Color-Space Deformation for Neural Network Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex neural networks are prone to overfitting, especially when they have a large number of parameters, as they tend to memorize training data rather than generalize well to new examples, and existing methods for mitigating overfitting through synthetic data generation are limited in increasing the size of the training set and are inferior to collecting more independent data.
Innovation Solution
A system that generates training images by performing color-space deformations on existing training images using an image processing module, which includes a random number generator, principal component analysis module, and arithmetic logic unit, applying deformations such as altering RGB channels and adding noise based on eigenvalues and random variables to create diverse and uncorrelated training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more training data is collected to mitigate overfitting, then the neural network's generalization ability is improved, but the labor and expense increase significantly
Solution Approach 1:
The patent creates synthetic copies of existing training images by applying color-space deformations. Instead of collecting new independent images, the system generates additional training samples by transforming the color values of existing images through principal component analysis and random variable manipulation in color space, thereby reducing data collection costs while maintaining training effectiveness
Solution Approach 2:
The patent transforms training images by changing their color-space parameters through principal component analysis. By manipulating the eigenvalues and eigenvectors of the color covariance matrix and adding random variables, the system generates diverse training samples with varied color characteristics, effectively increasing training data quantity without physical data collection
2Ease of manufacture
If synthetic training data is generated from existing data, then data collection cost is reduced, but the generated images are highly correlated with original images and the training set size increase is limited
Solution Approach 1:
The patent moves the data augmentation process from spatial domain transformations to color-space domain transformations. By performing principal component analysis on color values and applying deformations in this additional dimension (color space rather than spatial space), the system generates more diverse and less correlated training samples, effectively increasing the usable training set size
3Productivity
If complex neural networks with large number of parameters are used, then the network's processing capability is improved, but overfitting becomes increasingly likely
Solution Approach 1:
The patent enables the training data to serve itself by automatically generating additional training samples from existing data through color-space deformations. This self-service approach to data augmentation provides more training examples without external intervention, helping complex networks generalize better without requiring proportional increases in manually collected data
Data Source
AI summary
A system and method for generating training images. An existing training image is associated with a classification. The system includes an image processing module that performs color-space deformation on each pixel of the existing training image and then associates the classification to the color-space deformed training image. The technique may be applied to increase the size of a training set for training a neural network.