Label Reliability Assessment for Image Object Recognition

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing techniques for recognizing objects in images face accuracy issues due to incorrect label information, which can lead to decreased performance of object recognition systems.

Innovation Solution

An information processing apparatus that calculates the reliability of label information and displays it alongside images, allowing users to efficiently review and modify label information, thereby correcting errors and improving data accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual labeling operation is repeated to ensure sufficient accuracy, then the accuracy of learning data is improved, but the time and labor required for data preparation increases significantly

Engineering Contradiction:

Improveaccuracy of label informationVSAvoidtime for manual labeling

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system automatically evaluates the accuracy of label information using a recognizer and reliability calculation unit, eliminating the need for continuous manual verification. The label information itself serves to evaluate its own reliability through the recognizer's assessment

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system provides feedback by calculating and displaying reliability values for each piece of learning data, allowing operators to identify and correct inaccurate labels efficiently. The reliability information guides selective review rather than exhaustive manual checking

Inventive Principle:

Principle #23Feedback

2Productivity

If label information is set by person without verification, then the productivity of data preparation is improved, but the reliability of learning data decreases due to potential errors

Engineering Contradiction:

Improvespeed of label information setupVSAvoidaccuracy of label information

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The reliability calculation unit acts as an intermediary between the manual labeling process and the learning system. It automatically assesses the quality of human-generated labels without requiring direct human verification of each label, maintaining both speed and reliability

Inventive Principle:

Principle #24Intermediary (Mediator)

3Reliability

If all label information is reviewed manually to ensure accuracy, then the reliability of learning data is improved, but the efficiency of the review process decreases

Engineering Contradiction:

Improveaccuracy of learning dataVSAvoidefficiency of label review

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

Instead of uniformly reviewing all label information, the system applies quality assessment locally to each individual label or image. The reliability calculation unit evaluates each piece of learning data independently, allowing selective review only where needed based on calculated reliability scores

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS10964057B2Information processing apparatus, method for controlling information processing apparatus, and storage medium

Publication Date: 2021.03.30 CANON KK

AI summary

An information processing apparatus comprising: at least one processor programmed to cause the apparatus to: hold label information regarding presence of a target object, the label information being set for the target object in an image; obtain a reliability of the label information; cause a display apparatus to display the label information and an image corresponding to the label information in the image, based on the reliability; accept an operation made by a user; and modify the label information based on the operation.