Q.

1 Apply the image degradation/restoration model to explain how an image affected by

Gaussian noise can be restored using inverse filtering. Illustrate with a suitable example.
Based on the provided text:
•
Image Degradation Model: Image degradation is modeled as an operator , together with an additive
noise term h(x,y), operating on an input image f(x,y) to produce a degraded image g(x,y)1. This can
be represented as g(x,y) = f(x,y) + h(x,y) when only additive noise is present2. The objective of
restoration is to obtain an estimate f̂(x,y) of the original image, as close as possible to f(x,y)1. A
diagram illustrating this process shows the input image f(x,y) passing through a Degradation block
(combining and Noise h(x,y)) to produce g(x,y), which then enters a Restoration filter(s) block to
produce the estimate f̂(x,y)3.
•
Noise Sources: Principal sources of noise arise during image acquisition and/or transmission4.
•
Gaussian Noise: Gaussian noise is one type of noise model5.... The distribution of amplitude values
in white Gaussian noise is Gaussian7.
•
Restoration in Presence of Noise Only: When degraded only by additive noise, g(x,y) = f(x,y) +
h(x,y), which in the frequency domain is G(u,v) = F(u,v) + N(u,v)2. Generally, the noise terms are
unknown, so subtracting them to obtain the original image is typically not an option2.
•
Frequency Domain Filtering: Periodic noise can be reduced significantly via frequency domain
filtering8.... In the case of periodic noise, it is sometimes possible to estimate N(u,v) from the
spectrum of G(u,v)10. If N(u,v) can be estimated, it can be subtracted from G(u,v) to obtain an
estimate of the original image10.
•
Spatial Filtering: Spatial filtering is the method of choice for estimating f(x,y) (denoising g(x,y)) when
only additive random noise is present10. Mean filters6... and order-statistic filters like the median
filter12... are examples of spatial filtering techniques for noise reduction discussed in the module6....
•
Inverse Filtering for Gaussian Noise: The provided text does not describe how an image affected by
Gaussian noise (a type of random noise5...) can be restored specifically using inverse filtering. It
mentions that subtracting unknown noise is typically not possible2... and highlights spatial filtering as
the method for additive random noise10. It mentions frequency domain filtering (specifically notch
filtering) for periodic noise8... and estimating/subtracting that noise10. Section 5.5 is mentioned for
discussing image restoration in the presence of both and h4, but the content of Section 5.5 is not
provided.
•
Example: The provided examples illustrating restoration in the presence of noise only (6...) use
spatial filtering (mean, geometric mean, median, adaptive median filters) on images corrupted by
salt-and-pepper or Gaussian noise. There is no example illustrating inverse filtering for Gaussian
noise.
(Page 318-340)

Q.2 Given an image affected by additive white noise, apply spatial filtering techniques to
perform image restoration. Compare the effectiveness of mean and median filters with
suitable examples.
Based on the provided text:
•
Additive Noise: The degradation process can include an additive noise term h(x,y)1, resulting in a
degraded image g(x,y) = f(x,y) + h(x,y) when only noise is present2.
•
White Noise: White noise has a frequency spectrum that is constant over a specified frequency
band3.... Its values can be statistically independent (e.g., White Gaussian noise)7.
•
Spatial Filtering for Denoising: Spatial filtering is the method of choice for restoring images degraded
only by additive random noise10.
•
Arithmetic Mean Filter: This is the simplest mean filter11. It computes the average value of the
corrupted image pixels in a defined neighborhood (Sxy) centered on the point (x,y), replacing the
original pixel value with this average11. This filter is the same as the box filter11.
•
Median Filter: This is the best-known order-statistic filter12. It replaces the value of a pixel with the
median of the intensity levels in a predefined neighborhood (Sxy)12. The pixel at (x,y) is included in
the median computation12.
•
Comparison and Effectiveness (based on examples):
◦
General: Median filters are popular because, for certain types of random noise, they provide
excellent noise-reduction capabilities with considerably less blurring than linear smoothing filters (like
mean filters) of similar size13.
◦
Salt-and-Pepper Noise: Median filters are particularly effective against bipolar and unipolar impulse
noise (like salt-and-pepper noise)13. Example 5.3 shows an image corrupted by salt-and-pepper
noise (P = P = 0.1)14. Filtering with a 3x3 median filter significantly improves the image, removing
most noise points, though repeated passes might be needed for very high noise levels, at the risk of
blurring14.... Figure 5.10 illustrates the result of median filtering on salt-and-pepper noise15. Figure
5.12 compares a 5x5 arithmetic mean filter, geometric mean filter, median filter, and alpha-trimmed
mean filter on an image with uniform and salt-and-pepper noise, visually showing differences in
performance16.
◦
Gaussian Noise: Example 5.4 shows an image corrupted by additive Gaussian noise6. Both a 7x7
arithmetic mean filter and a 7x7 geometric mean filter smoothed out the noise, but at the cost of
significant blurring6. The differences between their results were primarily in the degree of blurring6.
◦
The provided text does not explicitly show the application of these filters on white noise, but white
Gaussian noise is mentioned7. The examples demonstrate their effect on specific types of random
noise like Gaussian and salt-and-pepper6.... Spatial filtering, using mean and median filters, is the
recommended approach for such noise10.
(Page 327-336, 730)

