image registration - uppsala university · image similarity – area based methods several options...

Image Registration

Computer Assisted Image Analysis II

Nataša Sladoje, [email protected]

Registration – To Read?!

Unfortunately, the course book does not cover all aspects of image registration. You will find pieces of it, however.

Matching 6.4

ICP 12.2.6

Transformations 5.2, 11.2

Interest points 5.3.10, 5.3.11

Examples of good reading material on the topic include: Image Alignment and Stitching: A Tutorial

http://www.cs.washington.edu/education/courses/cse576/05sp/papers/MSR-TR-2004-92.pdf

Image Registration Methods: A Survey http://library.utia.cas.cz/prace/20030125.pdf

oExercises









http://library.utia.cas.cz/prace/20030125.pdf

http://library.utia.cas.cz/prace/20030125.pdf

Registration – What is it?

“register” “To adjust so as to be properly aligned.”

“fusion” “something new created by a mixture of qualities, ideas, or things”

“warping” “become or cause to become bent or twisted out of shape, typically as a result of the effects of heat or dampness”

“matching” “a person or thing that resembles or corresponds to another”

y = j(x)

X

Y

x1

x2

y1

y2

Registration – When is it used?

Art and Photography

Stitching photos together, panoramic images

Astronomy

Stitching photos together and fusion of different

wavelengths


Robotics

Finding and tracking objects in a scene (2-D)

Chemistry

Finding similar images of molecules


“Word Spotting”

Matching and aligning words in

old handwritten historical

documents

Enables searching in large

document collections without

the need for manual translation

F. Wahlberg et al. 2011


Medical Imaging

study skin changes over time

colonoscopy (1-D)

Pre operative vs post operative

Fusion of different modalities

CT-MRI

CT-Ultrasound

PET-CT

…

And much more …

Holmberg et al. 2008

Nain et al. 2002

Registration – An ill posed problem

An ill-posed problem: Many degrees of freedom compared to available data. Given a set of pixel intensities, find a vector (deformation) for each position – a

number of unknowns is greater than the number of constraints

Rotational symmetry of objects affects uniqueness of the solution.

Solution: Restrict the set of possible transformations φ

y = j(x)

X

Y

x1

x2

y1

y2

Image registration flow chart

Registration – How to do it?

Choose starting

parameters

Transform Study

Evaluate cost

function

Converged?

Choose new set of parameters

Reference Study

No Yes


1) Define the set of allowed transformations φ

2) Define a useful measure of similarity

3) Choose a practical optimization procedure

y = j(x)

X

Y

x1

x2

y1

y2

GEOMETRIC

TRANSFORMATIONS

Geometric Transformations

Translation (2-D, 3-D, … N-D):

y = x + t

y = j(x)

1100

10

01

1

2

1

2

1

2

1

x

x

t

t

y

y

xt

yT

10

I


Rigid transformation

Rigid transformation + mirroring

y = Rx + t

y = Rx + t

y = j(x)


Affine transformations

y = Ax + t

y = j(x)


Projective (2-D):

A

222120

121110

222120

020100 , hyhxh

hyhxhy

hyhxh

hyhxhx

Similarity Affine Projective

Geometrical Transformations

Intensity Interpolation

Both position and signal value is interpolated during registration

Without proper resampling of the image values (color/intensity/…), the image might

look ugly and the registration may perform poor

Nearest neighbor

Bi- (2-D) or trilinear (3-D) interpolation

Bicubic or spline based interpolation, Windowed sinc, …

In Matlab: imrotate, imresize, interp2, interp3, …

Intensity value interpolation enables sub-pixel precision registration

Nearest neighbor Bilinear Bicubic





y = j(x)

X

Y

x1

x2

y1

y2

Different kinds of image fusion

Similar images

Same camera

Same modality

Same patient

Less similar images

Different brightness or contrast

Different lighting (e.v. day/night)

Different modalities (CT/MRI/Ultrasound)

Photo vs a drawing

SIMILARITY MEASURES

Image similarity – area based methods

Several options exist to measure (dis)similarity between the transformed source image φ(X) and the target Y: Sum of squared differences (dissimilarity)

Correlation (similarity)

Mutual information (similarity)

These are based on pointwise comparisons between pixels in φ(X) and Y

Compare pixel values in φ(X) and Y,

sweep and average/sum/aggregate

over the whole image φ(X) where it

overlaps with Y

Sum of Squared Differences

Pros:

Simple, intuitive and fast

Cons:

Signals must have exactly the same brightness and

contrast

Does not work for different modalities

2)(),( ii yxYXSSD

The correlation coefficient

How can we find this template?

Linear intensity changes are common in

practice. SSD is not appropriate here.

Normalized Cross-Correlation can be used.


Pros:

Corrects for intensity changes (common in practice)

Efficient to evaluate

Cons:

Does not work for (very) different modalities.

