Abstract
This article addresses qualitative descriptions of image which characterise hierarchy and configuration of image structure. The qualitative descriptions are derived from trajectories of stationary points in the Gaussian scale-space and the gradient field of the scale-space image, which are called respectively “stationary curves” and “figure field”. The stationary curves and figure-field fluxes define the scale-space hierarchy, which is explicitly described as a single tree by introducing a point at infinity. The configuration of image structure at fixed scale is represented by a pseudograph which is obtained from the figure field. The Voronoi tessellation is also employed to extract boundaries of image segments.
Introduction
In computer vision, understanding the "structure" of an image is difficult because features exist at multiple scales—from fine textures to large anatomical shapes. While we can blur an image to remove noise using Gaussian scale-space, a formal method is needed to describe how features at a fine scale relate to those at a coarse scale. Existing methods often struggled to create a unified, mathematically consistent "tree" of these features. Specifically, when features disappear (annihilate) as the image gets blurrier, it wasn't always clear how to link them to a "parent" feature. This paper provides a complete qualitative description that captures both the hierarchy (how features merge) and the configuration (how they are arranged spatially).
Methodology
The authors leverage several key mathematical concepts to build this framework:
Gaussian Scale-Space and Stationary Curves
The image is progressively blurred using the linear diffusion equation, creating a 3D volume (2D space + 1D scale). Within this volume, the authors track "stationary points"—locations where the gradient is zero, including local maxima (peaks), local minima (valleys), and saddle points (passes). The trajectories of these points as scale increases are called stationary curves.
The Figure Field and Pseudographs
The "figure field" is defined as the negative gradient of the image, representing a fluid-like flow from peaks to valleys. At any fixed level of blur, the flow of the figure field is used to construct a pseudograph. In this graph, vertices represent maxima, and edges are flow lines passing through saddle points. This captures the spatial configuration of features.
The Point at Infinity
A novel contribution is treating the dark background of an image as a "local minimum at infinity." This ensures that every feature can be connected into a single, rooted hierarchical tree, solving a major consistency issue in scale-space theory where features might otherwise seem to disappear without a clear parent.
Key Insights
- Topological Consistency: The "point at infinity" ensures the Euler characteristic of the image graph remains constant, making the mathematical description robust.
- Saddles as Connectors: Saddle points act as the "glue" in the figure field, defining the boundaries and relationships between different peaks.
- Conservation of Intensity: The figure field follows a conservation law, allowing image structure to be analyzed like a physical flow system.
- Weighted Voronoi Tessellation: Image segmentation is achieved by assigning regions to specific maxima, weighted by the scale at which those features eventually disappear.
Results
The theory was demonstrated using synthetic blobs and complex BrainWeb MRI slices.
- Hierarchical Mapping: Complex MRI slices were transformed into clean, branching trees showing how small brain structures merge into larger anatomical regions as scale increases.
- Structural Graphs: As an image blurs, its corresponding pseudograph simplifies predictably, eventually collapsing toward a single point.
- Effective Segmentation: The Voronoi tessellation based on annihilation scales produced segmentations matching human intuition of "dominant" brain parts, outperforming simple gradient-based edge detection in blurred images.
Conclusion
By combining stationary curve tracking with the figure field and the concept of a point at infinity, this paper establishes a robust topological framework for multi-scale image analysis. This approach provides a mathematically consistent way to describe both the hierarchy and spatial arrangement of image features across scales.
Summary updated