Q.3 Explain the basic components involved in the image degradation/restoration model.
Illustrate with a simple diagram.
Based on the provided text:
•
Input Image f(x,y): This is the original, ideal image that we want to restore1.
•
Degradation Operator : This represents the process or system that degrades the image1. The nature
of this operator can vary.
•
Additive Noise Term h(x,y): This term represents unwanted signal or interference added to the image
during acquisition or transmission1.... When degradation is only due to additive noise, the degraded
image is the sum of the original image and the noise: g(x,y) = f(x,y) + h(x,y)2.
•
Degraded Image g(x,y): This is the observed image, which is the result of the original image being
operated on by and corrupted by additive noise h(x,y)1.
•
Restoration Filter(s): This is the process applied to the degraded image g(x,y) to produce an
estimate of the original image1.
•
Estimated Image f̂(x,y): This is the output of the restoration process, aiming to be as close as
possible to the original image f(x,y)1.
Diagram (Figure 5.1):
(Page 318, 319, 327)1…​
​

Q.4 Define edge detection. Explain how gradient-based methods are used for detecting edges
in an image.
Based on the provided text:
•
Definition of Edge Detection: Edge detection is based on detecting sharp, local changes in
intensity21. Edge pixels are pixels where the intensity of an image changes abruptly, and edges (or
edge segments) are sets of connected edge pixels21. Edge detectors are local image processing
tools designed to detect these edge pixels21.
•
Role in Segmentation: Edge-based segmentation is a principal approach in which boundaries of
regions are detected based on local discontinuities in intensity22. Ideally, edge detection yields pixels
lying only on edges, but in practice, noise, breaks, and illumination issues introduce
discontinuities23.
•
Stages: Edge detection techniques aim for three objectives: response to true edges, minimum false
positives, and close localization to the true edge24. The process involves detecting edge points
(local operation) and edge localization (selecting points that are members of an edge)25. Edge
linking algorithms are typically used after detection to assemble edge pixels into meaningful
edges23.
•
Gradient-Based Methods:
◦
The provided text mentions edge detection by gradient operators in Section 3.626.
◦
It states that an edge image can be computed as the magnitude of the gradient or the absolute value
of the Laplacian27.... This magnitude image can then be thresholded to produce a binary mask
highlighting "strong" edge pixels27....
◦
Example 10.15 uses the gradient magnitude image, thresholded at a high percentile (e.g., 99.7%), to
create a mask29. This mask is then used to select pixels from the original image whose histogram is
then used for global thresholding27....
◦
Example 10.10 shows detecting strong horizontal and vertical edges by thresholding the gradient
magnitude image30.
◦
The sources describe using the result (magnitude) of gradient computation27... but do not provide
the mathematical definition of the gradient operator or the detailed steps on how it is computed for
detecting intensity changes. The term "gradient operators" is mentioned26, but their specific
application and mechanism for edge detection are not fully explained in the provided excerpts,
beyond stating that the magnitude of the gradient is used to create an edge image27....
(Page 701, 702, 712, 715, 730, 734, 737, 755, 757)21…​
​

