Scale space And Anisotropic Diffusion

With linear low pass filtering and isotropic diffusion we let the pixel
values to be mixed with across the edges, so the images become
blurred. So, it becomes inconvenient to detect the edges. So, the
concept of anisotropic diffusion comes. The main concept of this is
IT JUST ALLOWS THE PIXELS TO BE MIXED IN THE SAME
SIDE OF BOUNDARY. It is based on the logic of calculus of
variation which finds extreme values of integrals of a function.
Formally, let
denote a subset of the plane and
be a family of gray scale images, then anisotropic diffusion is
defined as

where denotes the Laplacian, denotes the gradient,

is the divergence operator and
is the diffusion
coefficient.
controls the rate of diffusion and is usually
chosen as a function of the image gradient so as to preserve
edges in the image. This does not let allow the pixels to diffuse
across the edges where c is high.

Edge using anisotropic diffusion

Edge detection using canny

Bibliography:

EDGE DETECTION BY WAVELET TRANFORM

Most appropriate for non stationary signal.
The basic functions vary both in frequency and spatial
range.
The wavelet transform is designed such a way that we get
good frequency resolution low frequency(average
intensity values) and we get high temporal resolution for
high frequency component (for the edges).
STEPS:
1. Start with a mother wavelet such as harretc.
2. The signal is then translated to shifted and scaled version of
mother wavelet..
3. If subsignal are small they are made zero without significant
change in the image.

Now , the low-low component (approximated image) can be

modified for edge detection.
The simplest method is to replace approximated image by zero,
this will remove all low frequency components and reconstruct

it. the remaining component is of high frequency (corresponds

to an edge).
Another method is to pass the approximated images through a
good edge detector(specially canny). Now modified
approximated pixels and the remaining are used in
reconstruction which leads to an edge.

An MRI image(a)

edge detected using

wavelet transform

Bibliography: [1] Mauro Barni. Document and Image Compression.

CRS Press, Taylor and Francis Group,
2006.
[2] Al Bovik. The Essentuial Guide to Image Processing. Elsevier,
Inc., second edition, 2009.
[3] Rafael C. Gonzalez and Richard E. Woods. Digital Image
Processing. Pearson Education,
Inc., Upper Saddle River, New Jersey, third edition, 2008.

MUMFORD-SHAH IMAGE
SEGMENTATION
Segmentation is levelling the images ,one way is to paint it with grey
values. Mumford-shah implements it.
While detecting an edge there maybe two types of penalty.
Penalty for edges
Penalty for errors
THE BASIC CONCEPT OF MUMFORD SHAH Is to write basic
formulation to compromise between a representation image
which is too far from the original image to a simplified image
which is not too far from the original image. There should not
also many segmentation.
This concept is also used in image registration.
algorithms for solving the general Mumford{
Shah model tend to be relatively complicated and computationally
expensive. As a simplication,
Mumford and Shah also considered a piecewise constant
formulation which is used in Chan vese segmentation.

CHAN VESE SEGMENTATION

While many segmentation methods rely heavily in some way on edge

detection, method by Chan and Vese ignores edges completely.
Instead, the method optimally fits a two-phase piecewise constant
model to the given image. The segmentation boundary is represented
implicitly with a level set function, which allows the segmentation to
handle topological changes more easily than explicit snake methods..
STEPS:
The Chan Vese method is inspired by the Mumford Shah model.
Mumford and Shah approximate the image by a
piecewise-smooth function as the solution of the minimization
problem.
The minimization problem requires minimizing over all set
boundaries . This is accomplished by
applying the level set technique introduced by Osher and Sethian .
Instead of manipulating boundaries
explicitly, it is represented as the zero-crossing of a level set
function.
Here we describe the semi-implicit gradient descent for solving
the Chan Vese minimization as developed
in the original work . This is not the only way to solve the problem.

Original noisy image

by canny

by chan vese
Bibliography: [1] J. Canny, \A Computational Approach To Edge
Detection," IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 8, no. 6, pp. 679{698, 1986.
http://dx.doi.org/10.
1109/TPAMI.1986.4767851
[2] S. Osher, J.A. Sethian, \Fronts propagating with curvaturedependent speed: Algorithms based on Hamilton-Jacobi
formulations," Journal of Computational Physics, vol. 79, no. 1, pp.
12{49,
1988. http://dx.doi.org/10.1016/0021-9991(88)90002-2

