Frequency-Domain Image Texture Augmentation for Domain Shift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face reduced accuracy and increased resource consumption due to domain bias and shift when trained on data from different sources, leading to inefficient retraining and processing requirements.
Innovation Solution
Image texture manipulation techniques are applied to modify the textures of image files in a data set to enhance consistency and similarity, allowing the model to generalize across domains and reduce the need for retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are trained on data from different sources, then the model can handle diverse data sets, but domain bias and shift reduce model accuracy
Solution Approach 1:
The patent transforms image data from spatial domain to frequency domain using Fourier transform, then manipulates texture parameters by modifying frequency components. This parameter transformation allows consistent texture representation across diverse images while maintaining domain adaptability, resolving the contradiction between handling diverse data sets and maintaining model accuracy
Solution Approach 2:
The patent introduces frequency domain representation as an intermediary between raw image data and model processing. By converting images to frequency domain, extracting texture features, and transforming back, the system creates a standardized intermediate representation that eliminates domain bias while preserving essential image characteristics, enabling accurate processing of diverse data sets
2Reliability
If machine learning models are retrained to improve accuracy on diverse data, then model accuracy increases, but processing and memory resource consumption increases
Solution Approach 1:
The patent performs preliminary texture manipulation and data augmentation before model training by generating synthetic training images with consistent texture properties. This pre-processing creates a standardized training data set that reduces domain bias, allowing the model to achieve high accuracy without requiring extensive retraining and reducing computational resources needed for iterative model improvement
Solution Approach 2:
The patent creates synthetic copies of training images by manipulating texture in the frequency domain. By generating augmented training data through Fourier transform-based texture manipulation rather than collecting and processing additional real images, the system reduces memory and processing resource requirements while providing sufficient training data for accurate model learning
Data Source
AI summary
Implementations described herein relate to image texture manipulation for machine learning data augmentation. In some implementations, a computer vision platform may transform first data of a first image file to a first frequency domain representation of the first image file. The computer vision platform may identify, based on the first frequency domain representation, a first subset of frequencies for the first image file. The computer vision platform may transform second data of a second image file to a second frequency domain representation of the second image file. The computer vision platform may identify, based on the second frequency domain representation, a second subset of frequencies for the second image file. The computer vision platform may generate a third image file based on the first subset of frequencies and the second subset of frequencies. The computer vision platform may store the third image file in a set of image files.