Class-Identity-Preserving Data Augmentation via GAN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks face overfitting issues due to limited training datasets, and existing data augmentation methods do not adequately address privacy concerns and insufficiency of available training images, especially for in-the-wild recognition systems.
Innovation Solution
A system and method for data augmentation that preserves class identity by generating photorealistic synthetic datasets using a generative adversarial network, conditioning on facial identities and varying semantic features like pose, lighting, and expression, to enhance model performance and training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data augmentation is used to increase training dataset size, then overfitting is reduced, but class identity may be lost in augmented samples
Solution Approach 1:
The patent applies parameter changes by selectively modifying certain attributes (pose, lighting, expression, makeup) while preserving the core class identity parameter. This allows data augmentation to increase dataset diversity and reduce overfitting while maintaining the fundamental characteristics that define each class, thus resolving the contradiction between improving generalization and preserving class identity.
Solution Approach 2:
The patent implements local quality by applying different levels of modification to different parts of the data. Critical identity-defining features are preserved with high fidelity, while non-critical features (such as lighting conditions, facial expressions, or pose) are varied to create diversity. This localized approach to quality control enables both overfitting reduction and identity preservation.
2Quantity of substance
If synthetic training data is generated to address data insufficiency, then model training is improved, but privacy concerns arise
Solution Approach 1:
The patent uses copying by generating synthetic training samples that replicate the statistical properties and visual characteristics of real training data without copying actual sensitive information. The generative model creates new, artificial examples that preserve the essential patterns needed for training while eliminating direct links to private individual data, thus increasing data quantity while mitigating privacy risks.
3Quantity of substance
If existing data augmentation methods are used, then training dataset size increases, but photorealism and class identity are not preserved
Solution Approach 1:
The patent introduces an intermediary element - a generative adversarial network - that acts as a mediator between the need for data quantity and the requirement for quality preservation. This intermediary learns the underlying data distribution and generates new samples that maintain photorealism and class identity, unlike traditional augmentation methods that directly transform existing data. The GAN intermediary ensures both scale and quality improvements.
Data Source
AI summary
Disclosed herein is a system and method for data augmentation for general object recognition which preserves the class identity of the augmented data. The system comprises an image recognition network an image generation network that take as input ground truth images and classes respectively and which generates a predicted class and an augmented image. A discriminator evaluates the predicted class and augmented image and provides feedback to the image recognition network and the image generation network.