LINE SEGMENT DETECTION

LSD is a linear-time Line Segment Detector giving subpixel accurate

results. It is designed to work on any digital image without
parameter tuning. It controls its own number of false detections: On
average, one false alarms is allowed per image . The method is based
on Burns, Hanson, and Riseman's method.
PROCESS:
The LSD algorithm takes a gray-level image as input and returns a
list of detected line segments. The algorithm can be described by
the following steps. The auxiliary image STATUS has the same size
as the scaled image, and is used to keep track of the pixels already
used.
1. Scale the input image to scale S using Gaussian sub-sampling
(=/S).
2. Compute the gradient magnitude and level-line orientation at each
pixel.
3. Build a list of pixels pseudo-ordered according to their image
gradient magnitude.
4. Set all pixels in the auxiliary image STATUS to the value NOT
USED.
5. Mark the STATUS of pixels whose gradient magnitude is less
than to the value USED.
6. For each pixel P in the list, starting with the ones with the
highest gradient magnitude, and with STATUS set to NOT USED,
do:
1. Starting from P as an initial pixel, grow a region R of
connected and NOT USED pixels, that share the same levelline angle up to a tolerance . Mark the STATUS of the pixels
in the region as USED.
2. Compute the rectangular approximation for the connected
region R of pixels found.
3. If the density of aligned points in the rectangle is less than
D, cut the region, until the density restriction is satisfied.
4. Compute the NFA(Number Of False Alarms) value for the
rectangle found.

5. Try to modify the rectangle to improve the NFA value.

6. If NFA(r)< (where =selected threshold value) add the
rectangle to the output list.
BIBLIOGRAPHY:
Rafael Grompone von Gioi, Jeremie Jakubowicz, Jean-Michel Morel,
Gregory Randall, LSD: a Line Segment Detector, Image Processing On
Line, 2012.
http://dx.doi.org/10.5201/ipol.2012.gjmr-lsd

INTERACTIVE SEGMENTATION

Some images are very much tough to segment as there is confusion

which part should be considered as background and which as
foreground. Here ,comes the concept of interactive segmentation.
Here, the user is allowed to tell which part of the image should be
considered as foreground and which as a background.
STEPS:
When an user marks the pixels for foreground as well as the
pixels for background forms a curve, from those curves we can
find the probability of the pixels of being in the fore as well as
background. But, this is not enough. It may happen that a pixel
inside the image looks like background.
Weighted geodesic distance: Along with the probability we
calculate the path from pixel to foreground as well as from pixel
to background. Whichever is shorter we take that.
Automatically create narrow bands and new scribbles to
differentiate between the foreground and background.

BIBLIOGRAPHY: https://www.ipisingapore.org/sites/default/files/styles/thumbnail_600x295/pub
lic/tech_offers/techinnovation_2014/interactive
%20segmentation%20toolbox.png?itok=VPXAS2vi,,coursera.com

Interactive off-line segmentation of moving

objects

in real traffic conditions

For real traffic management the segmentation using assuming
difference between pixels is not going to work. we have to use
method which can incorporate appropriate information about
regions, borders, occluding and shape. One of such modern methods
is Graphcuts which is able to distinguish occluded objects with the
same brightness levels as the background.

The nodes in our case represent pixels and we have two

terminals in the graph. Source terminal tagged as S is the
representation of all pixels corresponding with an object in the
picture and Sink terminal tagged as T which is the
representation of the background. Each pixel is connected to
both S and T terminals.
Non-negative weights we are assigned to each of the edges
e which contain n-links and t-links.
The s-t cut c is a subset of edges in a graph. The cut
divides nodes between terminals of a graph g(c)=(v,/c)
This divides the picture to object and background. The energy
cost is given as a sum of weights placed on the cut.
|c|= we

Interactive segmentation based on Graph-Cuts:

The user marks regions in the picture that belongs to the
object and regions that definitely are parts of the background
. Number of marked regions depends on the user and
the type of the task. Segmentation consists of three parts:
1. For each pixel inside the object is given a value that
represents similarity of its intensity with the model.
Low values represent better results.
2. For each pixel inside the background is given the
value that represent similarity of its intensity with the
model. Low values represent better results.
3. For each pair of neighbouring pixels, where one is
inside the object and the second is outside, is given a
value that represents the similarity of intensities of
both pixels. Low values represent their similarity.
Image analysis strategy:

This setting defines three regions that represent:

- Part of the object (it is the expansion of the seed line and
it is shown on Fig.2, it is the region labelled by 1)
- Part of the background (it is the frame of the scanning
region shown on Fig.2, it is frame labelled by 2)
- Unallocated region (it is shown in Fig.2 and it is labelled

by 0)
The algorithm finds the best border between regions
representing the object and background in unallocated region
labeled by zeros.

Motion tracking strategy:

Prediction is the main feature typical for motion tracking.
Extrapolation is the simplest form of trajectory prediction
where the position of the object in the next frame is given by
the position on previous frame shifted by vector (dx, dy)
representing the differences of positions on last two frames.
There are two important parameters in the model.
v = (x, y)
where x, y defines the position of point B.
The difference vector
d = (dx,dy)
this represents differences of individual parameters in last two
frames. The predicted values determining the model position,
rotation, shape for the present and for the next frame are given
Vt+1=Vt+ d

segmented object occluded by car by

graph cut method.

