost of data preparation increases enormously
Solution Approach 1:
The patent creates synthetic defect images by copying and combining normal image regions with defect patterns. Instead of manually collecting real defect images, the system generates artificial defect data by overlaying defect patterns onto normal images, thereby creating a large dataset without the time-consuming process of manual data collection and annotation.
Solution Approach 2:
The system automatically generates defect images by combining normal images with defect patterns using computer algorithms. This self-service approach eliminates the need for manual data annotation and preparation by human experts, significantly reducing labor costs and making the process scalable without proportional increases in expense.
3Quantity of substance
If traditional data augmentation methods (shift, inversion, rotation, zooming) are applied to increase learning data, then the amount of learning data is increased, but the accuracy of defect identification is insufficiently improved
Solution Approach 1:
The patent performs preliminary preparation of normal image data by pre-processing and storing it in a database. This preliminary action allows the system to quickly generate defect images on-demand by combining pre-processed normal images with defect patterns, eliminating the need for time-consuming real-time processing when creating the learning dataset.
Solution Approach 2:
The patent creates synthetic defect images by copying and combining normal image regions with defect patterns. Instead of manually collecting real defect images, the system generates artificial defect data by overlaying defect patterns onto normal images, thereby creating a large dataset without the time-consuming process of manual data collection and annotation.
4Measurement precision
If learning data is prepared for each defect type to improve identification accuracy, then the accuracy of defect classification is improved, but the complexity and time required for data preparation increases
Solution Approach 1:
The patent employs a universal defect pattern library that can be applied to multiple defect types. Instead of creating separate datasets for each defect category, the system uses a single set of defect patterns that can be overlaid onto normal images to generate various defect scenarios, thereby simplifying the data preparation process while maintaining the ability to classify different defect types.
Solution Approach 2:
The patent performs preliminary preparation of normal image data by pre-processing and storing it in a database. This preliminary action allows the system to quickly generate defect images on-demand by combining pre-processed normal images with defect patterns, eliminating the need for time-consuming real-time processing when creating the learning dataset.
Data Source
AI summary
A learning data generation device that can generate learning data suitable for learning of an identification model. The learning data generation device has a function of cutting out part of first image data as second image data, a function of generating a two-dimensional graphic corresponding to the area of the second image data and representing a pseudo defect, a function of generating third image data by combining the second image data and the two-dimensional graphic, and a function of assigning a label corresponding to the two-dimensional graphic to the third image data. By using the third image data for learning of the identification model, a highly accurate identification model can be generated.