Active Learning Image Selection for Balanced Inspection Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge of selecting a subset of images for labeling in machine learning models for inspection systems is cost-prohibitive and can lead to unbalanced models, hindering their performance in production environments.

Innovation Solution

An active learning framework that manages image data and employs algorithms to select and label images efficiently, using strategies like K-Means clustering, outlier detection, and uncertainty sampling to train inspection models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large number of images are labeled by experts for training machine learning models, then model accuracy is improved, but labeling cost and time increase significantly

Engineering Contradiction:
Improvemodel accuracyVSAvoidlabeling time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system uses active learning algorithms to automatically select which images should be labeled by experts, enabling the system to serve itself by identifying the most valuable training samples without human intervention in the selection process

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

Instead of labeling all available images or using a fixed subset, the system performs partial labeling only on the most informative images identified through active learning, achieving high model accuracy with a smaller, strategically selected subset of labeled data

Inventive Principle:
Principle #16Partial or excessive action

2Loss of energy

If a subset of images is selected for labeling to reduce cost, then labeling cost is reduced, but model balance and performance may deteriorate

Engineering Contradiction:
Improvelabeling costVSAvoidmodel performance
Core Design Contradiction:
Loss of energyVSReliability

Solution Approach 1:

The active learning system continuously monitors model performance and uses this feedback to iteratively select the next batch of images that will most improve model accuracy, ensuring that each labeling investment maximizes model performance

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system dynamically adjusts selection criteria and labeling strategies based on model performance metrics and data distribution analysis, changing parameters such as selection thresholds and sampling strategies to maintain model balance while minimizing labeling costs

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If all available images are processed through multiple selection strategies, then comprehensive image selection is achieved, but processing time and computational resources increase

Engineering Contradiction:
Improveimage selection comprehensivenessVSAvoidprocessing speed
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The system performs preliminary processing of images including transformations and metadata generation before they are needed by selection strategies, so that when multiple strategies are applied, the images are already prepared and ready for rapid evaluation

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system merges multiple image selection strategies and combines their results, allowing comprehensive image selection through diverse criteria while managing computational overhead by coordinating the strategies efficiently

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4268125B1Active learning management system for automated inspection systems
Publication Date: 2026.02.25 3M INNOVATIVE PROPERTIES CO
  • EP4268125B1 patent drawingFigure 1
  • EP4268125B1 patent drawingFigure 2A
  • EP4268125B1 patent drawingFigure 2B

AI summary

An example method for selecting product images for training a machine-learning model includes obtaining product images to include in an image population; receiving an indication of an image selection strategy for determining if a product image is to be included in a set of images of interest; determining image transforms based on configuration data for the indicated image selection strategy, wherein the image transforms perform image manipulation operations to obtain transformed image data for each of the product images in the image population; selecting a subset of images from the image population for inclusion in the set of images of interest based on the indicated image selection strategy and the transformed image data; determining one or more descriptive labels and applying the one or more descriptive labels to the respective sets of images; and training an inspection model for a product inspection system based on the labeled images.