Style- and Content-Matched Image Synthesis for Small Datasets

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing image-based machine-learning models struggle with small training datasets, leading to inefficiencies and inaccuracies due to insufficient training images and lack of diversity, resulting in overfitting and discounted original dataset features.

Innovation Solution

A style matching system that utilizes a generative machine-learning model to expand a small image dataset by selecting and synthesizing images from a catalog of stored images, ensuring the expanded dataset maintains the same style and content distribution as the original, thereby enabling accurate and efficient training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a small image dataset is used for training, then training time and computational resources are reduced, but the model cannot be accurately trained due to insufficient training images

Engineering Contradiction:

Improvetraining efficiencyVSAvoidmodel accuracy

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent uses a generative machine-learning model to create synthetic images that copy the style and content characteristics of the original small dataset. These generated images serve as artificial copies that expand the training dataset without requiring additional real images, thereby providing sufficient training data while maintaining the original dataset's integrity and characteristics.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent transforms the original small dataset by applying style transfer and content modification parameters to generate diverse variations. By changing parameters such as style attributes, content compositions, and visual characteristics while preserving the core data distribution, the system creates a expanded dataset that maintains fidelity to the original while providing sufficient diversity for accurate model training.

Inventive Principle:

Principle #35Parameter changes

2Device complexity

If a small image dataset is used, then data processing complexity is reduced, but the dataset lacks the diversity needed to generate robust models

Engineering Contradiction:

Improvedata processing complexityVSAvoidmodel robustness

Core Design Contradiction:

Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent performs preliminary style analysis and content extraction on the original small dataset to establish style profiles and content characteristics before generating new images. By pre-processing and understanding the stylistic and content parameters of the original data, the system can subsequently generate diverse variations that maintain consistency with the original dataset's characteristics while expanding diversity.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent expands the dataset by adding a new dimension of stylistic variation while maintaining the original content structure. By introducing style transfer as an additional dimension of diversity beyond the original limited variations, the system generates images that explore new stylistic spaces while preserving the core content relationships, thereby enhancing model robustness without fundamentally altering the data processing complexity.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

3Quantity of substance

If larger image datasets are combined with small datasets, then training data quantity increases, but the models produce results fitted to the larger datasets while discounting the original small dataset

Engineering Contradiction:

Improvetraining data quantityVSAvoidoriginal dataset features

Core Design Contradiction:

Quantity of substanceVSLoss of information

Solution Approach 1:

Instead of combining the small dataset with unrelated larger datasets, the patent generates synthetic images that specifically copy and preserve the stylistic and content features of the original small dataset. This ensures that the expanded training data maintains fidelity to the original data distribution and characteristics, preventing the loss of original dataset features while still increasing data quantity.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The generative machine-learning model acts as an intermediary that translates the original small dataset into an expanded set of training images. This intermediary process ensures that the expansion maintains the statistical properties and feature distributions of the original data, serving as a faithful mediator that preserves original characteristics while providing the quantity needed for robust training.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12354324B2Generating large datasets of style-specific and content-specific images using generative machine-learning models to match a small set of sample images

Publication Date: 2025.07.08 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

The present disclosure relates to utilizing a style-matching image generation system to generate large datasets of style-matching images having matching styles and content to an initial small sample set of input images. For example, the style-matching image generation system utilizes a selection of style-mixed stored images with a generative machine-learning model to produce large datasets of synthesized images. Further, the style-matching image generation system utilizes the generative machine-learning model to conditionally sample synthesized images that accurately match the style, content, characteristics, and patterns of the initial small sample set and that also provide added variety and diversity to the large image dataset.