Flattnes of similarity measure maxima.

In Fourier domain:

Phase correlation for translation

Find a peak in the inverse of the Cross power spectrum of the two

FTs of the images.

Position indicates displacement.

)()(

)()(

YFXF

YFXF

• When data come from different modalities, there is no

linear relationship between voxel intensities and a simple

similarity measure as SSD or correlation will not suffice.

Cost Function for Dissimilar Images

Histograms

0 2 2

1 4 5

1 4 5

3 3 2

3 6 5

3 6 7

X

Y

0

2

4

6

0 1 2 3 4 5 6 7

0

1

2

3

0 1 2 3 4 5 6 7

Hx(k)

Hy(k)

Joint Histograms

0 2 2

1 4 5

1 4 5

3 3 2

3 6 5

3 6 7

X

Y

k/m 0 1 2 3 4 5 6 7

0 0 0 0 0 0 0 0 0

1 0 0 0 0 0 0 0 0

2 0 0 1 0 0 0 0 0

3 1 2 1 0 0 0 0 0

4 0 0 0 0 0 0 0 0

5 0 0 0 0 0 1 0 0

6 0 0 0 0 2 0 0 0

7 0 0 0 0 0 1 0 0

X

Y

Hxy(k,m)

The Joint Histogram

The Joint Histogram of images Y and φ(X):

H(k,m; Y,φ(X)) = Number of pixel positions with intensity k in image Y and

intensity m in image φ(X)

Joint intensity histogram

Before

registration

After registration

Spect

Spect

MR

M

R

Example: Fused MR and Spect images

o Mutual Information measures images similarity:

o H(k, m) is a value in the joint histogram

o Hx(k) and Hy(m) are values in the ordinary histograms for image X and image Y, respectively

o Mutual information measures how much information two sources X and Y have in common

(amount of uncertainty in Y reduced when X is known)

o Mutual information can be expressed in terms of entropy of distributions.

MI = H(k,m)logHxy (k,m)

Hx(k)× Hy (m)

æ

è ç ç

ö

ø ÷ ÷

k,m

å

Mutual Information

Hxy

Hy

Hx

Entropy(X)

Entropy(Y)

MI(X,Y)

• The MI registration criterion states that the images

are geometrically aligned when MI(A,B) is maximal.

• For example, let A be an MRI-scan and B a

SPECT-scan. If we know the MRI intensities, the

uncertainty of the SPECT intensities is minimal

when the scans are aligned.

Mutual Information

Joint intensity histogram

Before

registration

After registration

Spect

Spect

MR

M

R

MI =

0.357

MI =

0.446

Mutual information – an Example

Area based similarity measures

Appropriate in absence of distinctive details in

the images

Based on image intensities (not e.g., shape or

structure)

Require rather similar images (same or

statistically dependent)

Limited to small deformations (translation and

small rotations) to work reasonably well.





y = j(x)

X

Y

x1

x2

y1

y2

Functionals for Image similarity

A cost functional evaluates a particular transformation φ

The goal is to minimize the cost: A low cost means that

we have a good fit

We can also include an additional cost to punish “bad”

inappropriate transformations

(Restricting the set of possible transformations is also a kind of punishment,

i.e. setting the cost to infinite to all other transformations)

Cost = image dissimilarity(φ(X),Y)

Cost = image dissimilarity(φ(X),Y) + deformation cost(φ)

Optimization

Non-convex functions have local minima

We are looking for the global one

global minimum local minimum

gradient decent

Simulated annealing

(eventually steps should

be smaller and less random)

Starting point

Area based methods – fitness landscapes

Image and template

SSD

Distance based on fuzzy set theory

1-NCC

FEATURE BASED METHODS

Feature points based registration

Instead of matching all pixels with all pixels, it may be

beneficial (faster, more robust) to match salient

structures (points/lines) in the two images.

1. Feature detection – find outstanding points in the images

2. Feature description – to improve pairing of feature points

3. Feature matching – pairing up feature points

4. Transformation estimation – actually match the two

images (possibly with outlier rejection)

Interest Point Detectors

Feature points:

Distinctive, frequently spread, easy to detect.

Invariant under image degradation.

Surrounded by rich content.

Feature detectors:

Corner detectors

Edge detectors

SIFT – Scale-Invariant Feature Transform

Iterative Closest Points – ICP

1. Associate feature points in X with their closest

neighbors in Y

2. Estimate a transformation of respective points in X to

Y (e.g. an affine transformation)

3. Transform points in X, i.e. X = φ(X);

4. Iterate steps 1-3 until convergence

Requires good initialization to not get stuck in local minimum!

