Training Label Image Correction via Trained Model Segmentation

Solution Overview

Problem

In machine learning for image segmentation, the creation of training label images is hindered by variations in threshold settings among creators, leading to inconsistent boundaries and the need for extensive manual corrections, making the process inefficient and inaccurate.

Innovation Solution

A method that uses a trained model to statistically average boundary variations in multiple training label images, enabling automatic correction of label areas based on comparison with a determination label image, thereby establishing consistent criteria for boundary definitions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If training label images are created by multiple label creators manually adjusting thresholds, then individual differences in threshold setting are introduced, but this leads to variations in boundary definitions across training label images

Engineering Contradiction:

Improvemanual threshold adjustmentVSAvoidboundary consistency

Core Design Contradiction:

Ease of operationVSManufacturing precision

Solution Approach 1:

The system performs self-correction by automatically identifying and correcting boundary variations in training label images using the trained model's segmentation results as reference, eliminating the need for manual intervention while maintaining consistency

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system uses feedback from the trained model's segmentation results to identify boundary variations in training label images and automatically corrects them, creating a closed-loop system that continuously improves boundary consistency

Inventive Principle:

Principle #23Feedback

2Manufacturing precision

If dedicated trained models are created to correct training label image boundaries, then boundary consistency can be improved, but this requires learning from highly accurate training label images created with consistent criteria, which takes a lot of effort

Engineering Contradiction:

Improveboundary consistencyVSAvoidcorrection effort

Core Design Contradiction:

Manufacturing precisionVSLoss of time

Solution Approach 1:

The system performs self-correction by automatically identifying and correcting boundary variations in training label images using the trained model's segmentation results as reference, eliminating the need for manual intervention while maintaining consistency

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The trained model's segmentation results serve as an intermediary reference to automatically correct boundary variations in training label images, avoiding the need for manual correction while maintaining consistency

Inventive Principle:

Principle #24Intermediary (Mediator)

3Productivity

If training label images with varied boundaries are used for machine learning, then the process can proceed without extensive manual correction, but the segmentation results may lack consistent criteria for boundary definitions

Engineering Contradiction:

Improvetraining efficiencyVSAvoidboundary consistency

Core Design Contradiction:

ProductivityVSManufacturing precision

Solution Approach 1:

The system performs self-correction by automatically identifying and correcting boundary variations in training label images using the trained model's segmentation results as reference, eliminating the need for manual intervention while maintaining consistency

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system uses feedback from the trained model's segmentation results to identify boundary variations in training label images and automatically corrects them, creating a closed-loop system that continuously improves boundary consistency

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS11830195B2Training label image correction method, trained model creation method, and image analysis device

Publication Date: 2023.11.28 SHIMADZU CORP

AI summary

A training label image correction method includes performing a segmentation process on an input image (11) of training data (10) by a trained model (1) using the training data to create a determination label image (14), comparing labels of corresponding portions in the determination label image (14) and a training label image (12) with each other, and correcting label areas (13) included in the training label image based on label comparison results.