Segmented Workpiece Images for Teacher Data Generation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing systems require a large number of non-defective and defective images for training a learning model, which is inefficient and resource-intensive.

Innovation Solution

A method of generating teacher data by segmenting image data into multiple areas, marking each area for predetermined information, and training the learning model using this segmented data, including background color variations and difference information, to reduce the number of required images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large number of non-defective and defective images are used for training, then the determination accuracy is improved, but the resource consumption and training complexity increase

Engineering Contradiction:

Improvedetermination accuracyVSAvoidnumber of images

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent segments a single workpiece image into multiple first areas (regions of interest) and generates multiple pieces of teacher data from one image by marking different areas with different ground truth labels. This segmentation approach allows the system to create multiple training samples from a single image, thereby reducing the total number of images needed while maintaining sufficient training data diversity for accurate defect detection

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by creating different markings for different areas within the same image. Each first area receives appropriate marking based on its content (defective or non-defective), allowing the learning model to learn from localized features and contexts. This enables more efficient use of training data by extracting multiple learning signals from different regions of a single image

Inventive Principle:

Principle #3Local quality

2Quantity of substance

If multiple pieces of teacher data are generated from one image data, then the number of required images is reduced, but the device complexity increases

Engineering Contradiction:

Improvenumber of imagesVSAvoiddata processing complexity

Core Design Contradiction:

Quantity of substanceVSDevice complexity

Solution Approach 1:

The system segments the image into multiple first areas and generates multiple teacher data pieces through automated marking processes. While this increases processing steps, the segmentation is systematic and rule-based, allowing for programmatic implementation that manages complexity through structure and repeatability

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system employs automated marking mechanisms that can identify and label areas independently without requiring extensive manual annotation for each generated teacher data piece. The marking process is self-service in nature, using algorithms to determine appropriate labels based on image content analysis, thereby reducing the need for manual intervention despite generating multiple data pieces

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS12444034B2Teacher data generation method, trained learning model, and system

Publication Date: 2025.10.14 CANON KK

AI summary

According to one aspect of the present invention, a method of generating teacher data for image recognition includes acquiring image data by capturing an image of a workpiece, and segmenting the image data into a plurality of first areas, marking whether predetermined information is included in each of the plurality of first areas, and generating a plurality of pieces of teacher data.