Training Image Generation Using Ideal and Degraded Image Pairs

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The generation of training data for machine learning models is hindered by the need for large amounts of actual images, which are prone to human error and contain inherent degradation elements, making them unsuitable as ideal teacher data.

Innovation Solution

A method to generate ideal images from captured images by suppressing degradation factors like lens and sensor characteristics, followed by creating degraded images based on target device parameters, to produce high-quality training data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If actually-shot images are used as training data, then data collection is simple, but the images contain degradation elements and human errors making them unsuitable as ideal teacher data

Engineering Contradiction:

Improveease of data collectionVSAvoidimage quality

Core Design Contradiction:

Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent creates ideal teacher data by generating synthetic images through rendering scenes with known ground truth information, rather than using actually-shot images. This copying approach produces perfect reference images without degradation elements while maintaining realistic appearance through simulated camera characteristics.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent introduces a rendering system as an intermediary between the real world and training data. This intermediary generates ideal images by computationally synthesizing scenes with known parameters, acting as a mediator that produces perfect teacher data without the flaws of actual photography.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Ease of operation

If actually-shot images are used as training data, then preparation is straightforward, but degradation elements cannot be added accurately or in a limited manner

Engineering Contradiction:

Improvedata preparation simplicityVSAvoiddegradation control accuracy

Core Design Contradiction:

Ease of operationVSMeasurement precision

Solution Approach 1:

The patent enables precise control of degradation elements by changing parameters during the rendering process. By adjusting rendering parameters such as camera characteristics, scene conditions, and environmental factors, the system can accurately add or remove specific degradation elements in a controlled manner.

Inventive Principle:

Principle #35Parameter changes

3Productivity

If actually-shot images are used as teacher data, then no additional processing is needed, but they cannot serve as ideal images due to included degradation elements

Engineering Contradiction:

Improvedata generation efficiencyVSAvoidteacher data quality

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent performs preliminary actions by generating ideal teacher data through rendering before the learning process begins. By pre-computing scenes with known ground truth and perfect image quality, the system prepares reliable teacher data in advance, eliminating the need for post-processing actually-shot images and ensuring high reliability from the start.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12592058B2Data generating method, learning method, estimating method, data generating device, and program

Publication Date: 2026.03.31 SONY SEMICON SOLUTIONS CORP

AI summary

An object is to obtain training data through an ideal image generated from an actually-shot image.A data generating method of generating data used in learning a model using a processor, the data generating method including: generating an ideal image corresponding to output data of the model from a captured image obtained by a predetermined device; and generating a degraded image corresponding to input data of the model from the ideal image.