itjim02(first review)

8/7/2019 ITJIM02(First Review)

1/13

Registering Images from an Image Ensemble

using Clustering Methods

Introduction:

Suppose you have several images all of the same content and you want to

register them all together. We call this collection of images an ensemble. The vast

majority of registration methods are designed to register only two images at a time.

It is not clear how to use these pairwise methods for ensemble registration. The

problem of registration becomes more difficult when the images come from

different sources. For example, a body part could be imaged with different

modalities In these cases, the image intensities cannot be compared directly

because, although the images depict the same content, they do so with different

transfer functions. We refer to such registration problems as multisensor

registration. The pairwise ensemble registration has the undesirable property that

the solution depends on which pairs of images are chosen and registered. We will

refer to this issue as selection dependency. Plotting the points for all the pixels

creates a scatter plot in this joint intensity space, and we refer to this scatter plot as

the joint intensity scatter plot, or JISP. The idea behind many multisensor

registration methods is to reduce the dispersion in the JISP. The implicit

assumption linking different images of the same object is that they are recognizable

as the same object because of some consistency by which intensities are assigned

to components in the image.


2/13

Literature survey:

Joint intensity scatter plot:

Plotting the points for all the pixels creates a scatter plot in this joint

intensity space, and we refer to this scatter plot as thejoint intensity scatter plot, or

JISP. The idea behind many multisensor registration methods is to reduce the

dispersion in the JISP. The implicit assumption linking different images of the

same object is that they are recognizable as the same object because of some

consistency by which intensities are assigned to components in the image. For

example, bones show up as black in an MR image, and white in a CT image.

Even though bones are rendered with a different intensity in each imaging

modality, we still recognize the similarity in global shape because the intensity

correspondence is consistent across many pixels. That is, pixels with intensities

near x in one image often correspond to pixels with intensities near y in the other

image. We call this correspondence an intensity mapping. An intensity mapping

need not be one-to-one. Indeed, there are lots of examples where two pixels withthe same intensity in one image correspond to different objectsand different

intensitiesin another image. Using our MR/CT example again, white matter and

gray matter are virtually indistinguishable in CT, yet yield noticeably different

intensities in T1-weighted MRI.


3/13

Scope of the Project:

To recognize an object based on a query image from template images. idea

behind many multisensory registration methods is to reduce the dispersion in the

JISP.

Module:

User Interface Design

Multi-pose image case

Effects

Recognization

Module Description:

User Interface Design:

In this module we design the windows for the project. These windows

are used to give input to the process and to display the output. We use the Swing

package available in Java to design the User Interface. Swing is a widget toolkit for

Java. It is part of Sun Microsystems' Java Foundation Classes (JFC) an API for

providing a graphical user interface (GUI) for Java programs.

Multi-pose image case

When more than a single image from the same object is available we assume

that the relative pose of the same object in two different images is known or can be

estimated. For two image cases, denote by and two unknown poses of an object

and by the known absolute difference between the poses. If the pair of templates

and a pair of candidates have a signal in common, the joint distribution is assumed

to follow Gaussian distribution with mean zero and covariance matrix composed of

four block matrices given in and If the pairs do not have a signal in common, their


4/13

joint distribution follows Gaussian distribution with mean zero and covariance

matrix, which is equal to with off-diagonal block matrices set to zero.

Consider two images, one overlaid on the other. Each pixel corresponds to two

intensity values, one from each of the two images. This 2-tuple can be plotted in

the joint intensity space, where each axis corresponds to intensity from each of the

images. Plotting the points for all the pixels creates a scatter plot in this joint

intensity space, and we refer to this scatter plot as the joint intensity scatter plot, or

JISP. The idea behind many multisensor registration methods is to reduce the

dispersion in the JISP. The implicit assumption linking different images of the

same object is that they are recognizable as the same object because of some

consistency by which intensities are assigned to components in the image. For

example, bones show up as black in an MR image, and white in a CT image. Even

though bones are rendered with a different intensity in each imaging modality, we

still recognize the similarity in global shape because the intensity correspondence

is consistent across many pixels. Here we present an efficient method for

multisensory ensemble registration. Our method is based on clustering in the JISP,

jointly modeling the distribution of points in the JISP as it estimates the motion

parameters. Density estimation of the clusters is modeled as a Gaussian mixture

model (GMM), and is established iteratively using an estimation-maximization

(EM) method. The motion parameters are also solved using an iterative Newton-

type method. The iterates of these two methods are interleaved, thereby solving the

two problems (density estimation and registration) in synchrony.


