Unit III Image Processing Techniques and Image Features

Introduction to Digital Image Processing

What doyou mean by digital image?

The digital image means 2D array of quantize intensity values or maybe a 2D array of
numbers. So digital image processing means how to manipulate these numbers.

Suppose, if I want to improve the visual quality of an image, then in this case, I have to
manipulate these numbers. These numbers mean the quantized intensity values. And suppose,
if I want to remove noises then also I have to manipulate these numbers. So, manipulation of
these numbers is mainly the image processing.

In case of the image acquisition, the light is converted into electrical signal. So, I am getting
analog signal, the analog image; the analog image can be converted into digital image by the
process of sampling and the quantization. So, I have to do the sampling along the x direction
and sampling along the y direction and that is called the spatial sampling.

After sampling, I have to do the quantization of the intensity values. So, to convert an analog
image into digital image, I have to do sampling and the quantization. So, what is digital
image, I will show you.
Digital image processing means processing of digital images on digital hardware, usually a
computer. So, this is the definition of digital image processing; that is the manipulation of the
quantize intensity values, the manipulation of the numbers.

Now, here I have shown a typical digital image processing sequence. So, if you see, the light
is coming from the source and it is reflected by the object and I have the imaging system.
So ,this imaging system converts the light photon into electrical signal.

And in this case, I am getting the analog image. So analog means it is a continuous function
of time. To convert the analog signal, the analog image into digital, I have to do the sampling
and the quantization. So, sampling in the x direction, sampling in the y direction and that is
called spatial sampling.

And after this, I have to do quantization of the intensity value. After this I am getting the
digital image. This digital image I can store into digital storage, I can process the image in the
digital computers, I can display the image, the processed image and I can store the image. So,
this is a typical digital image processing sequence.

The same thing I have shown here in this figure. So, image sensors are available for image
acquisition and after this, I will be getting the digital image and that digital image, I can
process by image processing hardware. And I have also image processing software, so the
process image can be displayed and it can be stored. This is the components of a general-
purpose image processing system.
The concept of image formation here, I have shown and the light is coming from the source
and it is reflected by the surface. So, I have shown the surface normal also. So, suppose, this
is my surface normal and I have the optics, optics is mainly the camera.

So, light is reflected by the surface and I am getting the image in the image plane, that is, the
sensors. So, sensor means it converts the light photon into electrical signal. This pixel
intensity value depends on the amount of light coming from the surface.

So, if I consider one pixel at this point, this pixel value, the pixel intensity value depends on
the amount of light reflected by the surface. So that means, that the pixel value actually
depends on the surface property.

The surface property is called the reflectance property of the surface, that is, the amount of
light reflected by the surface. So, this is a typical image formation system.
So, I have shown in this case, this is the electromagnetic spectrum. The source of illumination
I have shown the gamma rays, X-rays, ultraviolet, infrared ray, microwaves, FM radio. And if
you see the wavelengths from 400 to 700 nanometers, this is the visible range of the
spectrum, the electromagnetic spectrum.

So, based on this source of illumination, I may have these types of images; may be the digital
photo, the image sequence used for video broadcasting, and something like the multi-sensor
data like satellite image, visible image, infrared images, and an image of them obtained from
the microwave bands, medical images like ultrasound, gamma-ray images, X-ray images, and
the radio band images like the MRI images, astronomical images. So, based on this
electromagnetic spectrum, I may have these types of images.

So here, I have given some examples. The first one is the X-ray, and next one is the photo
that is obtained in the visible region of the electromagnetic spectrum, the next one is the
infrared image, another one is the radar image; so, this is the radar image; ultrasound image,
and the mammograms. So different types of images.

And some of the applications of image processing, here I have shown. The applications like
multimedia, medical imaging, that is, medical image processing and medical image analysis.
I have many applications even in the forensic in the biometrics, like the fingerprint
recognition, the iris recognition, video surveillance, remote sensing. There are many
applications of image processing.

And one important application of image processing is image fusion. If you see, I have
considered in the first example, I have two images, one is the MRI image another one is the
PET image. So, these two images are fused and I am getting the fuse image. The image fusion
process is defined as gathering all the important information from multiple images, and their inclusion into
fewer images, usually a single one. So that means the important information from MRI images
and the PET images, we have considered and the redundant information I am neglecting, I
am not considering. So, this is the fused image. In this fused image, I have more information
as compared to the original source images. So, I have two images, one is the MRI image,
another one is the PET image, and the fused image I am getting from these two images.

