DSP Club

Cepstrum Presents-
Talk on Image Processing

Speakers: Turaga Pavan

Nishant Mohan
Purpose of Talk
 Preview of the field
 Starting point for future IPians
 Knowledge of Modern Image Processing
Plan of Talk
 Basics of image storage
 Major issues in IP
 Some Matlab functions
 Interesting Readings
 What’s on in IIT-G
Image as a Signal
 Image – 2D information of light
intensity.
 Information Source, Focussing Agent,
Sensor.
 Nearly always digital.
Digital Images
 2D array of numbers – representing
intensity.
 Numbers stored as bits.
 Generally used – 8 bit/pixel, 16
bit/pixel.
 Gray Scales
 Low intensity-
Lower gray scale
value
 Highest possible
gray level=2n-1
Color Image
 Primary Colors- Red, Green, Blue.
 All colors can be formed by specifying
relative values of R,G & B components.
 Each component is separately stored
like gray levels (3D matrix).
 Hence color images can occupy 3 times
the space required by b/w images.
Image Storage Formats
 Popular formats : BMP, JPG, GIF, TIFF
etc
 Different formats – Different levels and
techniques of image compression
 Commercial Issues - Registered
formats
Compression
 Resolution of standard medical images
is : 1024x1024
 A complete data set may contain 100
slices
 Storing as 8-bit/pixel
 Computer memory required = 800 Mb
 Hence compression is practically must
How to Compress
 Psycho-visual redundancy
 Spatial redundancy
 Mathematical models – Information
theoretic redundancies
 Transform coding
Matlab Function
 imread: reads image data in major
formats
 imwrite: writes image from given data
 imshow: shows image in matlab
window
Till Now ………
 How to look at image as a signal
 Need for compression
 Why is compression possible

Q?
Challenges in Image
Processing

 Image Enhancement
 Image Segmentation
 Image Understanding – retrieving
required information from an image.
Image Enhancement
Common Problems with images –
 Noise

 Blur

 Low Contrast
Noise Removal
 Image Signal is generally base-band
and noise has a broad spectrum
 Noise is removed by spatial filtering
techniques (similar to 1D signals)
Noise reduction example
Image De-blurring
 Basically, a smoothed version of an
image is called blur
 De-blurring is equivalent to enhancing
high frequency components of the
image
 High Pass filters are used to get the
high frequency components and added
back to the original image
De-blurring
Histogram Processing
 The histogram of a digital image with gray
levels from 0 to L-1 is a discrete function
h(rk)=nk, where:
 rk is the kth gray level
 nk is the # pixels in the image with that gray level
 n is the total number of pixels in the image
 k = 0, 1, 2, …, L-1
 Normalized histogram: p(rk)=nk/n
 sum of all components = 1
Image Enhancement in the
Spatial Domain
 Histogram Processing
 Histogram is good indicator of image contrast
 Histogram equalization is technique that
allows the image to take all gray scale values
nearly equally
 Done by mapping value of each pixel to an
appropriate value
Matlab Functions
 hist: Gives histogram of a sequence
 histeq: Performs histogram equalisation
Till now ………….
 Compression and Storage
 Requirement of Enhancement
 Histogram an useful function
 Use of histogram for image
enhancement

Q?
Image Segmentation
 One of the most studied problems in
Image Processing
 Identification of pixels belonging to the
image
 Or marking the boundary of the object
Segmentation Methods
 Thresholding – Decisions based on gray
levels
 Region Growing – Gray level +
connectivity of object
 Active Contours – Gradient +
Connectivity + smoothness of shape
 Thresholding Method
 Basic Idea:
Select a threshold.
Pixels above threshold -> region 1
Pixels below threshold -> region 2

Main challenge -> selection of threshold

 Thresholding Method
Advantages:
1) Simplicity of Implementation
2) Useful if histogram has well separated
and distinct peaks

Disadvantages:
1) Poor noise performance.
2) Sensitive to intensity inhomogeneities
Thresholding Method
Deformable Models
 Motivated by mechanical concept of
energy minimization.
 An initial contour is defined that moves
under influence of pre-defined forces.
 Forces: Internal -> control shape of
contour. External -> control the position
of contour.
Deformable Models
 Internal energy: first derivative->
elasticity. Second derivative-> rigidity
 External energy: Minimum at points where
gradient is high.
 Hence above energy definitions give
necessary contour characteristics.
Deformable Models
Deformable Models
 Entropy based Segmentation
 The technique is inspired by information
theory approach
 We define Gray-level Image Entropy as:
J
GIE  H [ p( g )]   p ( j ) H [ p j ( g )]
j 1

- GIE measures homogeneity of

the segmented region.
Entropy based Segmentation

Original Stage-1
Entropy based Segmentation
Our Hybrid Approach
 Deformable models
- Good noise performance
- Use of local gradient information
- Needs initialization
 Entropy based Approach
- Region based method
- No need for initialization
- Not robust in presence of noise.
Our Hybrid Approach
Our Hybrid Approach
Our Hybrid Approach
Matlab Function
 edge: Finds edges using different
methods ( Canny’s method is very
popular)
Till now ……….
 Compression and Storage
 Enhancement and use of histogram
 Image segmentation
 Use of deformable models and entropy
based histogram

Q?
Medical Imaging
 Aim is to observe internal structures
 Penetration of x-ray allows internal
observation
 Information obtained is only average
absorption by the body
 How to get internal 3D picture ……..
………. Tomography
Tomography
 Technique of obtaining 3d image by 2d
projections from different angles
 Radon showed – absorption value of
each point in the plane can be
calculated using total absorption of rays
at different angles
 Various slices obtained to get 3d info
Tomography
Tomography
Other Techniques
 Magnetic Resonance Imaging – Use of
Radio Frequencies and spin of H nuclei
 Positron Emission Tomography – Use of
radioactive material
 ………… many more
Computer Vision
Pattern Recognition
Image Understanding
Computer Vision
Pattern Recognition
Image Understanding
How do we understand images ?
Object Recognition
Character Identification
Face Detection
Describing Image “Content”
 How does a computer know what an
image means ?
 Descriptors/ Features : Numbers that
try to capture some feature of the
image
 Examples: Grayscale/Color Histogram,
Wavelet Statistics, Cooccurence Matrix,
Corellogram.
Measuring Similarity
 When are two images similar ?
 Extremely subjective
 Do we recognize objects based on
similarity ?
Content Based Image
Retrieval
 Aim: To retrieve images similar to a
given image from a database of images
 Not Google Image Search !!!
 What is similarity ? Can it be
mathematically described ?
Result of CBIR

Courtesy: [Link]
Face Detection
Challenge
Tracking
 Goal: To track a moving target
Almost unlimited scope
 Character recognition
 Facial Gesture Analysis
 Biometrics
 Artificial Navigation
 Surveillance etc…
So we saw …
 What is an image?
 How images are stored ?
 Basic Image operations
 Applications in Biomedical Images and
Computer Vision
Interesting Readings
 “Digital Image Processing” , Woods &
Gonzalez.
 “Fundamentals of Digital Image Processing”,
A.K. Jain.
 IEEE Transactions on Image Processing,
Pattern Analysis and Machine Intelligence,
Medical Imaging.
 The Computer Vision Homepage –
[Link]
[Link]/afs/cs/project/cil/ftp/html/[Link]