Learning Data Generation for Defect Identification Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image identification devices struggle to achieve high accuracy in detecting various defects due to insufficient and inadequate learning data, particularly when images featuring defects that the device is not well-trained on.

Innovation Solution

A data generation apparatus and method that generates new learning data by associating corrected determination results with images or composite images, specifically addressing erroneous determinations by adding information about the presence or absence of defects, thereby improving the identification device's training and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If learning data is accumulated to improve identification accuracy, then identification accuracy is improved, but the time and resources required for data preparation and processing increase

Engineering Contradiction:
Improveidentification accuracyVSAvoiddata preparation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system uses the identification device's determination results as feedback to automatically generate new learning data. When the identification device makes a determination on an image, the system automatically creates corrected learning data based on that determination, eliminating the need for manual data collection and preparation while continuously improving identification accuracy through iterative learning.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system enables the identification device to improve itself autonomously by generating its own learning data from its determination results. The learning data generation unit automatically creates training data using the images and determination results produced by the identification device, allowing the system to self-enhance without external intervention or additional manual data preparation.

Inventive Principle:
Principle #25Self-service

2Adaptability or versatility

If learning data includes sufficient types of images with defects to handle various defect types, then the identification device can handle various defects, but the complexity of data collection and management increases

Engineering Contradiction:
Improveability to handle various defectsVSAvoiddata collection and management complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The identification device automatically generates diverse learning data covering various defect types through its own determination process. As the device processes different images and makes determinations, the system automatically creates corresponding learning data entries, enabling the device to handle various defect types without requiring complex manual data collection and management systems.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system prepares learning data in advance by automatically generating it from the identification device's determination results on existing images. This preliminary generation of learning data ensures that the device is pre-trained on diverse defect types before actual inspection tasks, reducing the need for complex ongoing data management while maintaining high adaptability.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3502966B1Data generation apparatus, data generation method, and data generation program
Publication Date: 2023.08.23 OMRON CORP
  • EP3502966B1 patent drawingFigure 1
  • EP3502966B1 patent drawingFigure 2
  • EP3502966B1 patent drawingFigure 3

AI summary

Provided are a data generation apparatus, data generation method, and a data generation program for accumulating, as learning data, an image that includes a feature on which an identification device could not output a desirable identification result. A data generation apparatus includes: an acquisition unit configured to acquire a result determined by a determination unit, which uses an identification device trained using learning data, as to whether or not an object to be inspected includes a part to be detected based on an image of the object to be inspected; an evaluation unit configured to evaluate whether or not the determination result is correct; and a generation unit configured to generate, if the evaluation unit has evaluated that the determination result is not correct, new learning data by associating at least one of the image and a composite image generated based on the image with information in which the determination result is corrected.