In a second example, I have considered the IR image and the image of the visible region of
the spectrum. These two images I am fusing and I am getting the fuse image. So, it has more
information as compared to the source images. And in this case, you will remember that I am
neglecting the redundant information, I am not considering the redundant information. So,
these are the applications of image processing.

So, what is analog image? Analog means it is a continuous function of time. This is the
definition of the analog image. So here I have shown one analog signal that is obtained from
the camera.
So, I have the horizontal lines and the vertical, the blanking signals; this is the, mainly the
video signal. So, this analog signal can be converted into digital signal by the process of
sampling and the quantization.

So, in TV standard, I know some of the standard like the PAL, the phase alternating by lines.
So, this is one TV standard. Another standard is you know, NTSC. This is another standard.

Phase Alternating Line (PAL) is a colour encoding system for

analogue television. It was one of three major analogue colour
television standards, the others being NTSC and SECAM. In most
countries it was broadcast at 625 lines, 50 fields (25 frames) per
second, and associated with CCIR analogue broadcast television
systems B, D, G, H, I or K.

So, in the PAL system, we have 625 lines; 625 lines. And in this case, we have considered the
25 frames per second; 25 frames per second. In NTSC, generally we use the 30 frames
per second. So, I am not going to discuss about theTV standard.

Now, let us consider, suppose a video signal, a video signal I am considering and suppose the
bandwidth of the video signal, bandwidth of video signal I am considering, suppose it is 4
megahertz. Let us consider this example, 4 megahertz. So, what will be my sampling
frequency then?

Bandwidth is 4 megahertz, so sampling frequency will be, sampling frequency, sampling

frequency will be my 8 megahertz, as per the Nyquist sampling theorem, so my sampling
frequency will be 8 megahertz.
Now, I want to count the number of pixels per frame. So, I want to count the number of
pixels per frame. So how many pixels per frame, that I want to determine. So, 8 megahertz,
8 × 106 , that is samples per seconds, samples per seconds; divided by, I am considering 25
frames per second, 25 frames per second. So, this will be equal to 320000. This is the digital
pixels per frame, pixels per frame, pixels per frame. So, this is approximately the 512×512.

So that means, the image size will be 512 ×512. The image is represented by M × N. This is
the size of the image. The image is represented by M × N, this is the image because it is a 2D
array of numbers. So, I have M number of rows and N number of columns. So, my rows are
like this, these are my rows, rows and I have the columns, like this.

So, the image is represented by M × N. So, in this example, I have seen that the size of the
image will be 512 × 512. So, this I have shown this example.

Now, this already I have explained how to get the digital image. So, in this diagram, you have
seen that I have shown one continuous tone image that is the analog image and after this, I
have to do the sampling; the sampling along the x direction and the sampling along the y
direction I have to do.

And after this, I have to do the quantization; the quantization of the intensity values. So, I
have done the quantization and after, this I am getting the 2D array of numbers. So, this is a
sample and the quantized image I am getting. So, this is a representation of a digital image.
The same thing I am showing the sampling and the quantization. So, sampling along the x
direction and the sampling along the y direction I have shown and after this, I am doing the
quantization of the intensity value.

So, sampling the 2D space on a regular grid and after this I have to quantize each sample
value. First, I have to do the sampling and after this I have to do the quantization. So, this this
is first I am doing the sampling and after this, I am doing the quantization. So, it is 2D array
of numbers and this is the digital image.

So, I have shown the values. The value is the 0 is the grayscale value. The 34 is the grayscale
value, this 102 is the pixel value. So, I am considering this is one digital image.
And in this case, already I have shown the image is represented by M × N. M × N is the size
of the image. So, M × N is the size of the image. So here I have shown the rows. The M
number of rows I am considering and the N number of columns starting from 0 to N −¿1 and
0 to M −¿ 1.

And this is the origin, the origin of the image, that is, the point is (0,0). And I have shown the
pixels; these are the pixels, the pixels of the image. These are the pixels of the image.
Now, let us consider the neighborhood of a pixel. So, in this figure, I have shown and a
particular suppose pixel, the pixel is P. And this pixel I am considering P, I have shown two
cases. One is the neighbourhood, the 8 neighborhood I have considered.

