Synthetic Image Generation for DNN Training Data Expansion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) require significant and costly large-scale training data, making them underutilized due to the expense and time-consuming nature of data collection.
Innovation Solution
A media platform that includes an image extractor to extract latent features and a generative model to generate synthetic images, which are used to train a DNN image generator, utilizing a guided imagination framework to optimize latent features and maintain class consistency, entropy, and diversity, thereby expanding a small dataset into a larger informative one.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large-scale training data is collected to improve DNN performance, then model accuracy is improved, but cost and time consumption increase significantly
Solution Approach 1:
The patent uses a generative model to create synthetic copies of training images based on latent features extracted from a small set of real images. These synthetic images serve as artificial training data, replacing the need to collect large amounts of real-world data while maintaining the statistical properties needed for effective DNN training.
Solution Approach 2:
The system performs preliminary extraction of latent features from real images before generating synthetic samples. By pre-processing the real data to extract essential features and then generating multiple synthetic variations, the system prepares training data in advance without requiring extensive data collection efforts.
2Reliability
If large-scale training data is collected to improve DNN performance, then model accuracy is improved, but expense increases significantly
Solution Approach 1:
The generative model creates synthetic training data copies from a small set of real images, eliminating the need for expensive data collection campaigns. The synthetic images preserve the essential characteristics needed for training while avoiding the costs associated with acquiring, annotating, and managing large-scale real-world datasets.
Solution Approach 2:
The system uses the extracted latent features from real images to automatically generate its own training data without external data sources. The generative model serves itself by creating the training dataset internally, reducing dependency on external data collection resources and associated costs.
3Loss of time
If synthetic images are generated to reduce data collection needs, then data collection time is reduced, but maintaining class consistency and diversity becomes challenging
Solution Approach 1:
The system incorporates feedback mechanisms where the generative model's output is evaluated against class labels and diversity metrics. The latent features are optimized to ensure that generated images maintain correct class classification while introducing sufficient variation for model training effectiveness.
Solution Approach 2:
The system manipulates parameters in the latent feature space to control the generation process. By adjusting these parameters, the system can generate images that maintain class consistency while introducing diversity in appearance, ensuring both fidelity to the original class and sufficient variation for robust training.
Data Source
AI summary
Automatic data generation includes extracting latent features from an input image, adding a perturbation to the latent features, applying the perturbed latent features to a pre-trained generative model, and training an image generator with images output from the generative model.