GAN-Based Virtual Defect Data for Imbalanced Vision Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of securing a sufficient amount of high-quality defect data for training deep learning models in vision inspection systems is exacerbated by the imbalance between non-defect and defect data, making it difficult to achieve reliable defect detection.
Innovation Solution
An image processing apparatus generates virtual defect data using a small amount of actual defect data through an AI model, specifically a Generative Adversarial Network (GAN), to balance the dataset and enhance training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning models are trained using actual defect data collected during production, then the model learns from real defect patterns, but the limited number of defect occurrences results in insufficient training data and data imbalance
Solution Approach 1:
The patent uses Generative Adversarial Networks (GANs) to create virtual copies of defect data that replicate the characteristics of real defects. The generative model learns from limited actual defect data and synthesizes additional defect images, effectively multiplying the available training data while maintaining realistic defect patterns and properties
Solution Approach 2:
The patent introduces an intermediary system consisting of the generative model and discriminative model that mediates between the limited actual defect data and the need for abundant training data. This intermediary GAN framework transforms scarce real data into abundant synthetic data that bridges the gap between available samples and training requirements
2Adaptability or versatility
If more defect data is collected to balance the dataset, then training data balance improves, but the difficulty of securing defect data increases due to limited defect occurrences in production
Solution Approach 1:
Instead of collecting more physical defect data from production lines, the system creates virtual copies through the generative model. This copying approach bypasses the collection difficulty entirely by synthesizing defect data computationally, maintaining dataset balance without requiring additional physical samples
Solution Approach 2:
The patent replaces the mechanical/physical process of collecting defect data from production lines with a computational process. The GAN-based generative system substitutes physical data collection with algorithmic data synthesis, eliminating the constraint of limited defect occurrences during manufacturing
3Productivity
If virtual defect data is generated using GAN models, then data imbalance is resolved and training efficiency improves, but the complexity of the data processing system increases
Solution Approach 1:
The patent performs preliminary data preparation by generating balanced training datasets before the actual model training begins. The GAN system pre-processes and balances the training data in advance, creating a ready-to-use balanced dataset that streamlines subsequent training operations and improves overall efficiency
Solution Approach 2:
The patent transforms the training data through parameter changes in the generative model, converting limited actual defect data into expanded virtual datasets. By adjusting parameters such as defect types, locations, and characteristics in the generative model, the system efficiently creates diverse training samples without manual intervention
Data Source
AI summary
A product inspection apparatus including a camera configured to capture an image of a product to be inspected; and a processor configured to extract defect information from a defect indicated by the captured image of the product, generate first virtual defect data including at least one of a location, a size and a shape of the defect included in the captured image, based on the extracted defect information, generate second virtual defect data by synthesizing the first virtual defect data with non-defect data representing the product without the defect, and generate final virtual defect data by inputting the second virtual defect data to an artificial intelligence generative model.