Corresponding to the first case, I am considering the neighborhood pixel the P1, P2, P3, P4,
P5, P6, P7, P8, that is called the 8 connected neighborhood. In the second case, if you see the
neighborhood pixel I am considering P2, P4, P6, P8. Corresponding to the center pixel, the
center pixel is x , y; this is center pixel.

So, if you see here, the first one is the 8-connected neighbourhood, this is the 8-connected
neighbourhood, 8-connected neighbourhood. The second one is the 4-connected
neighbourhood, the 4-connected neighborhood. So, this 4-connected neighborhood I can
write like this N4 P, 4-connected neighborhood.

So, this pixel I have to consider, this pixel I have to consider, this pixel I have to consider,
this pixel I have to consider that means it will be x, y - 1, that is the one pixel; {(x, y + 1), (x
+1, y); (x - 1, y)}. So, this is the 4-connected neighborhood.

Similarly, you can define the 8-connected neighborhood. So, in the, this is the neighborhood
of the 4-connected neighborhood, the second one is the 8-connected neighborhood.

Now, you can see, corresponding to this neighborhood, I can determine the Euclidean
distance, the city block distance, and also the chessboard distance. So how to define the
distances?
This Euclidean distance, it is defined like this. Euclidean distance, Euclidean distance is
suppose dE; Euclidean distance is dE. So, suppose I am considering two points (i, j) and (k, l).
1
So, [( i−k )2 +( j−l )2 ]2 ,this is the Euclidean distance.

Another distance is L1 distance, that is called the city block distance, city block distance. The
city block distance I can show it is suppose, the d4 distance, city block distance suppose. So,
it is the distance between (i, j) and (k, l). So, it is¿ i−k𝗏+¿ j−l𝗏¿, this distance is called the
L1 distance. It is the City block distance or the, it is called the L1 distance, L1 distance.

Another distance is the chessboard distance or sometimes it is called a Chebyshev distance.

So, I can define suppose the it is d8 between two pixels (i and j), (k, l). So, I have to consider
this distance; the distance is maximum value I have to determine, ¿ i−k𝗏+¿ j−l𝗏¿.

So, these distances I have defined. One is the Euclidean distance, one is the city block
distance, another one is the chessboard distance.

So, corresponding to these distances you can see, in the first figure, this figure, I have
determined the distance between the center pixel, the center pixel is this pixel and the
neighborhood pixels. You can see the distance is 1.41 to the diagonal pixel and to this pixel it
is the distance is 1, to this pixel it is 1.41; like this I am getting the distances. All the
distances I can compute to the neighborhood pixel.

Similarly, if I consider the city block distance corresponding to the center pixel, the center
pixel is this, I can find a distance between the neighborhood pixel. So, if I find the distance
between the center pixel and this pixel the distance will be 2. This distance will be 1, this
distance will be 2, like this, I am determining the distance. I am calculating the distance.

In the third case, I am considering the chessboard distance. So, corresponding to the center
pixel, this center pixel, I am determining the distance between the neighborhood pixels and in
this case, I am getting 1 1 1 1, like this distance I am getting. So, you can see these distances I
can compute.
So I can give one example. Suppose, if I want to determine the distance between d, two
points A and B, that is, the, suppose the Euclidean distance I can determine between two

points, √( x A − x B)2 +( y A − y B )2 .

So suppose this xA is some, xA is suppose 70 minus 330 suppose xB is 330, I am taking one
example, whole square plus suppose this yA is 40 and yB is 228 suppose whole square. So,
this value I am getting it is 335 almost 335 this value. This value is 335. This is the Euclidean
distance I can determine.

And the L1 distance, if you can compute L1 distance, then this distance will be d (A, B) will
be you can see this is∣x A − x B∣+¿ y A − y B 𝗏¿. So, you will be getting the value something like
448. Also, you can determine the chessboard distance, also you can determine.

The Chessboard distance d (A, B) between two points, thus maximum you have to take and it
is max ¿, ¿ y A − yB𝗏¿). So, this value you will be getting something like 260. You can verify
this one.

So, in this example, I have shown how to calculate the Euclidean distance, city block
distance, and the chessboard distance. So, this is the important the neighborhood of the pixel
a particular pixel.
So, what is digital image? The digital image is nothing but the 2D array of numbers
representing the sample version of an image. The image defined over a grid and each grid
location is called the pixels.

