Learning Model Contour Evaluation for Incomplete Elongated Masks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing learning models for image processing in appearance inspection struggle to accurately evaluate performance when dealing with elongated subjects, as incomplete masks can occur, leading to difficulties in assessing their effectiveness using metrics like mIoU.

Innovation Solution

A method and system that evaluates learning models by comparing first and second contour information, calculating inter-contour distances, and using mean values and standard deviations to assess model performance, allowing for appropriate evaluation even with incomplete masks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If segmentation processing is performed using a learning model to identify pixels belonging to a subject, then the subject can be detected and classified, but the mask may stick out from or become discontinuous for elongated subjects, making it difficult to appropriately evaluate the learning model's performance

Engineering Contradiction:
Improveevaluation accuracyVSAvoidmask completeness
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent segments the evaluation into two independent parts: contour information extraction and performance evaluation. The contour information is extracted separately from the mask, allowing the evaluation to focus on contour accuracy rather than being affected by mask completeness issues. This segmentation enables appropriate evaluation even when masks are incomplete or discontinuous for elongated subjects.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts contour information from the mask and uses it as a separate evaluation criterion. By taking out the contour information from the mask and using it independently, the evaluation can assess the learning model's performance based on contour accuracy rather than being compromised by mask stick-out or discontinuity problems.

Inventive Principle:
Principle #2Taking out (Extraction)

2Productivity

If mean Intersection over Union (mIoU) is used as the evaluation metric, then the overall area overlap can be measured, but the metric is unlikely to largely degrade even when part of an elongated object is missing, making it insufficient for evaluating contour accuracy

Engineering Contradiction:
Improveevaluation efficiencyVSAvoidcontour evaluation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments the evaluation into two independent parts: contour information extraction and performance evaluation. The contour information is extracted separately from the mask, allowing the evaluation to focus on contour accuracy rather than being affected by mask completeness issues. This segmentation enables appropriate evaluation even when masks are incomplete or discontinuous for elongated subjects.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces local quality evaluation by measuring the distance between contours at specific locations. Instead of using a global metric like mIoU that averages over the entire area, the patent evaluates local contour accuracy by calculating distances at corresponding points between the predicted and actual contours, providing precise measurement of contour quality.

Inventive Principle:
Principle #3Local quality

3Measurement precision

If contour information is extracted and evaluated separately from the mask, then the learning model can be appropriately evaluated based on contour accuracy, but additional processing steps are required to extract and compare contour information

Engineering Contradiction:
Improvecontour evaluation accuracyVSAvoidevaluation process complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent performs preliminary action by extracting contour information from the mask before the evaluation process. The contour information is extracted in advance and stored, so that during evaluation, the system can directly compare the extracted contour information with the ground truth contour information without performing complex real-time processing. This preliminary extraction simplifies the overall evaluation process.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates a copy of the contour information from the predicted mask and compares it with the ground truth contour information. By copying the contour information and storing it for later comparison, the evaluation process becomes more straightforward, as the system only needs to compare the copied contour information with the reference rather than performing complex real-time analysis.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4583045A1Method, learning model evaluation system, and program
Publication Date: 2025.07.09 OMRON CORP
  • EP4583045A1 patent drawingFigure 1
  • EP4583045A1 patent drawingFigure 2
  • EP4583045A1 patent drawingFigure 3(a)~3(d)

AI summary

Provided are a method, a learning model evaluation system, and a program that can appropriately evaluate a learning model for identifying a subject to be detected. A method includes: by a computer, inputting a subject image in which at least one subject to be detected is imaged to a learning model and acquiring first contour information indicating a contour of the at least one subject to be detected in the subject image; acquiring second contour information indicating a contour of the at least one subject to be detected and serving as a reference for evaluating the first contour information; acquiring a condition to be satisfied by the first contour information; and generating, based on the first contour information, the second contour information, and the condition, evaluation information in which performance of the learning model is evaluated.