Neural Network Object Sample Selection for Class Imbalance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural networks used for object detection face performance issues due to class imbalance in training data, where variations in the number of object samples from different classes lead to poor performance, especially for classes with fewer samples.

Innovation Solution

A method for selecting object samples from multiple datasets by determining an importance score for each sample, defining thresholds, and balancing the number of samples across classes to ensure equal representation while maintaining high quality and relevance, thereby alleviating class imbalance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If object samples are selected from multiple datasets with different properties, then the training data diversity is improved, but class imbalance occurs causing poor performance for classes with fewer samples

Engineering Contradiction:
Improvetraining data diversityVSAvoidperformance for classes with fewer samples
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent changes the parameter of sample selection by introducing importance scores and thresholds. Instead of uniformly selecting samples, it dynamically adjusts which samples are selected based on their importance scores, ensuring that samples from underrepresented classes with high importance scores are included while maintaining balance across classes.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements a feedback mechanism by calculating the variation of sample numbers across classes and using this information to adjust the selection process. The method iteratively selects samples and monitors class balance, ensuring that the final training set achieves both diversity and balanced representation.

Inventive Principle:
Principle #23Feedback

2Reliability

If the number of object samples from different classes is balanced, then the performance for classes with fewer samples is improved, but the quality and relevance of selected samples may be compromised

Engineering Contradiction:
Improveperformance for classes with fewer samplesVSAvoidquality and relevance of selected samples
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent introduces importance scores as a new parameter to evaluate sample quality and relevance. By setting importance score thresholds, the method ensures that only samples meeting minimum quality standards are selected, while still achieving class balance through the scoring mechanism.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If importance score thresholds are set high to ensure sample quality, then the relevance of selected samples is improved, but the number of available samples for balancing decreases

Engineering Contradiction:
Improverelevance of selected samplesVSAvoidnumber of available samples
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent uses feedback from the variation calculation to adjust the selection process. By monitoring the balance across classes, the method can identify when additional samples are needed and selectively include samples with importance scores at or near the threshold, thereby maintaining quality while achieving sufficient quantity for balance.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4187438A1Object sample selection for training of neural networks
Publication Date: 2023.05.31 AXIS
  • EP4187438A1 patent drawingFigure 1
  • EP4187438A1 patent drawingFigure 2A~2B
  • EP4187438A1 patent drawingFigure 3

AI summary

The present invention generally relates to method for selecting object samples (208) for training of a neural network (200) from more than one dataset (202, 204, 206) comprising annotated object samples of at least two object classes ("Class A-D"), the method comprising: determining (S102) an importance score for at least a portion of the annotated object samples; defining (S104) a set of importance score thresholds (210); and selecting (S106) a number of annotated object samples (208) from each object class that fulfill a respective importance score threshold, and that provides the smallest variation of the number of object samples between the object classes, to be used for training of the neural network (200).