So, I have shown the pixels, these are the pixels, and represented by a finite grid and its
intensity data is represented by a finite number of bits. So that concept I can explain like this.

Suppose, the image is represented like this M × N; the size of the image is M × N, that is, M
means the number of rows, N number of columns. And the grayscale value suppose, I am
considering L number of grayscale values. So, 0 to L−1, grayscale value that can be
normalized between 0 to 1.

Suppose, the L number of grayscale levels, L is equal to 2K. So now, the number of bits,
number of bits required to store an image is equal to M × N × K. That is the number of bits
required to store an image is equal to M × N × K.

Now, what is the definition of the dynamic range? Suppose, I have an image, the highest
pixel value and a lowest pixel value I know. So, the difference between this is called a
dynamic range. The contrast of an image depends on the dynamic range. So, for a good
contrast image, the dynamic range will be more.

Now, the resolution of an image, resolution of an image. So how to define the resolution of
the image? So, one is, one resolution is the spatial resolution, the spatial resolution. The
spatial resolution means, I can say the number of pairs of lines per unit distance. This is the
definition of the spatial resolution.

And another one is the intensity resolution. So, intensity resolution, I can say that the
definition is smallest discernible, smallest discernible changes in the intensity level. So,

suppose if I consider 8 bits, so, 28=256. That means, how many intensity levels I am using?
The 256 numbers of intensity level I am using.

So spatial resolution means the number of pairs of lines per unit distance. That means, so how
many pixels, if I consider these are the pixels, so how many pixels per unit distance that is
called spatial resolutions.

And intensity resolution means the smallest discernible changes in the intensity level. So how
many intensity levels I am using that is the intensity resolution. And if I consider the video, in
video I have to consider, suppose, if I consider a video, in video I have to consider how many
frames, frames per second? How many frames per second that corresponds to the temporal
resolution.

So, for video, I have temporal resolution and for the image, I have spatial resolution and the
intensity resolution. And for the binary image, the binary image is represented by 1 bit. So
that concept I can show you.
(Refer Slide Time: 26:22)

So here, I have shown the pixels and the intensities. This is the two dimensional array of
numbers and here, I have shown the pixel value 255, 18, 19. So these are the pixel values.

(Refer Slide Time: 26:57)

So already I have defined the resolution of an image, one is spatial resolution, another one
intensity resolutions. So, suppose if I consider, suppose 0 to 255 levels; 0 to 255 levels. So
how many, how many bits we need?

For this, we need 8 bits. And if I consider 0 to 127 then I need 7 bits; if I consider 0 to 63
then in this case, I need 6 bit; if I consider 0 to 31 then in this case, I need 5 bits; if I consider
0 to 15 then in this case, I will need 4 bits; if I consider 0 to 7 then I need 3 bits; if I consider
0 to 3, I need 2 bits; and if I consider 0 to 1, only two levels, 0 to 1 means two levels, so I
need only one bit.

So, this is called the binary image, binary image. So, this is the binary image. So binary
image needs only 1 bit. So, a binary image is represented by1 bit and a grayscale, gray-level
image is represented by 8 bits.

(Refer Slide Time: 28:19)

And the image is represented by f (x, y). What is f (x, y)? f (x, y) means the intensity at a
position the position is (x, y). This is the definition of image. The image is represented by f
(x, y).
So here, I have shown the image as a function because image is nothing but this f (x, y); f (x,
y). So, the intensity value I have shown f (x, y) here.

And there is another definition of a digital image. In this definition, I have shown the rows
and the columns. So, I have shown the pixels. So, in this definition, if you see, the image is
represented by these parameters, one is x and y coordinate; that is the spatial coordinate. And
if I consider the stereo image, then one important parameter is the depth information. So that
corresponds to the depth information.

So, suppose if I have two images obtained by two cameras, so from two images, I can
determine the depth information. So, in this case, the z represents the depth information. And
if I consider the color image, so lambda means the color information. And if I considered a
video, video means the number of frames per second. So I have to consider temporal
information, so t is the temporal information.

So image is represented by x, y, z, lambda, t; x and y is the spatial coordinate, z is the depth

information, lambda is the color information, and t is the temporal information. That is one
definition of digital image.
And some of the typical values, number of rows, number of columns. The 256, 512, 525, 625
like this, these are the number of rows; number of columns 256, 512; and gray-level values
something like 2, 64, 256 like this, we have the values.

