Synthetic Data Generation for Visual Inspection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current visual inspection methods using machine learning models face challenges in accurately detecting defects due to the high cost of collecting diverse samples for training, particularly when defect and background types are imbalanced, leading to reduced accuracy and increased costs.
Innovation Solution
A data generation system that uses machine learning to generate pseudo samples for underrepresented combinations of defect and background types, allowing for the automatic creation of diverse training data without the need for extensive real-world sample collection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If diverse samples for all combinations of defect types and background types are collected for training, then the accuracy of visual inspection is improved, but the cost and time for sample collection increase significantly
Solution Approach 1:
The patent uses image synthesis technology to generate synthetic images that replicate real defect samples. The synthesis apparatus creates artificial images by combining background images with defect images, producing training data that mirrors real-world combinations without requiring actual physical samples for every possible defect-type and background-type combination. This copying approach enables comprehensive training data generation at reduced cost and time.
Solution Approach 2:
The patent introduces a synthesis apparatus as an intermediary between the training data requirement and the final inspection system. This intermediary generates pseudo-samples that serve as training data, mediating between the need for diverse comprehensive samples and the practical limitations of collection costs and time. The synthesized images act as stand-ins for rare or difficult-to-obtain real defect samples.
2Ease of manufacture
If machine learning models are trained using imbalanced samples, then the training process is simplified, but the model accuracy deteriorates due to inability to detect rare defect combinations
Solution Approach 1:
The synthesis apparatus creates synthetic training samples by combining background images with defect images, generating a balanced and diverse training dataset. This copying mechanism ensures that rare defect combinations are represented in the training data without requiring actual collection of all possible combinations, thereby maintaining training simplicity while improving detection accuracy for rare cases.
Solution Approach 2:
The patent controls parameters such as the number of synthesized samples, defect types, background types, and their combinations to create a balanced training dataset. By adjusting these parameters, the system can generate appropriate sample distributions that prevent imbalance issues while maintaining training efficiency, ensuring the model learns to detect all defect types including rare combinations.
3Measurement precision
If detection parameters are set for all possible defect and background combinations, then detection accuracy for all combinations is improved, but the complexity of setting and maintaining parameters increases
Solution Approach 1:
Instead of manually setting detection parameters for each combination, the patent uses synthesized training images to automatically train machine learning models. The synthesis apparatus generates training data that covers all combinations, and the model learns detection patterns from these synthetic examples, eliminating the need for manual parameter configuration for each scenario while maintaining comprehensive detection accuracy.
Solution Approach 2:
The patent replaces manual parameter setting mechanisms with automated machine learning training. Rather than requiring experts to configure detection parameters for every defect-type and background-type combination, the system uses synthesized training data to automatically train neural networks that perform detection, substituting mechanical parameter configuration with automated learning processes.
Data Source
Figure 1~2A
Figure 2B~3
Figure 4~5
AI summary
A technique reduces the costs for collecting various samples of data. A data generation system generates a first pseudo sample including a first feature of a type corresponding to an input value using a trained first generator, a second pseudo sample including a second feature of a type corresponding to an input value using a trained second generator, and a new sample including the first feature and the second feature by synthesizing the generated first pseudo sample and the generated second pseudo sample.