[go: up one dir, main page]
Scribd
Upload
106 views9 pages

Convex Optimization in Image Processing: Ernie Esser

This document discusses convex optimization in image processing. It describes how inverse problems in imaging can be formulated as variational models that define an objective function to minimize. This function incorporates data fidelity and regularization terms. Optimization techniques like iterative methods can be used to find the global minimum when the objective function is convex. While nonconvex problems are harder, sometimes they can be approximated by convex models to leverage efficient convex optimization algorithms. Overall, the document explains how modeling imaging problems as convex optimization allows them to be reliably solved.

Uploaded by

M Rameez Ur Rehman
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
106 views9 pages

Convex Optimization in Image Processing: Ernie Esser

This document discusses convex optimization in image processing. It describes how inverse problems in imaging can be formulated as variational models that define an objective function to minimize. This function incorporates data fidelity and regularization terms. Optimization techniques like iterative methods can be used to find the global minimum when the objective function is convex. While nonconvex problems are harder, sometimes they can be approximated by convex models to leverage efficient convex optimization algorithms. Overall, the document explains how modeling imaging problems as convex optimization allows them to be reliably solved.

Uploaded by

M Rameez Ur Rehman
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Convex Optimization in Image

Processing
Ernie Esser

5-5-2010

1
Inverse Problems

Let f be some given measurements of an image h

Goal: Find h given f and knowledge of the kinds of measurements taken

Some Examples:
Denoising: f = h + noise
Deblurring: f = blurred h + noise
Super Resolution: f = low resolution h + noise

2
Variational Models

Model the problem by defining a function F (u) on images u such that


F (u) is large when u is a poor solution
F (u) is small when u is a good solution

Data Fidelity: F (u) should be smaller when u is consistent with the


measurements f

Regularization: Data fidelity is usually not enough by itself. To make the


problem well posed, add an assumption about u:
smooth u
piecewise constant u
sparse u, etc...
F (u) should be smaller when the assumption is better satisfied

3
Optimization
Represent images as real M N matrices or as vectors in RM N .

Find u RM N that minimizes F (u)

Calculus Approach: Solve F (u ) = 0

Iterative Approach: Find uk u k = 1, 2, 3, ...

Issues:
linear versus nonlinear
differentiability
number of variables
constraints on u
local versus global minima
convex versus nonconvex

4
Convexity

In fact the great watershed in optimization isnt between linearity and


nonlinearity, but convexity and nonconvexity.
- R. T. Rockafellar

F ((1 s)u1 + su2 ) (1 s)F (u1 ) + sF (u2 ) for all u1 , u2


0s1

F(u2)
F(u1)

5
Local Min versus Global Min
If F is convex, Local Minimum Global Minimum

Smooth, Nonconvex Convex, Nonsmooth

u 0 u1 u 2 u3 u 0 u1 u2u3
Steepest descent, Proximal point method
"go downhill", finds global min
finds local min

Proximal point method diagram from Bertsekas and Tsitsiklis

6
Solving Convex Problems
There are efficient algorithms for convex optimization
Image processing problems modeled as convex optimization problems
can be reliably solved

Deblurring Example: F (u) was defined to be a convex function that


encourages data fidelity and prefers piecewise constant u

Original image Blurry/Noisy Recovered


(This was solved using an iterative method that is a generalization of the proximal point method.)

7
Nonconvex Problems
Nonconvex problems are much harder in general

Example: Blind Deblurring (dont know how image was blurred)


Example: Registration
Z
Want to align h1 and h2 F (v) = (h1 (x+v(x))h2 (x))2 dx

1.5
h1(x) 2500
h2(x)

2000

1500
F()

1000

0.5

500

0 0
0 10 20 30 40 50 60 70 80 90 100
0 10 20 30 40 50 60 70 80 90 100

x

h1 (x) and h2 (x) F (v) in the case when v(x) =


(translation)
8
Convex Approximation
Approaches for nonconvex problems often require convex optimization
for subproblems
Sometimes can approximate a nonconvex model by a convex one

Convex image registration example:


h h
2 1
0 0
10
10
20
30
20
40

30 50
60
40 70

0 10 20 30 0 20 40 60 80

h1(x+v(x)) displacement v
0 0
10
10
20
30
20
40

30 50
60
40 70

0 10 20 30 0 20 40 60 80

The good: based on convex model, so can find global minimum


The bad: slow to compute minimizer due to many extra variables
(had to essentially double dimensionality of problem to make it convex)
9

You might also like