Introduction to Digital Image Processing Using MATLAB

Raymond Phan – Ph.D. Candidate
Department of Electrical and Computer Engineering
Distributed Multimedia Computing Research Lab
EPH 408
rphan@ee.ryerson.ca
Some slides were taken from Prof. R. A. Peters PPT slides:
http://www.archive.org/details/Lectures_on_Image_Processing

Topics Covered in this Presentation
 1st Hour: 6:10 p.m. – 7:00 p.m.
Introduction to Digital Images and in MATLAB
 Quick intro to myself
 What is MATLAB?
 Where to get MATLAB and how to run
 Basic I/O  Reading and writing images
 Accessing pixels and groups of pixels
 Resizing Images
 Rotating Images
 …break for 10 minutes!

 2nd Hour: 7:10 p.m. – 8:00 p.m.
Operations on Digital Images
 Simple contrast and brightness enhancing
 Intro to image histograms
 Advanced enhancing using image histograms
 Intro to convolution in images
 Blurring / Smoothing images
 Edge detection
 Sharpening images

 3rd Hour: 8:10 p.m. – 9:00 p.m.
Applications of Image Processing
 Segmenting simple objects
 Noise Filtering
 Simple image stitching using template matching
 NOTE!
 For a more comprehensive MATLAB tutorial, check:
http://www.ee.ryerson.ca/~rphan/ele532/MATLABTutorial.ppt
 You can access the slides, images and code at:
http://www.rnet.ryerson.ca/~rphan/IEEEDIPTalk

 1st Hour: 6:10 p.m. – 7:00 p.m.
Introduction to Digital Images and in MATLAB
 Quick intro to myself
 What is MATLAB and where do I get it / find it?
 Intro to Digital Images
 Basic I/O  Reading and writing images
 Accessing pixels and groups of pixels
 Resizing Images
 Rotating Images

Introduction of Myself
 Started in 2002 in the B.Eng. – Computer Engineering
program
 Program only started at this time  Relatively new
 Graduated in 2006
 Started my M.A.Sc. in 2006 – ELCE
 Finished in 2008
 Winner – Ryerson Gold Medal (2008 – SGS)
 Started my Ph.D. in 2008 – ELCE
 Will finish before my back gives out
 Currently a 4th year Ph.D. Candidate
 2010 NSERC Vanier Canada Graduate Scholar

Introduction of Myself (2)
 Research Interests
 Digital Image Processing, Signal Processing,
Multimedia, Computer Vision, Stereo Vision, 3DTV, etc.
 M.A.Sc. Thesis – Content-Based Image Retrieval
System
 Featured in the Toronto Star (January 4th, 2008)
 Searching for images using actual images, rather than
keywords
 Current Ph.D. Thesis – Faster and more accurate 2D to
3D conversion

What is MATLAB?
 MATLAB  Stands for MATrix LABoratory
 Created by Cleve Moler @ Stanford U. in 1970
 Why? Makes linear algebra, numerical analysis and
optimization a lot easier
 MATLAB is a dynamically typed language
 Means that you do not have to declare any variables
 All you need to do is initialize them and they are created
 MATLAB treats all variables as matrices
 Scalar – 1 x 1 matrix. Vector – 1 x N or N x 1 matrix
 Why? Makes calculations a lot faster (will see later)

What is MATLAB? (2)
 How can I get and/or use MATLAB?  3 ways:
 1) Find it on any of the ELCE departmental computers
 Log in to your account, go to Applications  Math 
MATLAB R2010b
 2) Use any Ryerson computer on to the ACS network
 Log in with your Matrix ID and Password, then go to Start 
MATLAB R2010b
 3) Install it on your own laptop
 Go to http://www.ee.ryerson.ca/matlab for more details
 You must be on the Ryerson network to sign up for an account
 After, you can download MATLAB from anywhere

 MATLAB uses the Image Processing Toolbox (IPT)
 Should already be installed with MATLAB!

