Anisotropic Diffusion

Summing of the article

scale-space and edge detection using anisotropic diffusion
By Pietro Perona
And
Jitendra Malik


Nerya Agam
Computer-Science
undergraduate

About the upcoming class:

In this class, we will introduce the basics of Anisotropic
diffusion, and its usage in scale-space actions and edge
detection in graphic images.

The focus of the lecture will be gaining intuition about
the Anisotropic Diffusion work process. Less focus will be
dedicated to Mathematical analysis of procedures. This
author thereby promises, to make his best effort and keep
you interested throughout the class.

Any attendant may ask any question at any time during
class.
Introduction
What is the main drawback of this
picture ?
-Image noise

In what ways does it bother us ?
- visually as humans
- lost detail
- renders the image unfit for most
of our desired image manipulation
filters

So, we are looking for an efficient method of noise
removal, that is able to clarify the image.

Methods of noise removal were known a long time
before the Anisotropic Diffusion case was first claimed
(1990)

Most of the methods are typically forms of applying a
blur filter to the image, as Blurring it hopefully results in
a smooth, noise-less product.

Such a product, with low noise levels, is a much better
starting point for any image related algorithm
Classic smoothing methods:

Based on convolving a Gaussian kernel with each and every
pixel. A typical kernel would be of size NxN.

The single pixels brightness value is determined by its own
original value, as well as the values of its neighbor pixels.

an appropriate definition of the transformation would be:

I (x,y) = I
0
(x,y) * G(x,y,t)







Making a gaussian kernel of size NxN:

Set the center of the kernel to be x=0, y=0 => (0,0)

Using the equation, set the values on every square of the
kernel:
Squares closer to the center (such as (1,2)) would get higher
values than squares which are close to the kernels border.

Note:
t = theoretically is sigma^2, tough in a C.s implementation it is
the size of the gaussian (3x3,7x7 etc)

2 4 5 4 2
4 9 12 9 4
5 12 15 12 5
4 9 12 9 4
2 4 5 4 2
Gaussian 5x5 (without normalization)
Taking a gaussian kernel of size NxN:

Set the center of the kernel to be x=0, y=0 => (0,0)

Using the equation, set the values on every square of the
kernel:
Squares closer to the center (such as (1,2)) would get higher
values than squares which are close to the kernels border.

The parameter t stands for the variance of the gaussian:

when t=0 => I(x,y) = I
0
(x,y) (just the ID transform)

As t grows, the brightness of a pixel in the resulting image
considers more and more of its neighboring values.

In this next example, notice how the blurring makes the grass
color almost completely homogeneous which in our case is
considered a desired effect.

This is because the smoothing operation is vary useful in
highly homogeneous regions, since it scales-down the
complexity of the region to a simple blub of very few
discrete brightness levels.

Using a few iterations of classic Gaussian Blurring, with kernel
size of 7x7, 9x9 and 13x13, gives the following product image:


While being a popular tool, Gaussian pass has its
downsides:

Loss of fine detail

Smoothing across boundaries

The first could be expected, as it is the direct trade-off of
smoothing

The second is problematic with no clear boundaries,
image segmentation proves to be difficult.

Here are some examples:

This ww2 photo is given here in its original form.
From a human eye point of view it has a good enough
quality.

Pay attention : some surfaces, like the runways asphalt or
the roofs of the buildings, do not have a smooth texture


(pacific us navy airbase, ww2)
Now, pass the airbase photo through an edge detecting
filter (sobel).

Result:
along with the correct edges, this product photo
contains many false edges. This is not a good enough
edge-marking product.
If we smooth the
image, we can
expect a lot less
noise

Smoothing the
original with
Gaussian 5x5
Running edge detection
Many of the false edges were smoothed.
Unfortunately, so were the true edges.

Using this linear smoothing, gives especially poor results in more
coarse images
(a.k.a low resolution images), as Blurring already makes an image a
lot more coarse.

This is because meaningful edges in a coarse image would be
smudged so much, that it is hard to determine where the real
meaningful edge is originally located.

