Single-Channel Image Inversion for Labeled Data Augmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The shortage of labeled training datasets poses a challenge in machine learning, particularly due to the complexity of problems to be solved, necessitating data augmentation to generate diverse datasets.

Innovation Solution

A data processing apparatus and method that generates a labeled training dataset through data augmentation by inverting pixel values of single-channel images and corresponding labels, using reversible transformations to maintain the physical quantity representation, ensuring effective learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If data augmentation is performed on existing labeled training datasets to generate diverse datasets, then the quantity of training data is increased, but the quality and effectiveness of learning may deteriorate due to generation of inappropriate augmented data

Engineering Contradiction:
Improvequantity of training dataVSAvoideffectiveness of learning
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent applies parameter changes by inverting pixel values of single-channel images through reversible transformations. This transforms the physical quantity representation (pixel values corresponding to light reception amounts) while maintaining the validity of the data for training purposes, thereby generating diverse yet effective training data.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent converts potentially harmful or meaningless data generation into beneficial training data by inverting pixel values. Instead of generating random or irrelevant data, the inversion process creates valid training examples that preserve the relationship between pixel values and physical quantities, turning a simple transformation into a useful augmentation technique.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

2Adaptability or versatility

If conventional data augmentation methods are used without considering physical quantity representation, then data diversity is increased, but the physical meaning and interpretability of the data are lost

Engineering Contradiction:
Improvedata diversityVSAvoidphysical quantity representation
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent performs parameter changes through reversible transformations of pixel values that maintain the physical quantity representation. By inverting pixel values in a controlled manner, the patent diversifies the data while preserving the relationship between pixel values and underlying physical quantities such as light reception amounts.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent creates copies of existing training data through inversion operations. These copied datasets maintain the essential physical meaning of the original data while providing diversity for training, effectively replicating the information structure with transformed values.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4538687B1Data processing device, data processing method, and program
Publication Date: 2026.04.22 FUJIFILM CORP
  • EP4538687B1 patent drawingFigure 1~2
  • EP4538687B1 patent drawingFigure 3
  • EP4538687B1 patent drawingFigure 4

AI summary

There are provided a data processing apparatus, a data processing method, and a program for generating, by data augmentation, a labeled training dataset with which a learning model can perform effective learning. A data processing apparatus (10) is a data processing apparatus (10) including a processor (14). The processor (14) generates an inverted image obtained by inverting pixel values of a single-channel image that constitutes a first labeled training dataset, as training data of a second labeled training dataset obtained by augmentation of the first labeled training dataset. The single-channel image is an image of which the pixel values are determined in accordance with a physical quantity sensed by light-receiving elements during imaging or an image of which the pixel values are determined by reversible transform on the physical quantity. The processor (14) generates an inverted label corresponding to the inverted image, as a ground truth label that constitutes the second labeled training dataset. The inverted label is generated based on a ground truth label of the single-channel image. The ground truth label of the single-channel image constitutes the first labeled training dataset.