Intro to Digital Images
Color images have 3 values per pixel;
Digital Image monochrome / grayscale images = 1
value/pixel.

a grid of squares,
each of which
contains a single
color

each square is
called a pixel (or
picture element)

Intro to Digital Images – (2)
Pixels
 A digital image, I, is a mapping from a 2D grid of
uniformly spaced discrete points, {p = (r,c)}, into a set
of positive integer values, {I( p)}, or a set of vector
values, e.g., {[R G B]T(p)}.
 Each column location of each row in I has a value
 The pair (p, I(p)) is a “pixel” (for picture element)
 p = (r,c) pixel location indexed by row r & column c
 I(p) = I(r,c)  Value of the pixel at location p
 If I(p) is a single number  I is monochrome (B&W)
 If I(p) is a 3 element vector  I is a colour (RGB) image

 Monochromatic Case:
 We call the values at each pixel intensities
 Smaller intensities denote a darker pixel
 Bigger intensities denote a lighter pixel
 Colour Case:
 Think of a colour image as a 3D matrix  First layer is
red, second layer is green, third layer is blue
 Why RGB? Trichromacy theory  All colours found in
nature can naturally be decomposed into Red, Green
and Blue  This is basically how CCD cameras work!
 The three element vector tells you how much red, green
and blue the pixel is compromised of (i.e. [R G B]T = [0
255 0]  No red, no blue, all green

Pixel : [ p, I(p)]

p = (r, c )  red  12 
Pixel Location: p = (r , c)
= (row # , col # ) I ( p ) = green  = 43
   
Pixel Value: I(p) = I(r , c) = (272, 277)  blue  61
   

 How do digital cameras take images (very basic)?
 Uses sampling and quantization
 What we see now through our eyes is continuous
 There is essentially an infinite amount of points that
comprise our field of view (FoV)
 Not good, because we want to store this information!
 We first need to sample the FoV  Transfer the FoV to
a rectangular grid, and grab the colour in each location
of the grid

Sampling and Quantization
pixel grid column index
row index

real image sampled quantized sampled &
quantized

 We’re not done yet! There are also an infinite number
of possible colours
 We will now need to quantize the colours
 Quantizing will reduce the total number of colours to a
smaller amount
 Key  Quantize accurately so that we can’t tell much
difference between the original image and the quantized
one
 A digital image is essentially taking our FoV and
performing a sampling and quantization
 Values are now discrete and positive

Sampling and Quantization
pixel grid column index
row index

real image sampled quantized sampled &
quantized

 Digital images store their intensities / colour values as
discrete and positive values
 Usually, digital images need 8 bits for B & W and 24-
bits for colour (8 bits for each primary colour)
 B & W – 0 for Black and 255 for White  All integers
 Colour – 0 to 255 for Red, Green and Blue  All integers
 Note: We can consider a colour image as three 2D images
 Without compression, files would be very large!
 Compression algorithms (PNG, JPEG, etc.) eliminate
extra information to reduce the size of the image

R/W Images in MATLAB
 So we have an image file… how do I access the info?
 Open up MATLAB and change working directory to
where image is stored
 Use the imread() function
 im = imread(‘name_of_image.ext’)
 Use single quotes, and type in the full name of the image
with its extension (.bmp, .jpg, etc.)
 im will contain a 2D matrix (rows x cols) of B&W values
or a 3D matrix (rows x cols x 3) of colour values
 Matrix corresponds to each pixel in the digital image for
B & W, or a colour component of a pixel in colour

R/W Images in MATLAB – (2)
 How do I access a pixel in MATLAB  B&W case?
 pix = im(row,col);
 row & col: Row & column of the pixel to access
 pix contains the intensity value
 Access elements in an array by round braces, not square!
 For you C buffs  Indexing starts at 1, not 0!
 How do I access a pixel in MATLAB  Colour case?
 pix = im(row,col,1);  Red colour value
 pix = im(row,col,2);  Green colour value
 pix = im(row,col,3);  Blue colour value
 3rd argument  3rd dimension of matrix
 Only grabs one colour value at a time!