And in this case, I have shown the dimensionality and resolution of an image. So, in this case,
I have shown the image, the image is 2 dimension and another case, I have shown 3
dimensions that is for the video, x and y coordinate and another one is number of frames per
second, that is, the temporal information.
So here, I have shown that dimensionality and the resolution of an image. So. image is 2
dimension and video is 3 dimension.

So in this example, I have shown, I have changed the spatial resolution. If you see the first
image, the first image is 256 × 256 pixels; the second image is 128 × 128, this image; the
third image is 64 × 64.

So that means I am decreasing spatial resolution and because of this, I am getting some
effects like checkerboard effect. This effect I have observed here and this is called a
checkerboard effect because I am decreasing the spatial resolution of an image.
Now, I am considering changing the intensity resolution that is called the gray-level
resolution. So, first example is I am considering the 8 bit image, next one is the 4 bit image,
next one is 2 bit image, and 1 bit image is nothing but the binary image. So I am changing the
gray-level resolution the intensity resolution.

This is another example. The initial, if you see the first image, this is 16 gray levels, second
one is the 8 gray levels, the next one is 4 gray levels, and this is the 2 gray levels that means
the binary image. So, if you see this, this portion, this portion, suppose if I consider, this
portion is not properly represented in these images, in these images. So, this effect is called
the false contouring effect.

So, I have two effects one is the checkerboard effect that is because of the spatial resolution,
and another one is a false contouring effect because of insufficient number of intensity levels.
So, I have given these two examples.
The image types I have shown here. I have RGB image, the grayscale image, and the black
and white image that is the binary image. So, in case of the color image, corresponding to a
particular pixel, I have 3 components; one is the R component that is the red, another one is
the G component that is the green; and the blue component that is the B.

So RGB value that is the primary colors. So, corresponding to a particular pixel, I have 3
components; R component, G component, and the Blue component. And this pixel, particular
pixel is called the vector pixel, because I have 3 components.
So, this is one example of the binary image, basically the binary is 1 and 0 only. So if I apply
the edge detection principle, the edge detection principle or edge detection, then in this case, I
will be getting the binary image. So, this is one example of the binary image.

The next example is the grayscale image. So, in this case, I am considering 256 numbers of
intensity levels and if you see, it is the grayscale image.

And after this, I have considered the color image. So, this is the color image because this is
called the vector pixels. In the vector pixels, I have three components; R component, G
component, and the blue component.
Now, I have different types of color models. I will discuss about the color models like RGB
color model, YCbCr color model, HIS color model. So, all these color models I will discuss
in the color image processing. Because I need 8 bits for the R components, 8 bits for the G
component, 8 bits for the B component, so it is called 24 bit video, that is the color image.

(Refer Slide Time: 34:02)

And this is the example of the true color image, here you see corresponding to a particular
pixel, I have the red pixels, the green pixels, and the blue pixels. So that means, pixel means
the red information, green information, the blue information corresponding to a particular
pixel. The red value, the green value and the blue value.
Another definition of indexed image. So, for the indexed image, I have one color map. So,
this is a color map. So, this is something like a lookup table, lookup table. So, corresponding
to these value, we have three values; one is the R value, this is R value, like G value, and a B
value; three values are available.

So, corresponding to this one, the corresponding to these RGB value, the index number is 6.
Like this, I have all the index numbers. So, this is the indexed image, if you see. So, I need a
color map and in the color map, you see the index number is associated with the color values,
R value, G value, and the B values. So, corresponding to 6, I have this RGB values.

So, types of digital image. So binary image, 1 bit per pixels; grayscale image, 8 bit per pixels;
true color or the RGB image, it is 24 bit per pixels; and for indexed image, 8 bit per pixels.
So, this is the case.
So next one is what is image processing? The image processing is that manipulation and
analysis of the digital image by digital hardware. And it is for the better clarity that is for the
human interpretations and automatic machine processing of the scene data. So that is image
processing.

So, I have given some examples here. So, one is image enhancement. Image enhancement
means to improve the visual quality of an image. So first if you see, I am just the compare the
two images, the input image and the output image. You can see that I am improving the
visual quality of the image.
So, some of the examples like I can increase the contrast of an image, I can improve the
brightness of an image. Like this, I can do many things, that is one example.

