image restoration using graphical models anna grim ...ising model murphyk/papers/intro_gm.pdf local...

19
Image Restoration using Graphical Models Margaret Thoren Anna Grim

Upload: others

Post on 31-Mar-2021

9 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Image Restoration usingGraphical Models

Margaret ThorenAnna Grim

Page 2: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Problem: Image Restoration

https://www.smapip.is.tohoku.ac.jp/~kazu/SMAPIP-KazuKazu/Summary/index-e.html

Page 3: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Introduction

https://blog.statsbot.co/probabilistic-graphical-models-tutorial-and-solutions-e4f1d72af189

Def: A graphical model consists of a collection of random variables corresponding to some graph.

Example:(1) Random Variables

(2) Graph

Page 4: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Converting an Image into a Graph

Page 5: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Ising Model

https://www.cs.ubc.ca/~murphyk/Papers/intro_gm.pdf

Local Energy:

Pairwise Energy:

● We associate a random variable with each pixel;

Page 6: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Ising Model

● The joint pdf corresponding to the random variables is energy based and defined to be

● Maximizing the probability is equivalent to minimizing the energy

Page 7: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Overview

https://slideplayer.com/slide/8088915https://www.slideshare.net/zukun/02-probabilistic-inference-in-graphical-models

Page 8: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Gibbs Sampling

https://slideplayer.com/slide/8141891/

Main Idea: Restoring pixels to values close to original observation and close to neighboring observations

Page 9: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Formulas for Gibbs Sampling

Restore image by drawing samples from the following distribution

Goal:

● Using properties of Markov Random Fields, we can simplify this distribution such that

Page 10: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Gibbs Sampling Algorithm(1) Initialization

(2) Sampling

(3) Check for Convergence

Page 11: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Coding in MATLAB

Page 12: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Belief Propagation

Main Idea: Message passing algorithm that approximates the marginal probability distribution where pixels take on values based on messages they receive from neighbors

http://ssg.mit.edu/nbp/

Page 13: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Formulas for Belief Propagation

Messages from node i to node j:

Belief at node j:

Page 14: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Intuition of Belief Propagation Formulas

Page 15: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Belief Propagation Algorithm(1) Initialization

(2) Update Messages

(3) Check for convergence

Page 16: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Messages to Restoration

•Once the messages have converged, we restore the

image by computing

Page 17: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Convergence

https://en.wikipedia.org/wiki/Tree_(graph_theory)

https://slideplayer.com/slide/5090395/

Page 18: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Belief Propagation Results

Page 19: Image Restoration using Graphical Models Anna Grim ...Ising Model murphyk/Papers/intro_gm.pdf Local Energy: Pairwise Energy: We associate a random variable with each pixel;Ising Model

Conclusions

Belief Propagation Gibbs Sampling