Manifold-Based Data Augmentation for Imbalanced 3D Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data augmentation methods struggle to dynamically simulate the deformation of three-dimensional objects for image recognition, limiting performance improvement when data sets are imbalanced across classes.
Innovation Solution
A data augmentation device that acquires source and target domain data, estimates a manifold structure, and generates new target domain data using the source domain's data distribution to balance class data sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is performed by perturbing hue, contrast, and other image properties within a limited range, then the class labels remain unchanged and data quantity increases, but the dynamic deformation of three-dimensional objects cannot be simulated and recognition performance improvement is limited
Solution Approach 1:
The patent introduces a pre-trained GAN model as an intermediary component that bridges the source domain and target domain. The GAN model generates augmented images by capturing complex data distributions and generating realistic variations that simulate dynamic deformations of three-dimensional objects, thereby resolving the limitation of conventional perturbation methods while maintaining class label consistency.
Solution Approach 2:
The patent changes the fundamental parameters of data generation from simple pixel-level perturbations (hue, contrast) to deep feature-level transformations using GANs. This parameter change enables the simulation of complex dynamic deformations while preserving class labels, as the GAN model operates on high-level semantic features rather than low-level pixel properties.
2Reliability
If data augmentation is applied to balance class data sets, then recognition performance improves, but existing methods cannot effectively handle classes with small number of data sets when data distribution differs between domains
Solution Approach 1:
The pre-trained GAN model serves as a mediator that transfers data distribution characteristics from the source domain to the target domain. For classes with few data sets, the GAN model generates synthetic samples that match the target domain's data distribution, enabling effective training even for imbalanced classes that would otherwise be underrepresented.
Solution Approach 2:
The patent uses the pre-trained GAN model to copy and generate synthetic data samples for classes with insufficient data. By copying the data generation capabilities from the source domain to the target domain, the system creates realistic synthetic samples that replicate the target domain's data distribution, thereby balancing class representations without requiring actual target domain data for all classes.
Data Source
AI summary
In a data augmentation device, a data acquisition means acquires two sets of source domain data of a predetermined class from a data group of a source domain, and acquires one set of target domain data of the predetermined class from a data group of a target domain data. An estimation means estimates a structure of a manifold representing a data distribution of the source domain by using two sets of source domain data. A data generation means generates new data of the target domain by using the one set of target domain data and the structure of the manifold.