Perona & Malik
suggestion
Perona & Malik suggestion
A smoothing algorithm has to stand to a criteria:

1. Causality no spurious detail should be generated while passing
from finer to coarser scales.

2. Immediate Localization at each resolution, the region
boundaries should be sharp and coincide with the semantically
meaningful boundaries at that resolution.

3. Piecewise Smoothing at all scales, intraregion smoothing
should occur preferentially over interregion smoothing.


Smoothing that respects this criteria should get better results
Maybe as good
as this:




Maybe better ?
In order to achieve this desired filtering, lets take a look at
some of the basic fundamentals, that compose the whole
algorithm



Starting with:
Non-Linear passes

Linear approach:

Treat every pixel with the exact same convolution.






Non-Linear approach:

Treat a pixel with varying intensity, depending on its
neighborhood qualities.

Here is an abstraction of the principal:

A Non-Linear equation, helpful to our cause

Let us say that we have a method E of knowing if a certain
point in the image is a part of an edge or not.

We can make a new transformation
a filter that smoothes inside a region, but skips the edges in the
image.

Generally speaking:
if (x,y) is a part of an edge apply little smoothing
if not a part of an edge apply full smoothing



To implement this specific Non-Linear approach, we need a
detector to tell us if (x,y) is a part of an edge or not.

Luckily, the Gradient of the brightness function is quite good
for that:








In the first part,
the gradient vector is (0, n)
while (n > 0)

In the second part,
the gradient vector is
(-m , m)
while (m > 0 )

These points where the
norm of the gradient is
high, could be treated as
edge points, end
therefore be applied less
blurring
The coefficient
Until now:
We have an edge/not edge estimation method E.

We need to create a function g, that controls blurring
intensity according to .

g has to be a monotonous decreasing function (why ?)
The coefficient

Perona & Malik trialed with two different g definitions:

K = 1
K = 4
K = 1
K = 4
The coefficient
Now, we have E and g. Let us define the coefficient.

The coefficient controls how much smoothing is done in (x,y)

Simplification:

- C(x,y,t) is large when x,y is not a part of an edge

- C(x,y,t) is small when x,y, is a part of an edge
Original
Linear Non-Linear
A Non-Linear pass has these properties:

Good intraregion smoothing

generally keeps the edges as they were, doesnt do much
interregion smoothing


So, going Non-Linear might be a part of the solution, but is not
good enough on its own (edges stay rough)


Continuing with:
Anisotropic Diffusion
Warning:

The following may strengthen
your intuition of the flow of a
smoothing process
Diffusion
Diffusion
describes the
spread of
particles
through random
motion from
regions of higher
concentration to
regions of lower
concentration.
(wiki)
Isotropic:

that is not dependent of direction.




Anisotropic:

depends on the direction applied
Koenderink and Hummel pointed out that the family
(family = product of different t values)
of derived images from the equation:

I (x,y) = I
0
(x,y) * G (x,y,t)

Can be viewed as:

the solution of the Heat Diffusion equation


Like heat, generated in a cold surrounding
the gradient size represents the margins of temperature.
As time advances on (t), each potent molecule (a hot unit)
spreads in the direction of its gradient vector.

Surprisingly, it is also valid for liquid diffusion and more.

The matter in an image is not heat, but brightness level
So, an image could be generalized to be a surface, where bright
spots are hot and dark spots are cold.

spreading is explained by Ficks law, were J is the quantity of
matter flowing through a check point x.


this is beyond this classs scope
Heat spread
Do you get the intuition ?
Anisotropic approach

Wanted:
A form of smoothing specifically edges, with out loosing
significance.

If found:
Then it could be combined with the non-linear smoothing
idea, to produce an all around smoothing filter.

Remember main drawback of the non-linear approach
we saw earlier, was lack of intra-edge smoothing.

Ideas ?

Use varying smoothing kernels !
The desired direction of the dynamic kernel is effected by
the local gradients direction.




It is perpendicular to it

If gradient is (a,b),
then a perpendicular vector to it,
can be (a, -b)

A dynamic kernel is contracted along the direction of the
normal, ending in an elipticall kernel.

Original Anisotropic Diffusion
Original
Linear isotropic diffusion
(simple gaussian)
Original
Non-linear isotropic
diffusion (edge skipping)
Original
Non-linear anisotropic diffusion