Q.5 Use color features to segment an image into distinct regions. Explain the process using a
suitable color model and show how segmentation is achieved.
Based on the provided text, we can explain segmentation using the RGB color model and vector
processing32:
•
Segmentation Definition: Segmentation is a process that partitions an image into regions33.... The
goal is to group pixels into subregions that satisfy certain properties34....
•
Color in Segmentation: Color is heavily used in segmentation, especially where distinguishing
regions based on intensity alone is difficult36. Color pixels can be interpreted as vectors37....
•
Segmentation in RGB Space using Color Vectors:
◦
Working directly with RGB color vectors generally yields better segmentation results than using HSI
space32....
◦
Approach: Suppose the objective is to segment objects of a specific color range in an RGB image32.
◦
Prototype/Average Color: Obtain a set of sample color points representative of the colors of interest
and estimate their "average" color, denoted by the RGB vector a32.
◦
Similarity Measure (Euclidean Distance): Use a measure of similarity to compare each RGB pixel in
the image to the average color vector a32. A simple measure is the Euclidean distance32. For an
arbitrary pixel z in RGB space, the Euclidean distance to a is given by D(z, a) = sqrt((zR - aR)² + (zG
- aG)² + zB - aB)²)40.
◦
Classification/Thresholding: Classify a pixel z as having a color in the specified range if the distance
D(z, a) is less than a specified threshold, D₀40. Points within the sphere of radius D₀ centered at a in
RGB space satisfy the color criterion40.
◦
Generating Segmented Image: Code the points satisfying the criterion (e.g., with white) and points
not satisfying it (e.g., with black) to produce a binary segmented image40.
◦
Refinement (using a Box): Instead of a sphere defined by mean and standard deviation for each
component, a more accurate method can use a rectangular box in RGB space defined by the mean
and variance of the sample points39. For example, the R dimension of the box could extend from aR
- 1.25 * sR to aR + 1.25 * sR, where aR is the mean and sR is the standard deviation of the red
component of the sample points39. A pixel is classified as belonging to the target color range if its
RGB values fall within this box39.
•
Example: Example 6.15 and Figure 6.42 illustrate segmentation in RGB space32.... Sample points of
the color of interest are enclosed by a rectangle in the original image [52(a)]. The segmentation
result [52(b)] shows the regions identified by coding points inside or on the surface of a box (derived
from the sample statistics) as white and others as black39. This method is shown to be more
accurate than the HSI approach in a comparative example39.
(Page 446-450, 66, 67, 104)32…

Q.6 Apply the RGB to HSI color model transformation to convert a given RGB color value.
Show all intermediate steps and final HSI values.
Based on the provided text:
•
HSI Model Purpose: The HSI (hue, saturation, intensity) color model decouples intensity from color
information (hue and saturation)41. It corresponds closely with how humans describe color and is
useful for algorithms based on natural color descriptions41.
•
RGB to HSI Conversion: The module states that HSI components can be derived from RGB42. This
conversion uses specific equations, mentioned as Eqs. (6-16) through (6-19) for RGB to HSI42 and
Eqs. (6-20) through (6-30) for HSI to RGB43. The intensity component is the average of the RGB
values at the corresponding pixel44.... Hue corresponds to angles, and saturation measures
purity41....
•
Showing Steps and Values: The provided syllabus excerpts mention the existence of the conversion
equations (6-16 through 6-19) but do not include the equations themselves. Without these specific
mathematical formulas, it is not possible to perform the conversion for a given RGB color value and
show the intermediate steps and final HSI values based only on the provided text.
•
Interpretation of HSI Components:
◦
Hue: Gray-level values in the hue image correspond to angles42. After normalization (e.g.,1), it's
typically scaled (e.g., to 8 bits or ) for display42....
◦
Saturation: Corresponds to saturation values, also typically scaled for display42. Darker values
indicate less saturated colors, progressively darker toward the white vertex of the RGB cube44.
◦
Intensity: Every pixel in the intensity image is the average of the corresponding RGB values44.... The
intensity axis in the RGB cube goes from black (0,0,0) to white (1,1,1)46.
(Page 414-418, 454, 455)41…

Q.7 Describe and compare the RGB and CMY color models. Where is each model commonly
used?
Based on the provided text:
•
RGB Color Model:
◦
Based on a Cartesian coordinate system47.
◦
Each color is represented by its primary spectral components: red, green, and blue47.
◦
The color subspace is represented by a cube, where RGB primaries are at three corners, secondary
colors (cyan, magenta, yellow) are at three others, black is at the origin (0,0,0), and white is at the
opposite corner (1,1,1)47.
◦
The grayscale (points of equal RGB values) extends along the line from black to white47.
◦
Colors are points on or inside the cube, defined by vectors from the origin47.
◦
Values are often normalized to the range11.
◦
RGB images consist of three component images, one for each primary color48.
◦
These three images combine on an RGB monitor to produce a composite color image48.
◦
Used for color generation, such as image capture by a color camera or image display on a monitor
screen41....
◦
A 24-bit RGB image (8 bits per component) is often called a full-color image, with over 16 million
possible colors48.
•
CMY and CMYK Color Models:
◦
CMY stands for Cyan, Magenta, Yellow49.
◦
CMYK adds Black (K) to CMY49.
◦
These models are used for color printing49.
•
Comparison:
◦
Basis: RGB is an additive model (combining light primaries to create color), while CMY/CMYK are
subtractive models (mixing pigments that subtract light) [Implicit from usage: RGB for displays
emitting light, CMY/K for prints reflecting light].
◦
Primaries/Secondaries: RGB uses Red, Green, Blue as primaries and Cyan, Magenta, Yellow as
secondaries (formed by combining two primaries)47. CMY uses Cyan, Magenta, Yellow as primaries.
◦
Usage: RGB is ideal for image color generation and display41.... CMY and CMYK are used for color
printing49.
◦
Representation Detail: The provided text describes the RGB model and its cube representation in
detail47..., while CMY and CMYK are primarily mentioned in the context of their application
(printing)49.
(Page 406, 407)41…

