Wasserstein barycenters and applications in image processing: Difference between revisions

From Optimal Transport Wiki
Jump to navigation Jump to search
No edit summary
No edit summary
 
(One intermediate revision by the same user not shown)
Line 11: Line 11:
<math>\sum_{i \in I} \lambda_i W_2( \mu_i, \mu)^2</math> over the space <math>\mu \in \mathcal{P}(\Omega)</math>. Here <math>W_2</math> denotes the 2-Wasserstein distance, which may be replaced with the <math>p</math>-Wasserstein distance, <math>W_p</math>, though this is not always as convenient.  
<math>\sum_{i \in I} \lambda_i W_2( \mu_i, \mu)^2</math> over the space <math>\mu \in \mathcal{P}(\Omega)</math>. Here <math>W_2</math> denotes the 2-Wasserstein distance, which may be replaced with the <math>p</math>-Wasserstein distance, <math>W_p</math>, though this is not always as convenient.  


The minimization problem above was originally introduced by Agueh and Carlier <ref>Martial Agueh, Guillaume Carlier. Barycenters in the Wasserstein space. SIAM Journal on
This minimization problem was originally introduced by Agueh and Carlier <ref>Martial Agueh, Guillaume Carlier. Barycenters in the Wasserstein space. SIAM Journal on
Mathematical Analysis, Society for Industrial and Applied Mathematics, 2011, 43 (2), pp.904-924.
Mathematical Analysis, Society for Industrial and Applied Mathematics, 2011, 43 (2), pp.904-924.
ff10.1137/100805741ff. ffhal-00637399f</ref>, who also proposed an alternative formulation of the problem above when there are finitely many measures <math>\{\mu_i\}_{i = 1}^N</math>. Instead of considering the sum over all probability measures <math>\mu \in \mathcal{P}(\Omega)</math>, one can equivalently consider the optimization problem over all multi-marginal transport plans <math>\gamma \in \mathcal{P}(\Omega^{N+1})</math> whose [https://en.wikipedia.org/wiki/Pushforward_measure push forwards] satisfy <math>(\pi_i)_{\#} \gamma = \mu_i</math> for <math>i \in \{1, \ldots, N\}</math> and <math>(\pi_0)_{\#} \gamma = \mu</math> for some unspecified probability measure <math>\mu</math>.
ff10.1137/100805741ff. ffhal-00637399f</ref>, who also proposed an alternative formulation of the problem above when there are finitely many measures <math>\{\mu_i\}_{i = 1}^N</math>. Instead of considering the minimum over all probability measures <math>\mu \in \mathcal{P}(\Omega)</math>, one can equivalently consider the optimization problem over all multi-marginal transport plans <math>\gamma \in \mathcal{P}(\Omega^{N+1})</math> whose [https://en.wikipedia.org/wiki/Pushforward_measure push forwards] satisfy <math>(\pi_i)_{\#} \gamma = \mu_i</math> for <math>i \in \{1, \ldots, N\}</math> and <math>(\pi_0)_{\#} \gamma = \mu</math> for some unspecified probability measure <math>\mu</math>. The problem then tries to minimize
 
<math>\int \left ( \sum_{i = 1}^N \lambda_i \| x_i - x_0 \|^2 \right ) d \gamma</math>


===Existence and Uniqueness===
===Existence and Uniqueness===
Line 23: Line 25:
In the case where all of the measures <math>\mu_i</math> are finitely supported measures, computing the Wasserstein barycenter reduces to a problem in [https://en.wikipedia.org/wiki/Linear_programming linear programming].
In the case where all of the measures <math>\mu_i</math> are finitely supported measures, computing the Wasserstein barycenter reduces to a problem in [https://en.wikipedia.org/wiki/Linear_programming linear programming].


When our family of measures <math>\{\mu_i\}_{i \in I} </math> consists of only one finitely supported measure <math>\mu_1</math> and we restrict the input probability measure <math>\mu</math> to finitely supported probability measures whose supports have at most <math>k</math> points, the problem of finding the Wasserstein barycenter is equivalent to the [https://en.wikipedia.org/wiki/K-means_clustering <math>k</math>-means] clustering problem.
When the family of measures <math>\{\mu_i\}_{i \in I} </math> consists of only one finitely supported measure <math>\mu_1</math> and we restrict the input probability measure <math>\mu</math> to finitely supported probability measures whose supports have at most <math>k</math> points, the problem of finding the Wasserstein barycenter is equivalent to the [https://en.wikipedia.org/wiki/K-means_clustering <math>k</math>-means] clustering problem.


