Pseudo Image Generation for Field-to-Simulation Domain Conversion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning systems face challenges in efficiently performing model learning due to the state difference between simulation and field environments, requiring extensive data preparation and labeling, and existing domain conversion methods lack accuracy in converting field images to simulation images.

Innovation Solution

A machine learning system utilizing a pseudo image generative model and multiple determination models to convert field images into simulation images, where the system learns to distinguish between real and generated images without extensive data labeling, improving conversion accuracy through competitive learning and feedback mechanisms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If domain conversion method using generative model is used to convert field images to simulation images, then the state difference between simulation and field environments is reduced, but the conversion accuracy of the generative model needs improvement

Engineering Contradiction:
Improveconversion accuracyVSAvoidimage conversion precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the generative model's output is evaluated by a discriminator that provides feedback signals. The system uses adversarial training where the discriminator critiques the generated images, and this feedback is used to iteratively improve the generative model's conversion accuracy, resolving the contradiction between reliability and measurement precision.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces a discriminator as an intermediary component between the generative model and the final output. This intermediary evaluates the converted images and provides guidance, acting as a mediator that helps improve the conversion accuracy without requiring direct manual intervention, thus resolving the accuracy precision contradiction.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If extensive data preparation and labeling is performed for model learning, then learning accuracy is improved, but the time and resources required increase significantly

Engineering Contradiction:
Improvelearning accuracyVSAvoiddata preparation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent implements self-service through automated adversarial training where the system generates its own training data and automatically labels it through the discriminator's evaluation. The generative model and discriminator work together in a self-reinforcing loop that eliminates manual data preparation and labeling, achieving high learning accuracy without significant time investment.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent performs preliminary action by pre-training the generative model with available data and then using it to generate additional training samples before final model training. This preliminary generation of synthetic training data reduces the need for extensive manual data collection and labeling, saving time while maintaining learning accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3667566B1Machine learning system, domain conversion device, and machine learning method
Publication Date: 2025.10.29 HITACHI LTD
  • EP3667566B1 patent drawingFigure 1~2
  • EP3667566B1 patent drawingFigure 3
  • EP3667566B1 patent drawingFigure 4

AI summary

There is provided is a machine learning system that efficiently performs learning of a model that domain-converts an image. The machine learning system includes a first determination model that determines whether an input image is a second domain image, and a second determination model that determines whether an extracted image obtained by extracting an area where an object is presented from the input image is an extracted image obtained by extracting an area where an object is presented from the second domain image. Either a pseudo second domain image or the second domain image is selected and input into the first determination model, and either a first extracted image in the pseudo second domain image or a second extracted image in the second domain image is selected and input into an image extracting unit. Learning of the first determination model is performed based on a first determination result of the first determination model, learning of the second determination model is performed based on a second determination result of the second determination model, and learning of a pseudo image generative model, that coverts a first domain image obtained by capturing an image for an object into the pseudo second domain image, is performed based on the first determination result and the second determination result.