 How can I get the RGB pixel entirely? Use the :
command
 pix = im(row,col,:);
 : means to grab all values of one dimension
 However, this will give you a 1 x 1 x 3 matrix… we just
want an array! Call the squeeze() command
 pix = squeeze(im(row,col,:));
 Now a 3 x 1 vector. To access R, G and B values, do:
 red = pix(1)  Red, gr = pix(2)  Green,
blue = pix(3)  Blue

 So I know how to get pixels; how can I modify them in
the image?
 Easy! Just go backwards
 For a B & W Image do:
im(row,col) = pix;
 For a colour image, do either:
im(row,col,1) = red;
im(row,col,2) = green;
im(row,col,3) = blue; or
im(row,col,:) = [red; green; blue] or
im(row,col,:) = rgb; %rgb - 3 x 1 vector

 How do I access a subset of the image?
 How do I grab a portion of the image and store it into
another variable?
 Do the following for monochromatic images:
im2 = im(row1:row2,col1:col2);
 Do the following for colour images:
im2 = im(row1:row2,col1:col2,:);
 This will grab a rectangular region between rows 1 and
2, and columns 1 and 2
 e.g., if I wanted to get rows 17 – 31, and columns 32 – 45
for colour, do: im2 = im(17:31,32:45,:);

 So I know how to get pixels; how can I display images?
 Use the imshow() command
 imshow(im);  im: Image loaded into MATLAB
 Shows a new window with the image in it
 If you do: imshow(im,[])
 For monochromatic: Smallest intensity becomes 0 and
largest intensity becomes 255 for display
 For colour: Apply the above for each colour channel
 Every time you use imshow, the image you want to
display is put in the same window… so what do you do?

 Use figure command to create a new blank window
 Then, run the imshow command to display the image
on the other window
 We can also do:
 imshow(im(:,:,1));  Shows red channel
 imshow(im(:,:,2));  Shows green channel
 imshow(im(:,:,3));  Shows blue channel
 (:,:,1) means grab all of the rows and columns for the
first layer (i.e. red), etc.

 When showing the red channel:
 Darker pixels mean there isn’t much red in that pixel
 Lighter pixels mean there is a lot of red in that pixel
 Same applies for green and blue!
 How do I save images to disk? Use imwrite()
 imwrite(im, ‘name_of_image.ext’, ‘EXT’);
 im  image to write to disk
 name_of_image.ext  Name of the image
 ‘EXT’  Extension of the file (‘JPG’, ‘BMP’, ‘PNG’, etc.)

Demo Time #1!
Reading in Images
Accessing Pixels
Changing Image Pixels
Obtaining a Subset
Display Images  Monochromatic and Colour
Displaying Colour Channels Separately
Writing Images to File

Resizing Images
 One common thing that many people do is resize
images
 i.e. Make an image bigger from a smaller image, or make
an image smaller from a larger image
 How do we resize images in MATLAB? Use the
imresize command  How do we use it?
out = imresize(im, scale, ‘method’); or
out = imresize(im, [r c], ‘method’);
 For both methods  im is the image we want to resize,
and out is the resized image

Resizing Images – (2)
 Let’s look at the first method:
out = imresize(im, scale, ‘method’);
 scale takes each of the dimensions of the image (# of
rows and columns), and multiplies by this much to
determine the output
 e.g. If we have an image that is 20 rows x 40 columns:
 If scale = 2, the output  40 rows x 80 columns
 If scale = 0.5, the output  10 rows x 20 columns
 method determines the type of interpolation when
resizing  Important when making an image bigger

 When we are making an image bigger, we are trying to
create an image with a lack of information present
 There are three main types of interpolation
 Nearest Neighbour  method = ‘nearest’
 Uses the best pixels that are near the original pixels and fills in
missing information
 Bilinear Interpolation  method = ‘bilinear’
 Uses linear interpolation in 2D to fill in missing information
 Bicubic Interpolation  method = ‘bicubic’
 Uses cubic interpolation in 2D to fill in missing information
 Usually, bicubic is known to have the best accuracy

 Now, let’s take a look at the second method for resizing
out = imresize(im, [r c], ‘method’);
 This routine will resize the image to any desired
dimensions you want
 You can customize how many rows and columns the
final image will have
 Example: To resize a 130 rows x 180 columns image to
65 rows x 90 columns, with bilinear interpolation, do:
out = imresize(im, [65 90], ‘bilinear’);
 We can also do!
out = imresize(im, 0.5, ‘bilinear’);