In this example, another example, I am doing the image enhancement. So first one is the
input image, the second one is the better-quality image that is the enhanced image.

In this example, I am considering the blurred image. So blurred image maybe because of
some motion blur. Suppose, the camera is moving, then in this case, I am getting the blurred
image. And in this case, I can apply some image restoration technique to deblur the image.
So, in this example, I have the blurred image and I can apply some image restoration
technique to get the deblur image.

In this example, I have considered a noisy image and after this, I am getting the filtered
image, a better-quality image.

Now, the image processing, what is actually image processing now I can show. Because
image is represented by f (x, y), what is f (x, y)? Intensity at the point, the point is x, y. Now,
in this case, I have g (x, y); g( x, y) is the output image and f (x, y) is the input image. So, I
can change the range of an image or I can change the domain of an image.
What do you mean by range of an image? The range of an image means, I can change the
intensity value, the pixel value I can change by using some transformation. Also, I can
change the domain of an image. Domain of an image means, I can change the spatial
positions; x and y coordinate I can change.

One is just changing the range of an image and another one is changing the domain of an
image. So, in this example, here you see, I am changing the range of the image. That means I
am changing the intensity value. In the second case I am changing the spatial coordinates.

So, I have shown the two examples. The first one is the input image. This is the input image
and I am changing the range of an image that is the pixel values.

In a second example is the input image and I am changing the domain of the image. In this
case I am doing the scaling of the image. So that means, this is the changing the domain of
the image and these are changing of the range of an image.
This is one transformation, here if you see I am doing some mapping. This pixel is mapped to
this point, this pixel is mapped to this point. So, this is my, this is my input image and this is
my output image. So, I can do the mapping, that means, I am changing the spatial
coordinates.

I can do rotation on an image, this is the rotation of an image. I can do the scaling of an
image. That means, I can change the domain of an image, I can say the spatial position of the
pixels.
Now, the linear point operation on an image. So, I can give one example here. Here you see g
(n) is the output image, f (n) is the input image. And in this case, I am considering one
operation, the operation is h, that is called a point operation, the linear point operation. The
point operation is mainly the memoryless, we do not need any memory.

Now, in this case, g (n) is equal to I can consider g (n) is the output image, P× f (n)+ K . So,
if I consider this expression, K is the offset and P is the scaling factor. Now, based on this the
point operation, how to get that negative image?

So ,in this case if i consider P is equal to minus one and K is equal to L minus 1, L means the
number of gray level values, how many levels I am considering, from 0 to L minus 1. So,

output is g( n)− f (n)+ L−1.

So, if I consider this transformation, then in this case, I will be getting the negative image. So,
this is my input image and then this is my output image.
Also, I can apply the nonlinear point operation, maybe the logarithmic operation I can apply.
point operation. So, this is about the nonlinear point operation and the linear point operations
on image.

Here, I am giving one example, the addition of images. So, n number of images I am
considering. After this, I am doing the addition, addition of n number of images. The second
example is the subtraction of two images.
If I do addition of the images what will happen, you see I have the noisy image. Suppose I
have number of noisy frames. If I do the averaging, then in this case, the noise can be
reduced, image averaging for noise reduction.

So just I am adding this one that means, I am actually averaging the images, that is, I am
doing to sum of n input images and corresponding to this I am getting this this example. So,
in this example, the image averaging for noise reduction.
And already I have explained that how to find a difference between two images. So, one
example is, suppose if I consider these two images, the first image is this. In this case, the
background is available; in the second case the background and the foreground is available.

Suppose the moving objects are available in the image, in the first case, there is no moving
objects. And in the second case, suppose, the background is also there, the moving object is
there.

So, in this case, if I do the subtraction of these images from the first image to second image,
then in this case, I will be getting the foreground, something like this, moving objects I can
determine. That is called a change detection.

So, this is called a principle of change detections. Thus, I have to subtract the images then in
this case I will be getting the moving objects in the image.
Like this, we can do some geometric image operations. So, I can do image translation like
this, I can change the coordinate; coordinate is n1, another one is n2. So, I can do the
translation of the image. This the coordinate is x and y actually, in this case, I am showing n1
and n2. So n1 minus b1 and n2 minus b2, that means I am doing the image translation.

