Image-Specific Back-Propagation Training for Defect Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing inspection devices in substrate processing systems face challenges in accurately distinguishing between defect images and false report images, leading to increased inspector burden due to degraded inference accuracy from over-training.
Innovation Solution
A training device and method that utilize different error back-propagation rates for defect and false report images during training, generating learning models with improved inference accuracy for defect images while preventing over-training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of trainings in the image evaluation device is increased to reduce false reports, then the inference accuracy for defect detection is improved, but over-training occurs and degrades the overall inference accuracy
Solution Approach 1:
The patent applies local quality by using different error back-propagation rates for different image types during training. Specifically, a first error back-propagation rate is used for defect images while a second error back-propagation rate is used for false report images, allowing the model to learn different levels of detail for different classes. This resolves the contradiction by enabling focused learning on defect detection without uniformly over-training on all image types.
Solution Approach 2:
The patent changes the training parameter (error back-propagation rate) based on the type of image being processed. By dynamically adjusting the learning rate parameter according to whether the image is a defect image or false report image, the system optimizes training efficiency and prevents over-training while maintaining high inference accuracy for defect detection.
2Adaptability or versatility
If the training model is trained equally on all image types, then generalization is maintained, but inference accuracy for specific defect types does not improve sufficiently
Solution Approach 1:
The patent implements local quality by applying different training intensities (error back-propagation rates) to different image types. Defect images receive a first error back-propagation rate that enables more precise learning of defect characteristics, while false report images receive a second error back-propagation rate that maintains generalization. This differential approach allows the model to achieve high accuracy for specific defect types while maintaining overall adaptability.
Data Source
AI summary
A training device includes a data acquirer that acquires training data including a first image of a first type and a second image of a second type that is different from the first type, a first trainer that repeats an epoch for causing the first image and the second image to be learned under a same condition multiple times, and a second trainer that repeats an epoch for training a trained learning model generated by the first trainer with a rate of error back-propagation in regard to the first image and a rate of error back-propagation in regard to the second image being made different from each other.