Rotating Images
 Suppose we want to rotate an image  How?
out = imrotate(im, angle, ‘method’);
 im: The image we want to rotate
 angle: How much we want to rotate the image
 Angle is in degrees! Positive angle denotes counter-
clockwise rotation, and negative angle is clockwise
 When rotating, there will inevitably be some missing
information  method is like before with resizing
 out: The rotated image
 Example: Let’s rotate CCW by 45 degrees by bilinear:
out = imrotate(im, 45, ‘bilinear’);

Demo Time #2!
Enlarging Images
Shrinking Images
Rotating Images

Cont. & Brig. Enhancement
 First real application  Brightness enhancement
 How do we increase /decrease brightness of an image?
 1 way: Just add or subtract a brightness to every pixel
 How? im2 = im + c; or
im2 = im – c;  c is any constant (0 – 255)
 Takes c and add / subtract to every pixel in the image
 Adding / subtracting makes the image brighter / darker
 Round off occurs if out of range (i.e. set to 255 if > 255/set to 0 if < 0)
 What about another way?
 We can also scale the image by a constant
 Do: im2 = c*im;
 If c > 1, increasing brightness
 If c < 1, decreasing brightness

Cont. & Brig. Enhancement (2)
 So we covered brightness… what about contrast?
 Contrast  How well you can see the objects from the
background
 When performing brightness enhancing, you’ll notice
that it looks “white washed out”  Poor contrast
 We can do a contrast enhancement to make objects
look better, leaving background relatively unaffected
 How? Use the power law  s = rγ
 r is the input pixel intensity / colour and s is the output
intensity / colour

 For each pixel, apply this equation to each of the
intensities / colours
 For colour images,
apply to each channel
separately
 Exponent γ determines
how dark or light the
output is
 γ > 1  Make darker
 γ < 1  Make lighter

 How do we apply the power law in MATLAB?
 Use imadjust
out = imadjust(im, [], [], gamma);
 im: Input image to contrast adjust
 Ignore 2nd and 3rd parameters  Beyond the scope of
our talk
 gamma: The γ exponent that we’ve seen earlier
 out: The contrast adjusted image
 Example use: If γ = 1.4, we do:
out = imadjust(im, [], [], 1.4);

Demo Time #3!
Contrast and Brightness Enhancement

Intro to Image Histograms
 We can perform more advanced image enhancement
using histograms
 Before we cover this… we should probably cover what
histograms are! So, what’s a histogram?
 It measures the frequency, or how often, something
occurs
 Let’s look at a grayscale image for now
 Expressed as H(x) = q, x is an intensity- [0,255] for 8 bits
 This tells us that we see the intensity value of x for a
total of q times

Intro to Image Histograms – (2)
 Example: Let’s take a look at an grayscale image

 There are ~1500 pixels with a gray level of around 10
 There are ~1200 pixels with a gray of around 170, etc.

 How do we create histograms in MATLAB?
 imhist(im); %im is read in by imread
 Assuming im is a grayscale image
 If we want this to work with colour images, we will have
to display the histogram of each colour channel by itself
 imhist(im(:,:,1)); % histogram for red
 imhist(im(:,:,2)); % histogram for green
 imhist(im(:,:,3)); % histogram for blue
 The (:,:,1) means that we should grab every row and
column from the red channel, etc.

 Grabbing all of the pixels in any channel will produce a
grayscale image, which can be used for imhist
 We can also call the histogram function by:
h = imhist(im);
 Will create a 256 element array, where the (i+1)’th element
contains how often we see the grayscale i.
 Histograms give us a good insight on image contrast
 If the histogram has too many values towards the left
 Image is too dark
 If the histogram has too many values towards the right
 Image is too bright
 If the histogram has too many values in the middle
 Image looks very washed out

Advanced Enhancement – (1)
 Above cases are when the image has bad contrast
 A good image should have good contrast
 i.e. The histogram should have an equal number of
pixels over the entire histogram (a flat histogram)
 If an image has bad contrast, we can try to correct it by
doing histogram equalization
 Tries to make the image’s histogram as flat as possible
for good contrast  Stretches the histogram out
 For you probability nerds  If you divide each
histogram entry by the total number of entries, this
forms a Probability Density Function (PDF)

