Semi-automatic Image Segmentation via Multi-scale Saliency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image segmentation methods require significant user input and are inefficient due to high computational complexity and robustness issues, particularly in semi-automatic color image segmentation.

Innovation Solution

A semi-automatic image segmentation system using saliency detection by color histogram contrast, which employs the expectation-maximization algorithm for unsupervised classification and a multi-scale scheme to estimate object masks, reducing user inputs and improving stability and computational efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If grab-cut method is used for color image segmentation, then segmentation accuracy is improved by incorporating smoothness term and local gradient analysis, but computational complexity increases and robustness deteriorates

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

Solution Approach 1:

The patent divides the image into multiple scales using a multi-scale scheme, processing the image at different resolution levels. This segmentation approach allows the algorithm to capture objects at various sizes while reducing computational burden at each scale compared to processing the full-resolution image directly with grab-cut.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter of image resolution by processing images at multiple scales rather than a single fixed resolution. This allows the system to adaptively select appropriate processing detail levels, reducing computational complexity while maintaining segmentation accuracy for objects of different sizes.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If grab-cut method is used for color image segmentation, then segmentation accuracy is improved by incorporating smoothness term, but robustness deteriorates

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidrobustness
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

By segmenting the processing task across multiple scales, the patent reduces the impact of local variations and noise that can cause grab-cut to fail. The multi-scale approach provides redundancy and stability, making the segmentation more robust to different image conditions while maintaining accuracy.

Inventive Principle:
Principle #1Segmentation

3Device complexity

If traditional thresholding method is used for segmentation, then computational simplicity is maintained, but segmentation accuracy deteriorates for multi-colored objects

Engineering Contradiction:
Improvemethod simplicityVSAvoidsegmentation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies a simplified version of grab-cut at multiple scales rather than using complex thresholding. This maintains relative computational simplicity while significantly improving accuracy for multi-colored objects through the multi-scale processing framework.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies a partial version of the full grab-cut algorithm at multiple scales rather than implementing the complete complex algorithm once. This partial application across scales achieves better accuracy with controlled computational cost.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3332356B1Semi-automatic image segmentation
Publication Date: 2021.04.21 SONY GROUP CORP
  • EP3332356B1 patent drawingFigure 1A~1B
  • EP3332356B1 patent drawingFigure 2A
  • EP3332356B1 patent drawingFigure 2B

AI summary

Semi-automatic (instant) color image segmentation is performed on an input image for an object located near the point given by the user (e.g., touch screen contact). The size, shape, and the number of colors of the target object which is about (lying around or surrounding) the user-selected point is determined, and determinations made for which region the pixels belong to, either foreground or background. A binary object mask is generated which can be utilized for removing the material surrounding that object, or inverting the mask to remove the material of the object.