The Extrema Edges

Pablo Andrés Arbeláez, Laurent D. Cohen

TL;DR

This paper introduces a new method for detecting edges in images by modeling them as discontinuities in a "mosaic" image derived from the image's extrema. This approach leverages minimal path theory and a novel "linear variation" energy to create a robust edge representation that preserves semantically important features and reduces noise.

What Problem Does This Paper Address?

Detecting and accurately representing edges in images is a fundamental challenge in computer vision. Traditional methods often rely on local image information and differential operators, which can be sensitive to noise and may struggle to capture the global structure of edges. While watershed transforms offer closed contours, they are typically applied to the gradient magnitude, which can be further degraded by noise and smoothing. This paper aims to address these limitations by proposing a new edge detection framework that is more robust to noise and better preserves the semantic content of edges.

Methods

The proposed method consists of two main parts: edge localization and edge valuation.

  1. Edge Localization (Extrema Partition):
    • The core idea is to model edges as discontinuities in a "mosaic" image. This mosaic is derived from an "extrema partition" of the image.
  2. The extrema partition is based on the theory of minimal paths and a novel energy functional called "linear variation."
    • Linear Variation: This is a generalization of the one-dimensional total variation. It measures the minimal variation of a function along paths connecting two points. For images, it's defined as the infimum of the variation of the image along all possible paths between two points.
  3. Extrema as Sources: The method identifies image extrema (local maxima and minima) as "sources." The linear variation energy is then used to partition the image domain into "influence zones" around these extrema. The "extrema partition" is formed by these influence zones and the boundaries between them.
  4. Extrema Mosaic: The extrema partition is then used to create an "extrema mosaic" (EM). This is a piecewise constant approximation of the image where each region (influence zone) is assigned a constant value, typically the image intensity at the extremum source. Edges are then defined as the discontinuities of this extrema mosaic.
  5. Edge Valuation (Contrast-Driven Hierarchy):
    • To assign a "saliency" or importance to the detected edges, the paper proposes a contrast-driven hierarchy of partitions.
  6. This involves building a Region Adjacency Graph (RAG) from an initial partition (e.g., the extrema mosaic).
  7. A dissimilarity measure, based on boundary information and contrast between adjacent regions, is used to define a merging strategy.
  8. This process generates a family of nested partitions, creating a hierarchy. The "saliency" of a pixel is defined as the highest level in this hierarchy for which the pixel belongs to a contour. This provides a measure of how significant an edge is based on the contrast it represents.

Datasets

The paper uses several test images for demonstration and evaluation, including:

  • A synthetic Gaussian blob.
  • A cameraman image.
  • A medical image (retinal fundus image) for demonstrating application in medical image analysis.

No new benchmarks or datasets were introduced.

Key Insights and Ideas

  • Edges as Mosaic Discontinuities: The fundamental insight is to redefine edges not as local gradient features but as discontinuities in a simplified, piecewise constant representation (the extrema mosaic) of the image. This shifts the focus from local derivatives to global structural properties.
  • Linear Variation as a Robust Energy: The introduction of the "linear variation" as an energy functional provides a way to incorporate global path information into edge detection, making it less susceptible to noise compared to purely local gradient-based methods.
  • Extrema as Semantically Meaningful Sources: Using image extrema as sources for the energy partition is motivated by their potential to represent semantically important features and atoms of the image.
  • Contrast-Driven Saliency: The hierarchy of partitions based on contrast provides a principled way to assign importance to edges, allowing for a more nuanced understanding of edge significance beyond simple intensity changes.
  • Preservation of Semantic Features: The method is shown to preserve semantically important characteristics like corners and junctions, which are often lost or degraded by traditional methods.

Key Results

  • The proposed method produces a set of closed contours that represent edges.
  • The extrema mosaic provides a piecewise constant approximation of the image that can reduce noise and enhance contrast.
  • Qualitative results on test images demonstrate that the extrema edges accurately capture contour information and semantically important features like corners and junctions.
  • The method is shown to be effective in reducing the blur of the original image by absorbing non-extremal components into neighboring regions.
  • In medical imaging, the approach successfully detected drusen in retinal fundus images, highlighting its potential for challenging applications.

Limitations and Future Work

  • Quantification and Subsampling: The paper notes that quantification and subsampling can lead to the loss of important parts of the isoenergy set, which can fragment contours.
  • Color Images: The current work focuses on monochrome images. The authors suggest that the approach can be extended to color images by considering the brightness component, and this generalization will be the subject of future work.
  • Further Applications: The authors intend to explore further applications of their approach, such as tracking the evolution of diseases in medical images over time.

Summary updated

Abstract

We present a new approach to model edges in monochrome images. The method is divided in two parts: the localization of possible edge points and their valuation. The first part is based on the theory of minimal paths, where the selection of an energy and a set of sources determines a partition of the domain. Then, the valuation is obtained by the creation of a contrast driven hierarchy of partitions. The method uses only the original image and supplies a set of closed contours that preserve semantically important characteristics of edges.
BibTeX
@inproceedings{arbelaez2003extrema,
  author = {Pablo Andrés Arbeláez and Laurent D. Cohen},
  title = {{The Extrema Edges}},
  booktitle = {Scale-space},
  year = {2003},
  doi = {10.1007/3-540-44935-3\_13},
}
Figure