Find pairs of closest

feature points in X & Y

Transform X closer to

Y, using matching pairs

Registration – Least Squares Fit

Example: 2D Affine Transform , y = Ax + b

x = [x1, x2]T and y = [y1, y2]T

x is a point before the transformation

y is a point after the transformation

For several pairs of before/after points (x, y),

the problem of finding A and b becomes a

linear equation system.

For more than three pairs (x, y), there is no

guarantee to find A and b. However, we can

seek a Least Squares Solution.

X

Y

Registration – Least Squares Fit

Arrange all X-points in a matrix, where each

column represent a point in X. Call it X.

Arrange all Y-points in a matrix, where each

column represent a point in Y. Call it Y.

Find a Least Squares Fit using

X

Y Quadratic cost = sensitive to outliers

RANSAC

RANSAC – RANdom Sample Consensus

A robust estimation technique

Non-deterministic

Similarities with Hough- and Radon transforms

Robustness with respect to outliers

Paradigm: Hypothesize and Test

RANSAC

An example. Find the translation that transforms

red dots to yellow.

There is one outlier present

One corresponding pair is

needed (at least) to determine

the transformation

Outliers

RANSAC

In RANSAC we randomly select a pair and use it

to estimate a transformation

Hypothesis No match

RANSAC

And continue to select pairs until we find a good

transformation where many points are matching

(the outliers will of course not match)

Hypothesis

Match!

RANSAC – Algorithm

N, minimal number of data points needed to

estimate a model;

K, maximum number of allowed iterations;

T, threshold for data fitting (what is an OK match?);

D, number of data points needed to consider a fit

as good.

RANSAC – Algorithm

1. Draw a random sample of N data points;

2. Use these N data points to estimate a model M;

3. Test if this model fits D data points, within a threshold T;

4. If it does, you have a global fit. Exit.

5. If it does not, and if you have iterated fewer than K times, return to 1.

Optional final step: Use all inliers in a

LSQ fit to fine tune the final transformation.

RANSAC – Comments

Obviously, random sampling may be inefficient.

RANSAC is non-deterministic, you may not find the

global optimal fit within the K allowed iterations.

With larger model complexities, N, RANSAC may

need a lot of iterations.

RANSAC imposes problem-specific thresholds.

RANSAC estimates only one model for a particular

data set.

May perform poorly with more than 50% outliers.

FEATURE DETECTION AND

DESCRIPTION

SIFT – scale invariant feature transform

SIFT, detect interest points:

1) L(x,y,σ) = G(x,y, σ)*I(x,y), where G is a Gaussian

2) D(x,y,σ) = L(x,y,kσ) – L(x,y,σ)

3) Find max/min in D(x,y,σ)

4) Interpolate subpixel positions of max/min

Images from: Implementing the Scale Invariant Feature Transform(SIFT) Method

by MENG and Tiddeman

SIFT, detect interest points (cont.):

4) Eliminate low contrast responses D(x,y, σ) < 0.03

5) Eliminate edge responses (r=10)

H =Dxx Dxy

Dyx Dyy

é

ë ê

ù

û ú

Tr(H)2

Det(H)<

(r +1)2

r


Now compute feature vectors (description)

Position a 4 x 4 grid

at the position of each interest point

scaled according to the detected scale

oriented according to the dominant gradient direction


In each part of the grid, compute an 8 direction

histogram over the gradient directions

Compose this into a 4 x 4 x 8 = 128 value vector

This feature vector is invariant to scale, rotation

and intensity changes


From: Programming Computer Vision with Python

SIFT features are invariant to scale, rotation and

intensity changes.

SIFT features are robust to noise and 3D

viewpoint changes.

Scale and position are given by the interest point

detection step.

Rotation invariance comes from choosing the

strongest gradient as a reference direction.


Feature point pairing

Distance (e.g. Euclidean) between the descriptors

Pick the most similar feature point.

Weighting based on reliability can be used.

Decreased performance (multiple matches)

observed in cases of high illumination changes and

non-rigid deformations.

Note that invariance and distinctiveness are to some

extent contradictory properties, and have to be in

some balance.

RANSAC applied

Correspondences based on feature descriptor alone

RANSAC filtered correspondences

Summary

Image registration is an optimization problem.

Define a similarity measure

Restrict transformation space

Choose an optimization method

We have observed parametric transformations.

Complexity increases if local deformations are allowed.

Area based registration

Feature points based registration

Feature detection

Feature description

Feature matching

Transformation estimation (ICP, LSQ, RANSAC)

Optimization is challenging

Chamfer matching – Lab2

Technique for finding best fit of edge points

from two different images by minimizing a

generalized distance between them.

Algorithm based on distance transform to

locate one-dimensional features (edges).

Good performance in close to correct positions,

but poor elsewhere. (Can be seen as a variant of

ICP)

Chamfer matching – Lab2

Start from edge image

Compute DT from edges

Find locations providing minimal error

image registration - uppsala university · image similarity – area based methods several options...

Documents