Mini-Batch Learning for Semantic Segmentation Rare Classes

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing mini-batch learning methods for semantic segmentation result in low rare class determination accuracy due to class biases in the learning input and annotation images, leading to inadequate learning of rare classes.

Innovation Solution

A mini-batch learning apparatus that calculates the area ratio of classes in the data, specifies rare classes, and adjusts the update level of the machine learning model by reducing the learning coefficient for rare classes to mitigate class bias, thereby improving overall class determination accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If mini-batch learning is performed using conventional methods, then learning efficiency is improved, but rare class determination accuracy deteriorates due to class bias in the data

Engineering Contradiction:
Improvelearning efficiencyVSAvoidrare class determination accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by differentiating the update process based on class rarity. The update unit performs different update operations depending on whether a class is identified as rare or common. For rare classes, the update operation is adjusted to prevent overwriting learned features, while common classes receive standard updates. This localized differentiation in update quality resolves the contradiction by maintaining high learning efficiency overall while specifically protecting rare class determination accuracy.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If the learning coefficient is reduced for rare classes, then rare class determination accuracy is improved, but overall learning speed deteriorates

Engineering Contradiction:
Improverare class determination accuracyVSAvoidlearning speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent implements local quality by applying different learning coefficients to different classes based on their rarity. The update unit identifies rare classes and applies a reduced learning coefficient specifically to those classes, while maintaining the normal learning coefficient for common classes. This localized adjustment improves rare class determination accuracy without significantly impacting the overall learning speed, as the majority of common classes continue to learn at full speed.

Inventive Principle:
Principle #3Local quality

3Ease of manufacture

If class bias in training data is not addressed, then training simplicity is maintained, but model fairness and accuracy for rare classes deteriorates

Engineering Contradiction:
Improvetraining simplicityVSAvoidmodel accuracy for rare classes
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent implements feedback by introducing a feedback unit that monitors the determination accuracy for each class. The feedback unit provides information about rare classes back to the update unit, which then adjusts the update operation accordingly. This feedback mechanism maintains training simplicity by automatically identifying and handling rare classes without requiring manual intervention, while simultaneously improving model accuracy for rare classes through targeted update adjustments.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3896647B1Mini-batch learning device, operating program for mini-batch learning device, operating method for mini-batch learning device, and image processing device
Publication Date: 2026.04.01 FUJIFILM CORP
  • EP3896647B1 patent drawingFigure 1
  • EP3896647B1 patent drawingFigure 2
  • EP3896647B1 patent drawingFigure 3(A)~3(B)

AI summary

There are provided a mini-batch learning apparatus capable of preventing a decrease in a class determination accuracy of a machine learning model for performing semantic segmentation, an operation program of the mini-batch learning apparatus, an operation method of the mini-batch learning apparatus, and an image processing apparatus. In a case where the operation program is started, a CPU of the mini-batch learning apparatus functions as a calculation unit, a specifying unit, and an update unit. The calculation unit calculates an area ratio of each of a plurality of classes in mini-batch data. The specifying unit specifies a rare class of which the area ratio is lower than a setting value. The update unit sets an update level of the machine learning model in a case where the rare class is specified by the specifying unit to be lower than an update level of the machine learning model in a case where the rare class is not specified by the specifying unit.