Shri Hari

Registration establishes feature-based correspondence between related image pairs

Oct 5, 2008

Unit IV

CT highlights specific features in the image of a human organ MRI highlights another set of features in the image of a human organ To enhance the medical image, it is essential to fuse the CT & MR images of almost identical cross sections of a human organ.
Oct 5, 2008 2

Registration establishes correspondence between

different shape-based features

of CT & MR images
Oct 5, 2008 3

Satellite image registration

To estimate wind speed & direction over sea surface, meteorologists take the help of two consecutive satellite images taken at half-hourly interval. A specific cloud cluster is identified in both the images. The shift of cloud cluster in the II image wrt the I one provides the wind velocity

Registration is also

Oct 5, 2008

A process of mapping between a temporal sequence of image frames

4

How do we match two images?????

Transformation needs a set of features in both the images. These features are known as

Oct 5, 2008

Establishing a correspondence

Me an s

Matching of identical shapes in the related image pair Geometric transformation of one image into another
Oct 5, 2008 6

The mapping between the related image pairs is achieved using

Oct 5, 2008

Image registration is the task of applying some transformations to two images so that they match as best as possible.

Oct 5, 2008

In many image processing applications, it is necessary to form a pixel-by-pixel comparison of two images of the same object field obtained from different sensors, or from same sensor but at different times

Oct 5, 2008

For this comparison, it is necessary to spatially register the images & thereby, to correct for relative translational shifts, rotational differences, scale differences & even perspective view differences

When talking about geometric transformations, we have to be very careful about the object being transformed.

Oct 5, 2008

10

Two basic steps in Geometric Transformations

A spatial transformation of the physical rearrangement of pixels in the image A gray level interpolation, which interpolation assigns gray levels to the transformed image
Oct 5, 2008 11

Change in view point

Relative motion between the camera & object planes

Introduces distortion in the feature of an image

When viewed from a non-fronto-parallel vantage point

Oct 5, 2008 12

The geometric transformation models are used to parameterize this object to image distortion

Certain features of the object shape that remain intact or preserved even after transformations e.g. The length of a vector remains unchanged even after rotation
Oct 5, 2008 13

Oct 5, 2008

14

Representation of a Point
A two dimensional point (x, y) or a three dimensional point (x, y, z) can be represented as a matrix

Oct 5, 2008

15

The homogeneous coordinates [ x y 1] represents the position vector [x y] in the physical xy plane
Oct 5, 2008 16

Is homogeneous coordinate representation of a point unique? There is no unique homogeneous coordinate representation!!! e.g. [6 4 2], [12 8 4], [3 2 1] all represent the physical point (3, 2)
Oct 5, 2008 17

I. Euclidean Transformations The Euclidean transformations are the most commonly used transformations. A Euclidean transformation is either a translation, a rotation, or a reflection.
Oct 5, 2008 18

Translation
Translate (a,b): (x,y) (x+a,y+b)

Oct 5, 2008

19

(x, y) (x, y) where (x, y) = (x+h, y+k)

Then, the relationship between (x, y) and (x', y') can be put into a matrix form

Oct 5, 2008

20

Therefore, if a line has an equation Ax + By + C = 0 After plugging the formulae for x and y, the line has a new equation Ax' + By' + (-Ah - Bk + C) = 0.

Oct 5, 2008

21

Oct 5, 2008

22

Oct 5, 2008

23

Oct 5, 2008

24

If a point (x, y) is rotated an angle a about the coordinate origin to become a new point (x', y'), the relationships can be described as follows:

Oct 5, 2008

25

Thus, rotating a line Ax + By + C = 0 about the origin a degree brings it to a new equation:

(Acosa - Bsina)x' + (Asina + Bcosa)y' + C = 0

Oct 5, 2008 26

Rotating an object about Y-axis

Rotate an object about Y-axis by y in a counterclockwise direction.
Z X * cos y * Y = 0 Z * sin y 1 0 0 sin y 1 0 0 cos y 0 0 0 X 0 Y 0 Z 1 1

Y= Y X

In this case the object is moved. Only x and z are changed while y stills the same.
Oct 5, 2008 27

Translation & Rotation can be combined

Rotates the point (x, y) by an angle a about the coordinate origin and translates the rotated result in the direction of (h, k). Oct 5, 2008 28

Rotation and translation are not commutative!

If translation (h,k) is applied first followed by a rotation of angle a (about the coordinate origin), we will have the following:

Are translation & rotation commutative?

Oct 5, 2008

29

Oct 5, 2008

30

Properties preserved under Euclidean Transformations

Lengths Angles areas
Euclidean transformations preserve length and angle measure. Moreover, the shape of a geometric object does not change - lines transform to lines, planes transform to planes, circles transform to circles, and ellipsoids transform to ellipsoids. Only the position and orientation of the object will Oct 5, 2008 31 change.

