Spectrum Mixup Synthetic Data for Face Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition systems face challenges in achieving high accuracy due to domain gaps, privacy concerns, long-tailed distributions, image quality inconsistencies, noisy labels, and lack of attribute annotations, particularly when using web-crawled datasets.
Innovation Solution
A method and system utilizing Spatial Augmentation and Spectrum Mixup (SASMU) techniques, combining spatial and frequency domain manipulations, to create an augmented synthetic dataset that bridges the domain gap between synthetic and real datasets without using real face images, employing Fourier Transform, Gaussian filters, and soft-assignment maps to enhance recognition model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If web-crawled datasets are used for training face recognition models, then model accuracy is improved, but privacy concerns worsen
Solution Approach 1:
The patent creates synthetic face images that copy the essential characteristics and statistical properties of real face images without using actual personal data. By synthesizing training data that replicates the distribution and features of web-crawled datasets, the system achieves high recognition accuracy while completely avoiding privacy invasion associated with real face images.
2Object-affected harmful factors
If synthetic data is used for training face recognition models, then privacy concerns are resolved, but domain gap challenges worsen
Solution Approach 1:
The patent employs parameter changes by adjusting statistical parameters such as mean, variance, and correlation structures to match those of real face datasets. By carefully tuning these parameters during synthetic data generation, the system reduces the domain gap between synthetic training data and real-world test data, thereby improving model reliability and generalization performance.
Solution Approach 2:
The system implements feedback mechanisms by evaluating the domain gap between synthetic and real data distributions, then using this information to iteratively improve the synthetic data generation process. This feedback loop ensures that the synthetic data progressively better matches real-world characteristics, reducing the domain gap and enhancing model performance.
3Measurement precision
If real images are used to bridge the domain gap, then model accuracy is improved, but privacy preservation efforts are compromised
Solution Approach 1:
Instead of using real images to bridge the domain gap, the patent creates enhanced synthetic images that copy and amplify the beneficial statistical properties of real data. The synthetic data generation process incorporates techniques to match the distribution, texture, and structural characteristics of real faces, achieving the same accuracy-improving effect without compromising privacy preservation.
Data Source
AI summary
The present invention provides a method for enhancing recognition model accuracy across source and target domains and a system thereof. The method includes the steps of: extracting amplitude and phase components from a source domain dataset and a target domain dataset, respectively; separating high-frequency components from the amplitude components of the target domain dataset and low-frequency components from the amplitude components of the source domain dataset; creating an augmented amplitude in a frequency domain by incorporating the high-frequency components separated from the target domain dataset into the low-frequency components separated from the source domain dataset; generating an augmented synthetic dataset based on the augmented amplitude and the phase components of the source domain dataset; and training the recognition model with the augmented synthetic dataset.