BIBLIOGRAPHY: [1] GREIG, D. PORTEOUS, B.SEHEULT, A.: Exact

Maximum A
Posteriori Estimation for Binary Images, J. Royal Statistical Soc.
Series B, vol. 51, no. 2, pp. 271-279, 1989
[2] ISHIKAWA, H. GEIGER, D.: Segmentation by Grouping
Junctions, Proc. IEEE Conf. Computer Vision and Pattern
Recognition, pp. 125-131, 1998
[3] BOYKOV , Y. KOLMOGOROV, V.: Computing Geodesics and
Minimal Surfaces via Gr

ACTIVE CONTOURS

The process of labelling the location in the image where the rate of
change is high is known as edge detection. From the gradient of the
image we get the info where there is major change in intensity in the
image. So, the gradient gives the local information. The goal of
active contour is to integrate those local measurement. Edge
integration is the process of combining local, perhaps sparse and non
contigious edges into meaningful long edge curve for segmentation.
There is relationship between local edge and active contours going to
compute it .

PROCESS:

Geometric active contour, or geodesic active contour (GAC)[8] or

conformal active contours[9] employs ideas from Euclidean curve
shortening evolution. Contours split and merge depending on the
detection of objects in the image. These models are largely inspired
by level sets, and have been extensively employed in medical image
computing. For example, the gradient descent curve evolution
equation of GAC is
[8]

This level set information enables active contours to change topology

during the gradient curve evolution. Where curve value is minimum it
suggests boundary.

1. Bibliography: Kass, M.; Witkin, A.; Terzopoulos,

D. (1988). "Snakes: Active contour models" (PDF).International
Journal of Computer Vision 1 (4):
321. doi:10.1007/BF00133570.
2. Jump up^ Snakes: an active model,Ramani
Pichumani,http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL
_COPIES/RAMANI1/node31.html

3. Jump up^ Dr. George Bebis,University of

Nevada,http://www.cse.unr.edu/~bebis/CS791E/Notes/Defor
mableContours.pd

A Real Time Morphological Snake Algorithm

Active contours, or snakes, are computer-generated curves that

move within images to find out salient image structures like object
boundaries. Energy based formulations using a level set approach
have been successfully used to model the snake evolution. The EulerLagrange equation associated to such energies yields to partial
differential equations (PDE) which are usually solved using level set

methods which involve contour distance function estimation and

standard methods to discretize the PDE. Recently a morphological
approach to snake evolution has been proposed. First, we observe
that the differential operators used in the standard PDE snake
models can be approached using morphological operations. By
combining the morphological operators associated to the PDE
components we achieve a new morphological approach to the PDE
snakes evolution. This new approach is based on numerical methods
which are very simple and fast. Moreover, since the level set is just
a binary piecewise constant function, this approach does not require
to estimate a contour distance function to define the level set.

Example of morphological snake algorithm

Bibliography:
[1] M. Aleman-Flores, L. Alvarez and P. Henriquez, \AmiLabContours:
A tool for image struc-ture segmentation," IADIS Multi Conference
on Computer Science and Information Systems,Freiburg, Germany,
2010. http://ami.dis.ulpgc.es/ami/amilabcontours.html
[2] L. Alvarez, L. Baumela, P. Marquez-Neila and P. Henriquez,
\Morphological Snakes,"CVPR 2010, San Francisco, U.S.A.
http://www.sciweavers.org/publications/ morphological-snakes

[3] V. Caselles, R. Kimmel, and G. Sapiro, \Geodesic active contours,"

International Jour-nal of Computer Vision, vol. 22, no. 1, pp. 61{79,
1997. http://dx.doi.org/10.1023/A:
1007979827043
//mw.cmla.ens-cachan.fr/~morel/JMMBookOct04.pdf.
http://www.ipol.im/pub/art/2012/abmh-rtmsa/

FAST EDGE DETECTION USING SUBBAND

DECOMPOSITION
The method uses a kind of two dimensional subband spectrum
analysis (2D-SSA) filter that is based on subband decomposition,
and it is very convenient to get the edge frequency spectrum of an
image after certain pre-processing. Comparing with
spatial methods, the method is less sensitive to noise. It is also
superior to the conventional frequency methods. Subband Spectrum
Analysis algorithm is derived from Subband Decomposition
DFT(SD-DFT), and it is more flexible and general

than SD-DFT. From 1D-SSA, we can get the 2D-SSA:

Generally an image has large amounts of data, and one needs a long
time to process it. In the fast recognition of a target image, it is
very important to reduce the processing time. All these possibilities
and requirements make the 2D-SSA method valuable
in fast edge detection.The parameters which are needed to be
regulated are M, k0 and l0. M represents the times of
decomposition, and it determines the bandwidth and
central frequency of 2D-SSA lowpass filter. k0, l0 are
the amounts of frequency shift respectively. A group
optimum value of M, k0 and l0 can be got through
regulating.

original image
subband decomposition
bibliography:
Hossen AN, Heute U
Fully Adaptive Evaluation of
Subband DFT. Chicago: Proc. of IEEE ISCAS, 655-8.
Hossen AN, Heute U (1994). Two-Dimensional Sub-Band
DFT Algorithm. NOPRSIG 94:163-7.

(1993).