Advanced Enhancement – (2)
 Histogram equalization seeks to modify the PDF so that
all possible events (pixels) are equally likely to occur
 How do we perform histogram equalization?
out = histeq(im); %im given by imread
 What about for colour images?
 You will have to perform histogram equalization on each
channel individually, then merge together
r = im(:,:,1); g = im(:,:,2); b = im(:,:,3);
out(:,:,1) = histeq(r);
out(:,:,2) = histeq(g);
out(:,:,3) = histeq(b);

Demo Time #4!
Displaying and Calculating Histograms
Contrast and Brightness Enhancement by
Histograms

Intro to Convolution
 Before we proceed, we need to understand what
convolution is for a digital image
 Probably learned in ELE/BME 532, ELE 792, etc…. Bleck!
 But! Actually very simple when dealing with 2D images
 Preliminaries:
 First, we need to create an N x N matrix called a mask
 The numbers inside the mask will help us control the
kind of operation we’re doing
 Different #s allow us to blur, sharpen, find edges, etc.
 We need to master convolution first, and the rest is easy!

Intro to Convolution – (2)
 Steps:
 1) For each pixel (r,c) in our image, extract a N x N subset
of pixels, where the centre is at (r,c)
 Example: 9 x 9 subsets shown below @ various locations

 Another example: Let’s grab a 3 x 3 subset, located at
(r,c) = (6,4)
 Pixels are a, b,
… g, and centre
is e

 2) Take each pixel in the subset, and do a point-by-
point multiplication with the corresponding location
in the mask

 Example: Our 3 x 3 subset has pixels:
a b c
G = d
 e f

g
 h i


 Our mask has the following quantities

z y x
H = w
 v u

t
 s r


 When we do a point by point multiplication, we will
now have 9 numbers:
 a*z, b*y, c*x, d*w, e*v, f*u, g*t, h*s, i*r
 3) Create an output image and:
 a) Add up all of these values
 b) Store the output at (r,c) (i.e. the same row and
column location as the input image) in the output image
 Example:
out = a*z + b*y + c*x + d*w + e*v + f*u + g*t + h*s + i*r,
then store out into (r,c) of the output image

 Essentially, convolution is a weighted sum of pixels
from the input image within the subset
 Weighted by the numbers in your mask
 Each pixel in the output image is this weighted sum
 Do this for each location of the input image and assign
to same location in the output image
 What about for colour!?
 Do convolution on each channel separately
 Note: If the output value is floating point (decimal),
we must round in order to make this an 8-bit number
 Here’s one more example to be sure you understand…

Blur / Smooth Images
 So now what? Let’s try blurring/smoothing images
 Blurring? Think of a camera that is out of focus
 Why should we blur? Minimize sensor noise
 Noise  High Frequency
 When we blur, we are essentially low-pass filtering,
eliminating high frequency content
 High frequency content  Edges
 By getting rid of the edges, we blur the image
 How do we blur an image? Try an averaging filter

Blur / Smooth Images – (2)
 Averaging? 1st, create an output image to store results
 1) For each pixel (r,c) in the image, look at a N x N
neighbourhood / subset of pixels that surround (r,c)
 2) Add up all pixel values in the neighbourhood and
divide by N2 (number of pixels in neighbourhood)
 3) Take this new value and store it into same (r,c)
location
 Notes:
 This works for B & W images
 What about for colour images?

 Remember, colour image can be seen as a 3D matrix
 3D matrix  2D matrices with layers
 1st layer  Red values
 2nd layer  Green values
 3rd layer  Blue values
 So, we can blur each layer independently, and use the
RGB values after the blurring of each colour layer
 The bigger the neighbourhood, the more the blurring
 How can we efficiently implement this?

