Image Generation Model Training via Content Information Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI-based image generation models often suffer from bias towards specific dataset domains due to camera characteristics, leading to distorted structural information and reduced reliability of generated images.
Innovation Solution
A method and system for training an image generation model by receiving a training image in a first domain style, extracting content information, generating augmented content information through image processing, and training the model to generate synthetic images in the first domain style from input images in a second domain style.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If an image generation model is trained on images from a specific domain, then the model can generate images in that domain style, but the model becomes biased toward that domain and produces distorted structural information
Solution Approach 1:
The patent applies parameter changes by transforming content information through multiple image processing operations (translation, rotation, flipping, scaling, cropping, brightness adjustment, saturation adjustment, noise injection) to create augmented content information. This modifies the training parameters and variations the model exposure to, enabling it to generate reliable images across different domain styles without being biased toward a single domain.
2Manufacturing precision
If the model is trained to generate images in a specific domain style, then the generated images match the target style, but detailed parts of images are not properly generated and structural information is distorted
Solution Approach 1:
The patent applies preliminary action by extracting content information from training images before generating augmented content information through various image processing operations. This preliminary extraction and augmentation of structural content information ensures that detailed parts and structural relationships are properly preserved and transferred to generated images, preventing distortion while maintaining generation quality.
3Reliability
If augmented content information is generated through multiple image processing operations, then the model learns more variations and generates higher quality images, but the training process becomes more complex
Solution Approach 1:
The patent applies segmentation by dividing the training process into distinct modules: content information extraction, augmented content information generation through separate image processing operations, and model training. This segmentation of the training process manages complexity by organizing operations into independent, manageable steps while still achieving comprehensive training data augmentation for improved reliability.
Data Source
AI summary
The present disclosure relates to a method of training an image generation model performed by at least one processor. The method of training an image generation model includes receiving a training image in a first domain style, extracting, by the at least one processor, first content information for the training image, generating, by the at least one processor, a plurality of pieces of augmented content information perturbed from the first content information as part of image processing by augmenting the first content information, and training an image generation model to generate a synthetic image in the first domain style from an input image in a second domain style different from the first domain style, wherein the training of the image generation model is based on the training image, the first content information, and the plurality of pieces of augmented content information.