Graph-Driven Diffusion and Random Walk Schemes for Image Segmentation.

Area of Science:

• Computer Vision
• Graph Theory
• Image Processing

Background:

• Image segmentation is a fundamental task in computer vision.
• Existing methods often rely on grid-based assumptions, limiting flexibility.
• Graph-based approaches offer a powerful framework for modeling complex relationships in data.

Purpose of the Study:

• To introduce novel graph-driven diffusion processes for image segmentation.
• To develop new algorithms inspired by epidemic propagation models.
• To enhance existing graph-based segmentation methods with degree-aware properties.

Main Methods:

• Formulating image segmentation as infectious wavefront propagation on an image-driven graph.
• Relating the Susceptible-Infected-Recovered (SIR) model to the Random Walker algorithm.
• Developing Normalized Random Walker and lazy random walker variants with a degree-aware term.

Main Results:

• Demonstrated reliability and computational efficiency of the proposed graph-driven approaches.
• Showcased dimensionality reduction capabilities.
• Validated the approach on pixel-level, node-level, and multidimensional data without grid constraints.

Conclusions:

• Graph-driven diffusion processes provide a flexible and effective framework for image segmentation.
• The degree-aware term enhances segmentation by considering node centrality.
• The proposed methods offer a robust alternative to traditional grid-based segmentation techniques.