Multi-Resolution Image Segmentation for High-Frequency Artifact Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing binary segmentation methods, particularly those based on deep learning and example-based approaches, are computationally expensive and produce low-quality results with high frequency artifacts, especially for 3D images, due to their O(n^2) complexity and reliance on large initialization datasets.

Innovation Solution

A multi-resolution iterative method that segments images by downsampling the original image and bimap to a lowest resolution, iteratively upsampling and segmenting at higher resolutions, using interactive binary segmentation algorithms to reduce computational complexity and improve quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If example-based binary segmentation methods are used, then segmentation quality improves, but computational complexity increases to O(n^2)

Engineering Contradiction:
Improvesegmentation qualityVSAvoidcomputational complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies multi-resolution segmentation by dividing the segmentation process into multiple resolution levels. The image is processed at progressively finer resolutions, with each level building upon the previous one. This divides the complex O(n^2) problem into smaller sub-problems that are computationally more manageable while maintaining segmentation quality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a resolution dimension to the segmentation process. Instead of performing segmentation at a single resolution, it adds the dimension of progressive resolution refinement, transforming the problem from a single-step complex operation into a multi-step process that leverages information from multiple scales.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Extent of automation

If deep learning techniques are used for segmentation, then automation improves, but computational cost increases

Engineering Contradiction:
ImproveautomationVSAvoidcomputational cost
Core Design Contradiction:
Extent of automationVSUse of energy by moving object

Solution Approach 1:

The patent segments the computational workload across multiple resolution levels, performing coarser segmentation first and refining progressively. This divides the heavy computational burden of deep learning into manageable stages, reducing peak computational cost while maintaining automation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary segmentation at lower resolutions before refining at higher resolutions. This preliminary action at coarser scales provides a foundation that reduces the computational burden of subsequent finer-scale segmentation, effectively pre-processing the problem to reduce overall computational cost.

Inventive Principle:
Principle #10Preliminary action

3Manufacturing precision

If high resolution segmentation is performed directly, then segmentation precision improves, but high frequency artifacts increase

Engineering Contradiction:
Improvesegmentation precisionVSAvoidhigh frequency artifacts
Core Design Contradiction:
Manufacturing precisionVSObject-generated harmful factors

Solution Approach 1:

The patent segments the frequency spectrum by processing at multiple resolution levels. Coarser resolutions capture low-frequency structural information while finer resolutions add high-frequency details progressively, preventing the accumulation of high-frequency artifacts that would occur with direct high-resolution segmentation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary segmentation at lower resolutions to establish the basic structure before refining at higher resolutions. This preliminary action at coarser scales filters out high-frequency noise and artifacts, allowing subsequent refinement to focus only on genuine high-frequency details without introducing spurious artifacts.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4641494A1Segmenting of image
Publication Date: 2025.10.29 DASSAULT SYSTEMES SA
  • EP4641494A1 patent drawingFigure 1
  • EP4641494A1 patent drawingFigure 2
  • EP4641494A1 patent drawingFigure 3

AI summary

The disclosure notably relates to a computer-implemented method for segmenting an image. The method comprises obtaining (S10) an original image (I0) to be segmented and an original bimap (Bini). The method comprises computing 520 one or more images (∀ k ∈ 1, ... , n, Ik) each by downsampling the original image (I0). The method comprises segmenting (S40) the lowest resolution image (In) based on a bimap (Bn) of the image (In) having the lowest resolution. The method comprises, iteratively (S50) for each determined image, starting from the lowest resolution image (In) and increasing the resolution (∀ k ∈ 1, ... , n in descending order) determining (S51) an upsampled segmentation mask Skuphaving a next resolution, determining (S52) a respective bimap (Bk-1) of the image (Ik-1) having the next resolution and segmenting (S53) the next resolution image (Ik-1) based on the respective bimap (Bk-1). Such a method forms an improved solution for segmenting an image.