NCI Encoder Training for Water Segmentation Generalization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current water segmentation models face challenges in generalizing to Out-of-Distribution (OoD) scenarios due to data scarcity, particularly in flood monitoring where real data may not adequately represent all forms of water appearances, and existing techniques for domain generalization are not effectively applied to downstream OoD segmentation tasks.

Innovation Solution

A computer-implemented method for training a machine-learning model that combines real and augmented training data, using a Non-Commutatively Invariant (NCI) encoder to map both real and synthetic data to a shared representation space, allowing the model to leverage synthetic data for improved generalization and efficiency in image analysis tasks like water segmentation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If synthetic data is used to augment training data, then the quantity of training data increases and model generalization improves, but the modality gap between synthetic and real data creates distribution shift that reduces reliability

Engineering Contradiction:
Improvequantity of training dataVSAvoidmodel generalization reliability
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent introduces a domain adapter as an intermediary component between the encoder and downstream tasks. This adapter learns to translate synthetic data representations into a domain-invariant space that is compatible with real data distributions, thereby bridging the modality gap without requiring direct mixing of synthetic and real data in the same representation space

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the training pipeline into separate components: a shared encoder that processes both synthetic and real data, and a domain-specific adapter for each downstream task. This segmentation allows each component to specialize - the encoder learns universal features while adapters handle domain-specific transformations, reducing the harmful effects of distribution shift

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If domain generalization techniques are applied to downstream OoD segmentation tasks, then adaptability to OoD scenarios improves, but existing techniques have not been effectively applied to this specific task

Engineering Contradiction:
Improveadaptability to OoD scenariosVSAvoidease of application to downstream tasks
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The patent creates a universal domain adapter framework that can be applied across multiple downstream segmentation tasks. The adapter is trained in a task-agnostic manner using domain generalization objectives, making it universally applicable to various OoD segmentation problems without requiring task-specific retraining or modification

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If real data is collected to represent all water forms, then measurement precision improves, but data collection becomes infeasible due to the vast variety of water appearances

Engineering Contradiction:
Improveprecision of water representationVSAvoiddata collection feasibility
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent uses synthetic data as a copy or simulation of real water appearances. Instead of collecting all possible real water images, the system generates synthetic images that replicate the visual characteristics of water under various conditions (different colors, textures, lighting, depths). These synthetic copies are then used to train the encoder to recognize water across diverse scenarios

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4528541A1A computer-implemented method, data processing apparatus, and computer program for image analysis
Publication Date: 2025.03.26 FUJITSU LTD
  • EP4528541A1 patent drawingFigure 1A~1B
  • EP4528541A1 patent drawingFigure 2
  • EP4528541A1 patent drawingFigure 3

AI summary

A computer-implemented method of training a machine-learning model for image analysis, the method comprising receiving an input training dataset comprising real training data and augmented training data; iteratively training an encoder of the machine-learning model to obtain a trained encoder, the iterative training comprising: training a discriminator model by minimising a discriminator loss function using the input training dataset to obtain a trained discriminator, the trained discriminator being configured to discriminate whether input data is real data or augmented data; and training the encoder by maximising a discriminator error using the training dataset and the trained discriminator to obtain a trained encoder, the trained encoder being configured to invariantly encode real data and augmented data to a same representation space.