==Applications==
==Applications==
===Barycenters in Image processing===
===Barycenters in Image Processing===
Barycenters have several applications in [https://en.wikipedia.org/wiki/Digital_image_processing image processing]. A basic example<ref>Claici, Sebastian. “Stochastic Wasserstein Barycenters.” Lecture, MIT, July, 2018.</ref> arises in [https://en.wikipedia.org/wiki/Handwriting_recognition handwriting recognition]. One may wish to determine what an expected letter looks like, but simply taking the pixel-by-pixel Euclidean average won't capture the geometric information of the letter. Instead, a <math>k</math>-means algorithm can compute what a prototypical version of that letter looks like.


As another application, barycenters can be used to interpolate between two or more images<ref>Santambrogio, Filippo. Optimal Transport for Applied Mathematicians. Birkhäuser., 2015.</ref>. For example, suppose we are given two images of similar objects, one with a darker gray-scale histogram <math>\mu_1</math> and the other with a lighter one <math>\mu_2</math>. One way to find the middle point of their histograms would be to take their Euclidean average <math>(\mu_1 + \mu_2)/2</math>, but this would result in a histogram with a distribution of some dark and some light shades as opposed to a gray-scaled version. The Wasserstein barycenter between these two histograms provides a gray image. 
Similar ideas can be used to enhance the contrast of an image.


==Generalizations==
==Generalizations==

Latest revision as of 07:14, 12 February 2022

In optimal transport, a Wasserstein barycenter [1] is a probability measure that acts as a center of mass between two or more probability measures. It generalizes the notions of physical barycenters and geometric centroids.


Introduction

Motivation

Barycenters in physics and geometry are points that represent a notion of a mean of a number of objects. In astronomy, the barycenter is the center of mass of two or more objects that orbit each other, and in geometry, a centroid is the arithmetic mean of all the points in an object. Given countably many points in with nonnegative weights , the weighted barycenter of the points is the unique point minimizing . Wasserstein barycenters attempt to capture this concept for probability measures by replacing the Euclidean distance with the Wasserstein distance of two probability measures, .

Definition

Let be a domain and  be the set of probability measures on . Given a collection of probability measures and nonnegative weights , we define a weighted barycenter of as any probability measure  that minimizes  over the space . Here denotes the 2-Wasserstein distance, which may be replaced with the -Wasserstein distance, , though this is not always as convenient.

This minimization problem was originally introduced by Agueh and Carlier [2], who also proposed an alternative formulation of the problem above when there are finitely many measures . Instead of considering the minimum over all probability measures , one can equivalently consider the optimization problem over all multi-marginal transport plans whose push forwards satisfy for and for some unspecified probability measure . The problem then tries to minimize

Existence and Uniqueness

Agueh and Carlier [3] demonstrated that when at least one of the measures is absolutely continuous, the Wasserstein barycenter exists and is unique.

Examples[4]

In the case where all of the measures are finitely supported measures, computing the Wasserstein barycenter reduces to a problem in linear programming.

When the family of measures consists of only one finitely supported measure and we restrict the input probability measure to finitely supported probability measures whose supports have at most points, the problem of finding the Wasserstein barycenter is equivalent to the -means clustering problem.

Applications

Barycenters in Image Processing

Barycenters have several applications in image processing. A basic example[5] arises in handwriting recognition. One may wish to determine what an expected letter looks like, but simply taking the pixel-by-pixel Euclidean average won't capture the geometric information of the letter. Instead, a -means algorithm can compute what a prototypical version of that letter looks like.

As another application, barycenters can be used to interpolate between two or more images[6]. For example, suppose we are given two images of similar objects, one with a darker gray-scale histogram and the other with a lighter one . One way to find the middle point of their histograms would be to take their Euclidean average , but this would result in a histogram with a distribution of some dark and some light shades as opposed to a gray-scaled version. The Wasserstein barycenter between these two histograms provides a gray image. Similar ideas can be used to enhance the contrast of an image.

Generalizations

Wasserstein barycenters are examples of Karcher and Fréchet means where the distance function used is the Wasserstein distance.

References

  1. Santambrogio, Filippo. Optimal Transport for Applied Mathematicians. Birkhäuser., 2015.
  2. Martial Agueh, Guillaume Carlier. Barycenters in the Wasserstein space. SIAM Journal on Mathematical Analysis, Society for Industrial and Applied Mathematics, 2011, 43 (2), pp.904-924. ff10.1137/100805741ff. ffhal-00637399f
  3. Martial Agueh, Guillaume Carlier. Barycenters in the Wasserstein space. SIAM Journal on Mathematical Analysis, Society for Industrial and Applied Mathematics, 2011, 43 (2), pp.904-924. ff10.1137/100805741ff. ffhal-00637399f
  4. Peyré, Gabriel and Marco Cuturi. “Computational Optimal Transport.” Found. Trends Mach. Learn. 11 (2019): 355-607.
  5. Claici, Sebastian. “Stochastic Wasserstein Barycenters.” Lecture, MIT, July, 2018.
  6. Santambrogio, Filippo. Optimal Transport for Applied Mathematicians. Birkhäuser., 2015.