Adversarial Model Training for Background-Robust Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for defect detection and feature identification struggle with reduced accuracy due to overrepresentation of certain defect and background combinations in training data, leading to reliance on background information rather than defect detection, and are not robust against differences in backgrounds or features.
Innovation Solution
A model generation apparatus that employs a three-step adversarial learning process involving a coder and two estimators to convert input data into feature amounts, where the first estimator focuses on the primary feature while the second estimator is trained to fail on secondary features, ensuring the model is robust against background variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data includes all combinations of defect type and background type, then estimation accuracy is improved, but data collection cost and time increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-processing training data to extract and separate defect information from background information before model training. This allows the model to learn defect characteristics without being confounded by background variations, achieving high accuracy without needing all possible defect-background combinations in the training set.
Solution Approach 2:
The patent extracts defect information from training images by separating it from background information. The defect extraction unit identifies and extracts defect regions, while the background extraction unit identifies background regions. This separation allows the model to focus on defect characteristics independently of background variations, resolving the contradiction between accuracy and data collection requirements.
2Reliability
If training data includes diverse background types, then model robustness improves, but the model may rely on background information rather than defect information for estimation
Solution Approach 1:
The patent extracts defect information from training images by separating it from background information. The defect extraction unit identifies and extracts defect regions, while the background extraction unit identifies background regions. This separation allows the model to focus on defect characteristics independently of background variations, resolving the contradiction between accuracy and data collection requirements.
Solution Approach 2:
The patent applies preliminary anti-action by deliberately reducing the estimation performance of the initial model on background information through adversarial training. The adversarial training unit trains the initial model to minimize background estimation accuracy, forcing the model to rely on defect information rather than background information for defect detection, thereby preventing background reliance before it becomes a problem.
3Productivity
If machine learning is performed with limited training data, then processing time is reduced, but estimation accuracy deteriorates due to unseen defect-background combinations
Solution Approach 1:
The patent applies preliminary action by pre-processing training data to extract and separate defect information from background information before model training. This allows the model to learn defect characteristics without being confounded by background variations, achieving high accuracy without needing all possible defect-background combinations in the training set.
Solution Approach 2:
The patent introduces an intermediary defect extraction unit that processes training images to extract defect information separately from background information. This intermediary processing step creates purified defect data that can be used for training, allowing the model to achieve high accuracy on unseen defect-background combinations without requiring extensive training data coverage.
4Productivity
If training data overrepresents certain defect-background combinations, then training convergence is faster, but the model fails to generalize to underrepresented combinations
Solution Approach 1:
The patent extracts defect information from training images by separating it from background information. The defect extraction unit identifies and extracts defect regions, while the background extraction unit identifies background regions. This separation allows the model to focus on defect characteristics independently of background variations, resolving the contradiction between accuracy and data collection requirements.
Solution Approach 2:
The patent changes the parameter representation by separating defect and background information into distinct features. Instead of training on raw images where defect and background are mixed, the system extracts defect features and background features separately, then trains the model to estimate defect information from defect features alone. This parameter separation allows the model to generalize across different background types without being biased by overrepresented combinations.
Data Source
AI summary
A model generation apparatus according to one or more embodiments executes operations, with respect to each of learning data sets. The operations includes training a second estimator so that an estimation result obtained from a second estimator conforms to second correct answer data; training a coder so that an estimation result obtained from the second estimator does not conform to the second correct answer data; and training the coder and the first estimator so that an estimation result obtained from a first estimator conforms to first correct answer data. The model generation apparatus executes operation of the training the second estimator and the training the coder alternately and repeatedly.