The second example is also if you see, the coordinate n1 is divided by c, the coordinate n2 is
divided by d. So that means I can zoom the image, zoom in and zoom out I can do. I will
explain this one zoom later on.

So, this is also a geometric image operation and this is one example of the image rotation that

[ sinθ ]
cosθ −sinθ
I can rotate the image. So, for the rotation, the transformation matrix is
cosθ

. So, by using this transformation matrix you can rotate an image.

Here I have, in this example I have shown the zoom. After zooming, I have done some
interpolation, nearest neighbour interpolation it is done. The second case is the bilinear
interpolation that is more accurate. So, I have shown two interpolation techniques, one is the
nearest neighbour interpolation, another one is the bilinear interpolation after zooming.

So, image processing steps I have. One is the image acquisition. After this, I have to get the
digital image. So, for this, I have to do the sampling and the quantization. Compression is
important because if I considered the raw image or the raw video, that is very bulky. So that
is why for storing the image, for storing the video, or for the transmission of the image or for
the transmission of the video, I have to do image compression, I have to do video
compression.

Next one is the image enhancement. Image enhancement is to improve the visual quality of
an image. So, one example is, I can improve the contrast of an image or maybe I can improve
the brightness of an image. That is image enhancement.

And restoration is also the meaning is same. To improve the visual quality of an image. So,
there is a basic difference between enhancement and the restoration. In image enhancement
that is mainly the subjective. It depends on the subjective preference of an observer.

So, I can say, the image is not good, the contrast is not good, and based on my observation I
can improve the visual quality of an image. That is subjective and there is no mathematical
theory for image enhancement.

But in case of the image restoration, I have mathematical models. So, I can give one or two
examples like this. Suppose I am taking some image and camera is moving, then in this case,
I am getting the blurred image. So, in this case, I can deblur the image by using the
restoration principle, the image restoration principle.

So, for this I have to develop a mathematical model and based on this mathematical model, I
can improve the visual quality of the image. So, for this, I have to apply image restoration.
That is image restoration is objective and image enhancement is subjective, the subjective
preference of an observer.

I can give another example of the image restoration. Suppose the object is not properly
focused then in this case, I am getting the blurred image. Then also I can improve the visual
quality of the image. I can de-blur the image.

And suppose if I take one image in the foggy environment, then in this case, I am also getting
not good quality image. Then by using this image enhancement principle or the restoration
principle, I can improve the visual quality of an image. So, you see the fundamental
difference between image enhancement and the restoration.

In image enhancement there is no mathematical model, general model but in image

restoration, I can develop some mathematical model and by using this mathematical model, I
can improve the visual quality of an image.
Next one is image segmentation. The image segmentation is the partitioning of an image into
connected homogeneous region. So, I can later on, I will discuss about the image
segmentation. So, that means I am doing the partitioning of an image into connected
homogeneous region.

And after this, for computer vision, I have to extract features. And after this, based on the
features, I have to go for recognition, the object recognition that is called the image
interpretation or the understanding.

So, this is a typical image acquisition systems. So, we have the CCD, the search coupled
device camera. Here you see the CCD chips. So the photon is converted into electrical signal.
After this, I am getting the video signals, that is, the analog signals. The analog signal can be
converted into digital, already I have explained about this.
So, this is the image sensor. The light photon is converted into electrical signal and after this,
we have to do the sampling and the whole. And after this, you can convert the analog signal
into digital image.

So, these are the, for the still image, these are the 512 ×512 some typical values. 256 × 256,
one typical values. For video like this 720 × 480, 1024 × 768 that is the high definition TV,
like this. And regarding the intensity resolution, 8 bits is sufficient for all but the best
applications needs 10 bits, television, production, printing.
And for medical imaging generally, the 12 to 16 bits images we consider. So, this is the case
and one important application is the medical imaging. So, we need more number of bits we
need more number of intensity levels.

And if you see why we need the compression. Suppose a raw video is very bulky. So suppose
in this example, if I considered a digital video, uncompressed video, 1024 × 768 that is the
size of the image frame.

Suppose if I consider 24 bit per pixel, why it is 24 per pixel? That is the, if I consider the
color, color image that is the 24 bit per pixels and 25 frames per second. That means 472
Mbps. So that is very difficult to store and also for transmission. So, that is why we have to
compress the image, the compression of the image is important in this case. So, I have given
this example.
And image enhancement, this already I have explained that to improve the visual quality of
an image, like this we can improve the contrast, we can sharpen the edges, we can remove the
noise, the image enhancement.