 Create a mask for convolution to efficiently do this
 Mask: Same size as desired subset / neighbourhood
 Mask will contain numbers used to generate our result
 Examples: 1
1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1
1 1 1 1 
1 1 1 1 1  1 1 1 1 1 1 1 
1 25 1 
1 1 1  1 1 1 1 1 
1 1 1 1 1 1 1 1 1
9 1 1 1 1 1 1 1 1 1 1 1 1
  1
 1 1 1 1
 81 1 1 1 1 1 1 1 1 1
 
1 1 1 1 1 1 1 1 1
3 x 3 Averaging 5 x 5 Averaging 1 1 1 1 1 1 1 1 1
1 1
Mask Mask  
1 1 1 1 1 1 1

9 x 9 Averaging
Mask

 These masks make sense. Why?
 Let’s look at the 3 x 3 case:
 Example: Suppose our 3 x 3 neighbourhood is:
1 2 3
G = 4
 5 6

7
 8 9

Our 3 x 3 averaging mask is:
1 1 1

1 1 1  9 9 9
1
H = 1 1 1 = 
1 1 1

9  9 9 9
1 1 1  1
  1 1
9
 9 9


 The output should be:
out = (1)(1/9) + (2)(1/9) + (3)(1/9) + (4)(1/9) + (5)(1/9) +
(6)(1/9) + (7)(1/9) + (8)(1/9) + (9)(1/9)
out = (1/9)(1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9)
out = (1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9) / 9
 … isn’t this exactly the same thing as averaging!?
 You are essentially adding up all of the pixels in the
neighbourhood, and dividing them by the total
number of pixels (9)
 Note: If we get a decimal number, we round to be sure
that we have an integer number

 Understanding was the hard part!
 In MATLAB, this is very easy!
 First, we need to create the mask
 We do this by calling fspecial()
mask = fspecial(‘average’, N);
 mask contains the averaging mask to use
 First parameter specifies we want an averaging mask
 N specifies the size of the mask  Bigger the N, more
blurring we get

 Next, call a command that will perform this multiply
and sum command for each pixel in the image
 Use the imfilter() command
out = imfilter(im, mask);
 out: output image (this case  averaged output)
 im contains the image we want to blur
 mask: Convolution mask to use (this case  averaging)
 Nice little note  imfilter() works on both
grayscale and colour images
 For colour, this automatically performs convolution on
each channel individually and combines after

 imfilter() performs image filtering using masks
 Essentially convolution
 Filtering  Produce an output image that changes the
frequency content of the image
 Blurring, Sharpening, Detecting Edges, etc.
 Why are masks important?
 You change the size of the mask, and values in the mask
and you get different results!
 Essentially how most filters in Adobe Photoshop and
GIMP work

Demo Time #5!
Blurring / Smoothing Images

Edge Detection
 Next? Edge detection… what is an edge?
 Edges are sudden discontinuities appearing in images
 A.K.A. sudden jumps in intensity or colour in images
 How can we detect edges? Very simple: Think Calc.
 Find the derivative of each point in the image
 When slope is very high, that means you found an edge
 But, derivative is only for 1D signals… what about 2D!?
 You must take the gradient  Derivative in both the x
and y directions
 You combine both of these directions to create your
final derivative function

Edge Detection – (2)
 How do I take the derivative? We perform convolution, but
with a different mask  Corresponds to discrete
approximation of the derivative
 Two possible masks we can use:
1 1 1 1 2 1
0 0 0
0
 0 0 
 
− 1 − 1 − 1 − 1 − 2
 − 1

 
Prewitt 3 x3 Mask Sobel 3 x 3 Mask
 These masks detect changes in the horizontal direction
 If there are no changes, then weighted sum of the pixels in 3
x 3 neighbourhood should have same value  Gradient = 0

 Prewitt Mask  Normal derivative
 Sobel Mask  Puts more weight on the central pixels
 How does this detect horizontal changes?
 When using the mask, if there is a change, there is a
huge change above the zero line, and below the zero line
 So when performing correlation, it can detect horizontal
changes
 How do we detect vertical changes?
 Take the transpose of each mask!
 Tranpose  Interchange rows & columns of the mask

 How do we combine the horizontal and vertical
gradient information?
 Remember from Vector Calculus:
2
 ∂G   ∂G 
2

| ∇G |=   +
 ∂y 

 ∂x   
 So, we take the horizontal gradient image, and square
each term, and do the same for the vertical gradient
image
 Now, add both images, and square root them  This is
our output image

