GAN-Based Synthetic Defect Data Generation for Pattern Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pattern recognition technologies face challenges in obtaining diverse and large amounts of high-quality training data, particularly for defect data in various patterns, which are often specific to each pattern and require separate operations.
Innovation Solution
A method is introduced to generate defect data for a target domain using defect data from a source domain by applying and reconstructing masks with a reconstruction algorithm, allowing for the training of the algorithm to produce defect and normal data for both domains, enabling efficient generation of training data across different patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate operations are performed to obtain defect data for each different pattern, then the accuracy of defect recognition is improved, but the time and complexity of data collection increases significantly
Solution Approach 1:
The patent creates virtual defect data by copying and transforming defect patterns from source domains to target domains. The GAN model generates synthetic defect images that replicate real defect characteristics without requiring physical collection from each pattern, thus maintaining recognition accuracy while eliminating time-consuming separate data collection operations.
Solution Approach 2:
The patent transforms defect data across different patterns by changing domain parameters through the GAN model. By learning the mapping between source and target domain parameters, the system generates defect data for various patterns from a single source, resolving the contradiction between pattern-specific accuracy and data collection efficiency.
2Reliability
If a large amount of diverse training data is collected for each pattern, then the quality of pattern recognition learning is improved, but the complexity and cost of data acquisition increases
Solution Approach 1:
The patent creates a universal data generation system that can produce training data for multiple patterns simultaneously. The GAN model learns from source domain data and generates defect data across various target patterns, making the data acquisition system multi-functional and eliminating the need for separate collection processes for each pattern.
Solution Approach 2:
The system enables self-service data generation where the GAN model automatically creates diverse training data without external intervention for each pattern. The model learns defect characteristics from source data and autonomously generates sufficient training examples for target patterns, reducing acquisition complexity while maintaining learning quality.
3Reliability
If defect data is obtained through separate operations for each pattern, then the authenticity of defect data is improved, but the productivity of training data generation decreases
Solution Approach 1:
The patent uses copying to replicate authentic defect characteristics from source domain data to target domain data through the GAN model. By copying the essential features and patterns of real defects while transforming them to different patterns, the system maintains data authenticity while achieving high-speed parallel generation across multiple patterns.
Solution Approach 2:
The patent performs preliminary learning of defect characteristics from source domain data before generating target domain data. The GAN model pre-learns defect patterns, textures, and features, enabling it to efficiently generate authentic-looking defect data for multiple patterns without requiring separate collection operations, thus improving productivity while maintaining authenticity.
Data Source
AI summary
A method includes inputting defect data of a source domain, to which a first mask is applied/unapplied to a reconstruction algorithm. The algorithm is trained to generate defect data of the source domain, to which the first mask is reconstructed. Normal data of the source domain is input to the algorithm, and includes data to which a second mask is applied, and data to which the second mask is not applied. The algorithm is trained to generate normal data of the source domain, to which the second mask is reconstructed. Normal data of a target domain is input to the algorithm, and the normal data of the target domain includes data to which the second mask is applied, and data to which the second mask is not applied. The algorithm is trained to generate normal data of the target domain, to which the second mask is reconstructed.