This is already I have explained. One is enhancement; restoration, image restoration, I have
already explained. And reconstruction, mainly we have the 2D images. So, if I want to get the
3D information, the reconstruction of the 3D information from the 2D projections, that is
called a reconstruction.
This concept I am going to explain maybe after 2, 3 lectures that is image reconstruction from
2D projections. That concept I will explain the radon transform. Radon transform I will
explain later on. So, construction of 3D information from the 2D projections that is used in
CT-scan, the computer tomography. So, this concept I am going to explain when I will
discuss the radon transform.

After this, we have to extract the features. Image features I have to extracts, maybe something
like the color feature, I can consider the texture, I can consider shape features, motion
features I can consider. And finally, we will go for the recognition or the classification of the
objects, that is the machine learning. This is the machine learning, feature extraction and
recognitions.

So, these are some examples, the image filtering. Original image, the noisy image, and the
filtered image. So, this is one example.
(Refer Slide Time: 50:36)

Second example is the image restoration. Degraded image after processing I am getting the
restored image. That is a very good quality image, this is the noisy image.

And for contrast improvement, there is a technique. This technique also I am going to explain
later on. That is by histogram equalization technique, I can improve the visual quality of an
image. That means I can improve the contrast of an image. So, this concept I am going to
explain when I will discuss the image enhancement principles.
This is one example of the contrast enhancement. This is not a good contrast image, this is
not a good contrast image but I think this is also good contrast image I can say. So, by using
the histogram equalization technique, I can improve the contrast of an image.

And after this, I have to extract features. The features may be something like edges I can
consider, the boundary of an objects I can consider. So, I can extract many, many features
from the image or maybe from the video.
And after feature extraction this one example is the edge detection technique. So, this is my
input image and I am extracting the edges of the image.

(Refer Slide Time: 51:45)

And the segmentation is the partitioning of an image into connected homogeneous region.
This homogeneity may be defined in terms of the gray value, color value, texture value, shape
information, and the motion information. So, based on this principle, based on these values, I
can define the homogeneity.
So again, I am going to discuss about the segmentation principle. So, suppose in my next
classes, so suppose if I want to do the partitioning of an image, so this is one partition, this is
one partition, this is one partition.

So, this portion is homogeneous, this portion is also homogeneous. If I consider this, portion
is homogeneous, this portion is homogeneous. Like this, this portion is also homogeneous.
So, homogeneity I can define in terms of gray value, color, texture, like this, I can define.

So, corresponding to this region, the gray value is almost same or maybe the color may be
same. So, like this I can do the partitioning of an image.

So, this is one example of the segmentation, image segmentation. This is my image, input
image and after this, I am doing the segmentation.
And finally, after feature extraction, I will go for the object recognition. So, this is my last
step that is I have to extract the features. The features maybe the color is a features, texture
maybe the features. So, I can extract different types of features from the image and after this,
I will go for the object recognition.

So first one is the feature extraction, feature model matching, hypothesis formation, and
object verification, I have to do this.
And finally, I will go for the image understanding. So maybe I can apply some supervised
technique and that is the part of the artificial intelligence or maybe the machine learning. So,
I am going to discuss about this in my last lecture the machine learning. So, this is the final
step the image understanding.

So, in this class, I have discussed about the concept of image processing, what is the meaning
of the digital image. And after this, I have discussed about the types of digital image, the
RGB image, the grayscale image, the binary image, index image. So, I have discussed about
this.

And also, the concept of the resolution, resolution of an image. So first, I discussed about the
spatial resolution, after this, I discussed the intensity resolution. After this, I discussed about
the image processing techniques. So, I can change the domain of an image, I can change the
range of an image. So, I have given some examples I can do the rotation of an image, I can do
the scaling of an image. Like this I can do many operations, I can do zooming also.

After this, I have considered the typical image processing system, the computer vision
systems. So first, I have to do some image processing and in this case, before image
processing, I have to get the image. The analog image can be converted into digital image by
sampling and the quantization.

After this, I can improve the visual quality of the image by using the enhancement and the
restoration techniques. And after this, I can extract the features from the image. And finally, I
will go for object recognition. That is the pattern classification or the machine learning
techniques I can apply.