Image Recognition Feature Data Registration via Subclass Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The degradation of distinguishability between classes in image recognition systems occurs when new feature data is added, leading to potential false identifications due to the proximity or overlap between classes in the feature space, especially when feature data varies with photographing conditions.

Innovation Solution

Registering new feature data in a class different from the existing class of the object, based on predetermined similarity conditions, to prevent degradation of distinguishability and enhance recognition accuracy by maintaining class separation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If new feature data is added to enhance variation and robustness against photographing condition differences, then recognition accuracy is improved, but distinguishability between classes degrades due to proximity or overlap between classes in feature space

Engineering Contradiction:
Improverecognition accuracyVSAvoiddistinguishability between classes
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies segmentation by dividing the class of a first object into multiple subclasses based on similarity to another object. When new feature data shows high similarity to an existing object, it creates a new subclass rather than adding to the original class. This segments the feature space to maintain distinguishability while still incorporating the new data, resolving the contradiction between enhancing variation and preserving class separation.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If feature data from images with different photographing conditions are registered to improve robustness, then the variation of feature data is enhanced, but false identification occurs due to overlap between classes

Engineering Contradiction:
Improverobustness against photographing conditionVSAvoididentification accuracy
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies local quality by evaluating the similarity between new feature data and existing objects on a case-by-case basis. Instead of uniformly adding all feature data, it locally adapts the registration strategy: high-similarity data triggers subclass creation while low-similarity data follows standard registration. This localized approach maintains robustness while preventing false identifications.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP2955666B1Image recognition device and method for registering feature data in image recognition device
Publication Date: 2020.04.29 OMRON CORP
  • EP2955666B1 patent drawingFigure 1
  • EP2955666B1 patent drawingFigure 2A~2C
  • EP2955666B1 patent drawingFigure 3

AI summary

Provided is a technology of suppressing degradation of distinguishability from another object when new feature data is added to an already-registered object. An image recognition device includes a database in which pieces of feature data of plural objects are registered while divided into classes in each object, an identification unit that identifies an unknown object, and a feature data registration unit that registers the feature data in the database. The database can set plural classes to the identical object. In adding new feature data to a first object already registered in the database, the feature data registration unit sets a new class other than the existing class to the first object, and registers the new feature data in the new class, when a second object satisfying a predetermined similar condition with the new feature data exists in the database.