II. The Plane Affine Transformations

Affine transformations are generalizations of Euclidean transformations. Under affine transformations, lines transforms to lines; but, circles become ellipses. Length and angle are not preserved. Affine transformations include: Scaling Shear and General affine transformations.
Oct 5, 2008 32

An Affine Transformation maps variables

(e.g. Pixel intensity values located at (x, y) in an i/p image into new variable (x, y) in the o/p image)

by applying linear combination of translation, rotation, scaling &/or shearing.

Oct 5, 2008

33

Scaling transformations stretch or shrink a given object and, as a result, changes lengths and angles. So, scaling is not an Euclidean transformation. The meaning of scaling is making the new scale of a coordinate direction p times larger. In other words, the x coordinate is "enlarged" p times. This requirement Oct 5, 2008 34 satisfies x' = p x and therefore x = x'/p.

Oct 5, 2008

35

Oct 5, 2008

36

Scaling can be applied to all axes, each with a different scaling factor. For example, if the x-, y- and z-axis are scaled with scaling factors p, q and r, respectively, the transformation matrix is:

Oct 5, 2008

37

The effect of a shear transformation looks like ``pushing'' a geometric object in a direction parallel to a coordinate plane (3D) or a coordinate axis (2D). In the following, the red cylinder is the result of applying a shear transformation to the yellow cylinder:

Oct 5, 2008

38

Oct 5, 2008

39

How far a direction is pushed is determined by a shearing factor. On the factor xy-plane, one can push in the x-direction, positive or negative, and keep the ydirection unchanged. Or, one can push in the y-direction and keep the x-direction fixed. The following is a shear transformation in the x-direction with shearing factor a:

Oct 5, 2008

40

The shear transformation in the y-direction with shearing factor b is

Oct 5, 2008

41

In space, one can push in two coordinate axis directions and keep the third one fixed. The following is the shear transformation in both x- and ydirections with shearing factors a and b, respectively, keeping the z-coordinate the same:

Oct 5, 2008

42

Expanding the matrix equation gives the following: x' = x + az y' = y + bz z' = z Thus, a point (x, y, z) in space is transformed to (x + az, y + bz, z) Therefore, the z-coordinate does not change, while (x, y) is ``pushed'' in the direction of (a, b, 0) with a factor z. Oct 5, 2008

43

The general affine transformation matrix has the following form

Compare this with all the previously discussed Affine transformations matrices
Oct 5, 2008 44

Rotations and translations included, all of them fit into this form The matrix form of an affine transformation is a 3-by-3 matrix with the third row 0, 0 and 1. Affine transformation has six degrees of freedom.
Oct 5, 2008 45

The general 3 x 3 transformation matrix for two-dimensional homogeneous coordinates can be subdivided into four parts

a, b, c, d produce scaling, rotation, reflection & shearing

Oct 5, 2008

p, q produce translation

m, n, s when p=q=0 & s=1, the homogeneous coordinate of the transformed position vector is always h=1 46

Properties preserved under Affine Transformations

Parallelism Ratio of lengths of collinear or parallel segments Ratio of areas Linear combination vectors

Oct 5, 2008

47

III. The Plane Projective Transformations

Most general linear transformations Use homogeneous coordinates It is the projectivity from one projective plane to another plane The transformation matrix must be non singular

Oct 5, 2008

48

Given a point in space in homogeneous coordinate (x,y,w) and its image under a projective transform (x',y,w'), a projective transform has the following form:

Oct 5, 2008

49

The transformation matrix T, has eight degrees of freedom

(since only the ratio of homogeneous coordinates is significant & there are eight ratios among nine elements of T)

Oct 5, 2008

50

Projective transformation can bring finite points to infinity and points at infinity to finite range.

Oct 5, 2008

51

Consider the following projective transformation for the point (x, y, w) = (1, 0, 1):

The transformation sends (x, y, w) to (x, y, w)

(1, 0, 1)

(1, -1, 0)

This projective transformation sends (1,0) on the xy-plane to the point at infinity in direction <1,-1> Oct 5, 2008

52

Scale in the x-direction using a scale factor 5 (i.e., making it five times larger). Followed by a rotation about z-axis 30 degree Followed by a shear transformation in xand y-direction with shearing factor 2 and 3, respectively. Followed by a transformation moving the point in the direction of < 2, 1, 2 >.
Oct 5, 2008 53

Find a matrix to perform the following transformations to an object:

Let the scaling, rotation, shearing and translation matrices be A, B, C and D, respectively. Then matrix H = DCBA gives the net effect:

Oct 5, 2008

54

Oct 5, 2008

55

Properties preserved under Projective Transformations

Concurrency Collinearity Order of contact (intersection, tangency, inflections) Tangent discontinuities Cusps
Oct 5, 2008

Cross - ratio

56