Vision Transformer Segmentation for Low-Contrast Material Phases

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing computer vision techniques for semantic segmentation produce inconsistent and inaccurate results, particularly with porous, amorphous, and low-contrast materials like carbon black-binder (CBD), and often require large amounts of annotated training data.

Innovation Solution

Employ deep learning techniques using transformer models, such as Segment Anything Model (SAM) and U-Net architectures, adapted for specific materials, and incorporate user prompts to enhance segmentation accuracy and efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional computer vision techniques are used for semantic segmentation, then the method is simple and easy to implement, but the segmentation accuracy and consistency deteriorate, especially for porous, amorphous, and low-contrast materials

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the segmentation task into multiple specialized transformer models, each dedicated to identifying a specific material phase. This division allows each model to be optimized for its specific material characteristics, improving overall segmentation accuracy for complex multiphase materials while maintaining manageable complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs parameter changes by adapting transformer models through fine-tuning with domain-specific training data. By adjusting model parameters through training on labeled images of specific materials (such as carbon black-binder, graphite, silicon, and aluminum foil), the models achieve improved accuracy for segmenting challenging porous and low-contrast materials.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If conventional semantic segmentation methods are used, then the computational resources required are minimal, but the segmentation performance on porous and low-contrast materials deteriorates

Engineering Contradiction:
Improvesegmentation consistencyVSAvoidcomputational resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent performs preliminary action by pre-training and fine-tuning transformer models on labeled training data before applying them to new segmentation tasks. This preliminary training phase enables the models to learn characteristic features of different materials, ensuring consistent and reliable segmentation performance when applied to new images, even for challenging porous and low-contrast materials.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by creating synthetic training data and using pre-trained models as foundations. The Segment Anything Model (SAM) serves as a base that is copied and adapted through fine-tuning on domain-specific data, allowing the system to leverage existing computational investments while achieving improved performance on specific material segmentation tasks.

Inventive Principle:
Principle #26Copying

3Adaptability or versatility

If a single transformer model is used for all materials, then the model complexity is low, but the ability to accurately segment multiple specific materials deteriorates

Engineering Contradiction:
Improvematerial-specific segmentation capabilityVSAvoidnumber of transformer models
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the segmentation task into multiple specialized transformer models, each dedicated to identifying a specific material phase. This division allows each model to be optimized for its specific material characteristics, improving overall segmentation accuracy for complex multiphase materials while maintaining manageable complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent achieves universality by using transformer models that can be fine-tuned for different materials. The same transformer architecture can be adapted to segment multiple different materials (carbon black-binder, graphite, silicon, aluminum foil) by training on domain-specific data, allowing one model family to serve multiple functions across different material types.

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

Data Source

PatentEP4648021A1Systems and methods for determining semantic segmentation of real-world objects
Publication Date: 2025.11.12 DASSAULT SYSTEMS AMERICAS CORP
  • EP4648021A1 patent drawingFigure 1
  • EP4648021A1 patent drawingFigure 2
  • EP4648021A1 patent drawingFigure 3

AI summary

Embodiments determine semantic segmentation of real-world objects. One such embodiment obtains an image of a real-world object. The obtained image is processed using a plurality of vision transformer (ViT) models to generate a plurality of segmentation masks. Each ViT model is configured to output a respective segmentation mask corresponding to a respective target material. Using a neural network combiner model, a multiphase semantic segmentation mask is generated based on the plurality of segmentation masks. The neural network combiner model is trained to integrate outputs of the plurality of ViT models. Semantic segmentation of the real-world object is thereby determined.