 How do we do this in MATLAB?
 1) Create our Prewitt or Sobel Horizontal Masks:
mask = fspecial(‘prewitt’); or
mask = fspecial(‘sobel’);
 2) Get the Vertical Mask:
mask2 = mask’;  ‘ transposes a matrix
 3) Convolve with each mask separately
dX = imfilter(im, mask);
dY = imfilter(im, mask2);
 4)Find the overall magnitude
mag = sqrt(double(dX).^(2) +
double(dY).^(2));

 Note: We must cast images to double for sqrt() func.
 Now, we’ve found the overall gradient… how do we find
an actual edge? Threshold the image
 What do we mean by threshold?
 Values greater than a threshold is an edge (white)
 Values less than the threshold are not edges (black)
 How do we detect edges in MATLAB? Very easy!
I = edge(im, ‘sobel’); or
I = edge(im, ‘prewitt’);
 So, we can pretty much ignore the previous slide, but I
put that in there so you can see where it comes from

 You can also call the routines this way:
I = edge(im, ‘sobel’, thresh); or
I = edge(im, ‘prewitt’, thresh);
 I specifies edges found in input (B & W image)
 im is the input image with edges to be found
 Second parameter specifies method of finding edges
 thresh determines threshold for finding edges
 If not specified, threshold will be found automatically
 How do we use the thresh variable?
 Choose an intensity (e.g. 128)
 Any gradient value > 128 will be labeled white

 Any gradient value < 128 is labelled black
 thresh is between [0,1], so take your threshold and
divide by 255 to use
i.e. I = edge(im, ‘prewitt’, 128/255);
 Output image will give you a black and white image 
White is an edge, black is not an edge
 Only works for B & W images. To find edges for colour
images, convert the colour image to a B & W image
 Use gray = rgb2gray(im);
 rgb2gray converts a colour image into B & W by
doing: I = (R + G + B) / 3  Each colour pixel is the
average of the red, green and blue components

 Sidenote: Here’s another way of finding the gradient
 Unsharp Masking
 Probably heard this terminology in CSI? So what’s
unsharp masking?
 Let’s go back to Signals and Systems
 Suppose we perform a low-pass filtering of an image 
Get a blurred version
 If we take the original image, and subtract its blurred
version, what are we doing? Removing all low frequency
components, so what’s left? High frequency!

 High frequencies are essentially edge information
 Edges are sudden jumps  Essentially high frequency
 How do we perform unsharp masking?
 1) Blur the image  Try using the following mask:
mask = fspecial(‘average’, 5); %5x5
 So… we blur! im_blur = imfilter(im,mask);
 2) Subtract the original image from the blurred
im_unsharp = im – im_blur;
 Now, why are edges useful? We can use them to
sharpen images  Will take a look after this demo

Demo Time #6!
Edge Detection in Images
Calculating an Unsharp Mask

Image Sharpening
 Last image enhancing topic  Image Sharpening
 What is image sharpening? Make the image look
“clearer”, “sharper”, make the details “pop out more”
 How do we do this?
 Edges are the “detail” behind the image
 Edges are high frequency
 How do we make images sharper?
 Find overall magnitude of the gradient for the image
 Add these values on top of the original image
 Result? Increase the visibility of the edges  Sharper

Image Sharpening – (2)
 Use imfilter with unsharp masking
 mask = fspecial(‘unsharp’);
 The above syntax creates an unsharp mask in such a way
where when you convolve, it will automatically subtract
the image with a blurred version of itself and add the
original image on top
 Good for both x and y direction changes
 How?
mask = fspecial(‘unsharp’);
sharp = imfilter(im, mask);
 This will perform the detection of abrupt changes, and
adding them on top of the image all in one go.

Demo Time #7!
Image Sharpening in Images

 3rd Hour: 8:10 p.m. – 9:00 p.m.
Applications of Image Processing
 Segmenting simple objects
 Noise Filtering
 Simple image stitching using template matching

Segmentation via Thresholding
 There are times when we want to separate objects (the
things we want) from the background  Segmentation
 To make this easier, we convert to a B & W image first if the
image is in colour. Else just leave it alone
 After conversion, most of the time the intensities of the
objects are distinctly different from background
 We can write code to save the pixels that match these
intensities, and disregard the rest
 Pixels that match  Set to white, else set to black
 Output image will be binary  White belonging to object,
and black belonging to background
 We can use this to mask out the background pixels