5/13

Effects

In this module we will give effects like dark ,shadow ,illumination or glossy

to the given input image and then the affected image will be given as input

for the Recognization process.

Recognization

Recognition is performed at the level of informative features extracted from

images of objects. The features are combined in vectors called templates.

Templates of different objects are stored in a library. We define recognition

channel, by analogy with communication channel, as the environment that

transforms reference templates of objects in a library into templates

submitted for recognition.

System Architecture:


6/13

Data flow diagram:


7/13

Use case Diagram:

I n p u t Im a g eM a k e T e m p l a t eR e c o g n i z a t i o n

O u t p u t I m a


8/13

Class diagram:

Im a g e In p u t

im a g e N a m e

x P o s i t io n

y P o s i t io n

i n p u t ( )

T e m p la t e

k e y

fo r m T e m p la t e ()

R e c o g n iz e

x p o s it i o n

y p o s it i o n

x 1 p o s it i o n

y 1 p o s it i o n

m a t c h ()

O u t p u t

im g n a m e

d is p la y ( )

Object diagram:

Im a g e In p u t

im a g e N a m e

x P o s i t io n

y P o s i t io n

T e m p la t e

im a g e 1

im a g e 2

n im a g e

R e c o g n iz e

x p o s it i o n

y p o s it i o n

x 1 p o s it i o n

y 1 p o s it i o n

O u t p u t

im g n a m


9/13

State diagram:

input

image

collect

image

recognize

output

make

template


10/13

Sequence diagram:

input output recognize template

image

cluster

match image


11/13

Collaboration diagram:

input output

recognize template

1: im age

2: c luster

3 : m atch im age


12/13

Conclusion:

Ensemble registration is the process of registering multiple images togethersimultaneously within a single optimization problem. This approach for

multisensor registration was not previously feasible because the high-dimensional

joint histogram was too large to store in memory. Instead, we use a Gaussian

mixture model to perform density estimation of the content in the joint intensity

space. This GMM model naturally leads to a cost function based on likelihood. We

formulate an optimization problem that has two aspects, developing solutions for

the density estimation and motion parameters in synchrony. Within each iteration,

we hold the motion parameters fixed and update the density estimation parameters,

and then hold the density estimation parameters fixed and update the motion

parameters.

Our experiments show that ensemble registration is more robust than

pairwise registration. The content shared by one pair of images might be quite

different from the content shared by another pair of images. The key is to leverage

all these correspondences simultaneously. Ensemble registration does exactly that,

implicitly coupling the content of all the images into one optimization problem.

The experiments also show that ensemble registration is more accurate than

pairwise registration. Not only does ensemble registration offer more image

correspondences (as described above), but it is also less susceptible to noise. This

benefit stems from the fact that the estimate of an entity gets more accurate as you

include more observations. Adding more images yields greater statistical

confidence.


13/13

Reference:

[1] M. Jenkinson and S. Smith, A global optimisation method for robust affine registration of

brain images, Med. Image Anal., vol. 5, no. 2, pp. 143156, 2001.

[2] A. Collignon, F. Maes, D. Delaere, D. Vandermeulen, P. Suetens, and G. Marchal,

Automated multi-modality image registration based on information theory, in Proc. Info Proc

Med Imaging, Y. Bizais, C. Barillot, and R. Di Paola, Eds., 1995, pp. 263274.

[3] W. M. Wells, III, P. Viola, H. Atsumi, S. Nakajima, and R. Kikinis, Multi-modal volume

registration by maximization of mutual information, Med. Image Anal., vol. 1, no. 1, pp. 3551,

1996.

[4] C. Studholme, D. L. G. Hill, and D. L. Hawkes, An overlap invariant entropy measure of 3D

medical image alignment, Pattern Recognit., vol. 32, pp. 7186, 1999.

[5] H. Neemuchwala, A. Hero, and P. Carson, Image matching using alpha-entropy measures

and entropic graphs, Signal Process., vol. 85, no. 2, pp. 277296, 2005

itjim02(first review)

Documents