Image Recognition Feature Alignment Between CG and Photographic Images

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The accuracy of image recognition using classifiers trained with computer graphics (CG) images is limited due to the feature differences between CG and photographic images, making it difficult to improve recognition accuracy with available training data.

Innovation Solution

A training apparatus and method that identifies intermediate features from both CG and photographic images, calculates offset features to adjust for differences, and trains a second classifier using post-offset intermediate features to enhance recognition accuracy on photographic images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If a classifier is trained using CG images, then training data availability is improved, but image recognition accuracy on photographic images deteriorates due to feature differences

Engineering Contradiction:
Improvetraining data availabilityVSAvoidimage recognition accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent introduces an intermediate feature space as a mediator between CG images and photographic images. By learning a mapping from CG image features to photographic image features through intermediate features, the system bridges the domain gap and enables accurate recognition on photographic images while training primarily on CG data

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms the feature representation parameters by learning domain-specific feature mappings. It changes the parameter space of intermediate features to accommodate both CG and photographic image characteristics, enabling the classifier to adapt from CG training data to photographic test data

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If more photographic images are collected for training, then image recognition accuracy is improved, but data acquisition difficulty increases

Engineering Contradiction:
Improveimage recognition accuracyVSAvoiddata acquisition difficulty
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent creates synthetic copies of photographic images using CG rendering. By generating realistic CG images that replicate the appearance and features of photographic images, it provides abundant training data without the need to collect actual photographic samples

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The intermediate feature mapping serves as a mediator that transfers knowledge from CG images to photographic images, eliminating the need for direct photographic training data while maintaining recognition accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3617991B1Learning device, image recognition device, learning method, and program
Publication Date: 2026.01.28 SONY INTERACTIVE ENTERTAINMENT LLC
  • EP3617991B1 patent drawingFigure 1~2
  • EP3617991B1 patent drawingFigure 3
  • EP3617991B1 patent drawingFigure 4~5

AI summary

Provided are a training apparatus, an image recognition apparatus, a training method, and a program capable of improving the accuracy of image recognition of a photographic image using classifiers that have been trained using CG images. An intermediate feature identifying section (44) identifies an intermediate feature. An offset feature identifying section (46) identifies an offset feature on the basis of a CG intermediate feature and a photographic intermediate feature. A post-offset intermediate feature identifying section (48) identifies a post-offset intermediate feature associated with a CG image on the basis of the intermediate feature associated with the CG image and the offset feature. A second-classifier training section (50) executes training of a second classifier (40b) using the post-offset intermediate feature associated with the CG image, the second classifier (40b) being configured to receive, in a case where image recognition is performed on a photographic image, the intermediate feature associated with the photographic image.