Segmentation via Thresholding (2)
 What do we do?
 Take a look at the histogram, and see which pixels are
predominantly belonging to the object, and of the
background
 We threshold the image using this information
 We then use this threshold map to figure out what pixels
we want to keep, and what we want to disregard

Demo Time #8!
Object Segmentation via Thresholding

Noise Filtering
 A common task in image processing is to eliminate or
reduce image noise from an image
 Image Noise: Pixels in an image that are corrupted
undesirably
 Pixels could be corrupted in the acquisition process, or
in transmitting, etc.
 How do we reduce image noise?
 Think in the frequency domain  Noise is essentially
high-frequency information
 If we blur the image, we would eliminate the noise, but
the quality would reduce  blurring details

Noise Filtering – (2)
 We can generate artificial noise and add these to
images
 Purpose is for research  Design good filters by
recreating the noise we would encounter in practice
 How do we generate artificial noise? Use imnoise
 We will be concerned with two ways of generating noise
out = imnoise(im, ‘gaussian’, mean, var);
out = imnoise(im, ‘salt & pepper’, prob);
 First method generates Gaussian-based noise
 Usually encountered in transmitting process

 Second method generates “salt & pepper” based noise
 Also known as impulsive noise
 Called this way, because for monochromatic images, it
literally looks like someone took salt (white pixels) and
pepper (black pixels) and shook it over the image
 For colour images, specks of pure red, green and blue
pixels appear
 We can use blurring to get rid of Gaussian noise, but
for impulsive noise, we need to use median filtering
 What’s median filtering? Like convolution, but we’re not
doing a weighted sum

 For each pixel (r,c) in the image, extract an M x N
subset of pixels centered at (r,c)
 Sort these pixels in ascending order, and grab the
median value
 The output image at (r,c) is this value
 How do we perform median filtering in MATLAB?
out = medfilt2(im, [M N]);

Demo Time #9!
Noise Filtering in Images

Template Matching
 Template matching is using a small test image, which
we’ll call a patch
 Objective is to automatically find the location of where
this patch is in the entire image
 Useful in a variety of applications: Image Retrieval, Eye
Detection, etc.
 What I’ll show you today is to do some basic image
stitching
 We will have two images of the same scene, that have
been taken at slightly different perspectives  Only
horizontal shifting is concentrated on here

Template Matching – (2)
 What’s the best way to do template matching?
 1) For each pixel (r,c) in the image, extract a subset that
is the same size as the template with its centre at (r,c)
 2) Perform a cross-correlation between the patch and
this subset
 3) Take this value and assign it to location (r,c) for the
output
 4) The best location of where the template is, is where
the maximum cross-correlation is
 We can perform (1) to (3) by doing:
C = normxcorr2(template, im);

 Output is a correlation map
 To find the row and column co-ordinates of where the
template best matches, do the following:
[row col] = find(C == max(C(:)));
 So, how do we do image stitching?
 1) Extract a region in either image that is common
between both  Template
 2) Find the co-ordinates of where this template is in
both images
 3) Determine how much vertical and horizontal
displacement there is between the two images

 4) Create an output image with dimensions that
encapsulate both images together
 5) Place one image on the left, then place the other
image by displacing it over by the horizontal and vertical
shift
 OK… let’s do some code!

Demo Time #10!
Simple Image Stitching

Conclusion
 MATLAB is a great tool for digital image processing
 Very easy to use
 This is not an exhaustive tutorial! There are many
more things you can do with MATLAB
 For more image processing demos, check out:
http://www.mathworks.com/products/image/demos.html
 Lots of cool image processing stuff you can find here
 For a more comprehensive MATLAB tutorial, check:
http://www.ee.ryerson.ca/~rphan/ele532/MATLABTutorial.ppt
 You can access the slides, images and code at:
http://www.rnet.ryerson.ca/~rphan/IEEEDIPTalk

Introduction to Digital Image Processing Using MATLAB

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (9)

Similar to Introduction to Digital Image Processing Using MATLAB

Similar to Introduction to Digital Image Processing Using MATLAB (20)

Recently uploaded

Recently uploaded (20)

Introduction to Digital Image Processing Using MATLAB