Semiconductor Defect Image Augmentation for Sparse Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in semiconductor manufacturing is the scarcity and difficulty in obtaining sufficient, high-quality training data for machine learning models due to the rarity and subtlety of defects, leading to inadequate defect detection sensitivity and misclassification.
Innovation Solution
A computerized system and method for generating synthetic defective images by identifying and pasting annotated regions from original images onto candidate areas in target images, enriching the training set for machine learning models to improve defect detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained using actual defect images only, then the model can learn from real defect patterns, but the scarcity and rarity of defects lead to insufficient training data and inadequate defect detection sensitivity
Solution Approach 1:
The patent creates synthetic defect images by copying and pasting annotated defect regions from original images onto target images. This generates additional training data without requiring more physical defect samples, directly addressing the scarcity of training data while maintaining realistic defect patterns for improved detection sensitivity
Solution Approach 2:
The system performs preliminary annotation of defect regions in original images, storing these annotated regions as reusable components. This preliminary action enables subsequent synthesis of multiple training samples from a limited set of actual defects, effectively multiplying the training data available
2Measurement precision
If machine learning models are trained with insufficient or low-quality training data, then the training process is faster and requires fewer resources, but the model produces misclassification and reduced defect detection accuracy
Solution Approach 1:
By copying annotated defect regions with their contextual information onto multiple target images, the system generates high-quality synthetic training samples that preserve real defect characteristics. This approach improves detection accuracy without requiring complex data collection processes
Solution Approach 2:
The system uses the original defect images themselves as the source material for creating synthetic samples, eliminating the need for external defect injection or complex defect generation apparatus. The annotated regions from the original images serve as self-contained training components
3Quantity of substance
If more actual defect samples are collected to improve training data quantity, then the model can be trained better, but defects are rare and collecting sufficient samples is time-consuming and resource-intensive
Solution Approach 1:
Instead of collecting more physical defect samples over time, the system copies existing annotated defect regions to generate multiple synthetic samples instantly. This digital replication process eliminates the time-consuming physical collection process while providing sufficient training data quantity
Solution Approach 2:
The system performs preliminary annotation and extraction of defect regions from available images, creating a reusable library of defect components. This preliminary action transforms the data preparation process, allowing rapid generation of training samples without repeated physical collection efforts
Data Source
AI summary
There is provided a system and method for defect examination on a semiconductor specimen. The method comprises obtaining an original image of the semiconductor specimen, the original image having a first region annotated as enclosing a defective feature; specifying a second region in the original image containing the first region, giving rise to a contextual region between the first region and the second region; identifying in a target image of the specimen a set of candidate areas matching the contextual region in accordance with a matching measure; selecting one or more candidate areas from the set of candidate areas; and pasting the first region or part thereof with respect to the one or more candidate areas, giving rise to an augmented target image usable for defect examination on the semiconductor specimen.