Q.8 Explain the basic concepts of erosion and dilation in morphological image processing.
How do they differ in terms of effect on binary images?
Based on the provided text:
•
Morphological Image Processing: The language of mathematical morphology is set theory, where
sets represent objects in images50.
•
Binary Images: In binary images, objects are represented as sets of 2D integer coordinates (Z²) for
foreground pixels50. Morphological operations involve interactions between images (sets) and
structuring elements (SE)51.
•
Erosion:
◦
Consider an object (set A) in a binary image and a structuring element (SE) B51....
◦
To perform erosion of A by B (denoted A B), the SE B is translated (slid) over the image51.
◦
At each position (translation) z, if the SE B is completely contained within set A (the object), the
location of the origin of B (often its center) is marked as a foreground pixel in the output image51....
◦
Formally, erosion is the set of all points z such that B translated by z is a subset of A52.
◦
Effect: Erosion shrinks or thins objects in a binary image51.... For example, applying erosion to a set
A can erode its boundary, making the object smaller51.... The result is controlled by the shape of the
structuring element52.
•
Dilation:
◦
While not formally defined with set notation in the provided text, dilation is introduced through an
example53.
◦
Dilation is used for applications like bridging gaps53. Example 9.2 shows an image with broken
characters where the maximum length of the breaks is known53. A specific structuring element is
used, and applying dilation with this SE results in the gaps being bridged53.
◦
Effect: Dilation effectively expands or thickens objects in a binary image, allowing it to connect
broken parts or fill small holes53.
•
Difference in Effect:
◦
Erosion shrinks/thins objects by removing pixels on the boundary or within smaller features based on
the structuring element's shape and size51.... It removes foreground pixels that cannot "contain" the
structuring element centered at that point.
◦
Dilation expands/thickens objects by adding pixels around the boundary or filling gaps/holes based
on the structuring element's shape and size53. It adds foreground pixels wherever the structuring
element "hits" the object.
◦
Erosion is used to remove small objects or thin boundaries, while dilation is used to fill holes, bridge
gaps, or thicken objects.
(Page 636, 637, 638, 653, 654, 656)50…

Q.9 What is the Hit-or-Miss transform? Explain its role in shape detection in binary images.
Based on the provided text:
•
The term "Hit-or-Miss transform" is not explicitly named in the provided excerpts. However, the
process described in Example 9.3 and Figure 9.12 precisely illustrates the core concept of the
Hit-or-Miss transform for shape detection.
•
Shape Detection Process (as described in Example 9.3):
◦
The objective is to detect a specific object or shape (let's call it D) within a binary image
(representing set A, the foreground)57.
◦
To detect shape D, two structuring elements (SEs) are designed57. Let B₁ be an SE that matches the
shape of the object D itself57. Let B₂ be an SE that matches the shape of the background
immediately surrounding the object D57.
◦
The process involves two erosions57:
1.
Erode the original foreground set A by B₁: (A B₁)57. This operation finds all locations where the
shape B₁ fits entirely within the foreground object(s) in A51....
2.
Erode the complement of the foreground set A (denoted Aᶜ) by B₂: (Aᶜ B₂)57. This finds all locations
where the shape B₂ fits entirely within the background region (the complement of the object(s)) in
A52. This effectively locates the background pattern surrounding the desired shape.
◦
Intersection: The result of the shape detection is obtained by taking the intersection of the two
erosion results: (A B₁) (Aᶜ B₂)57.
◦
Role in Shape Detection: This intersection marks the location of the origin (reference point) of the
target shape D within the image57. By designing B₁ to match the foreground shape and B₂ to match
its specific background context, this combined operation allows for the precise detection of
occurrences of that particular shape D and its surrounding background pattern in the binary image57.
This method relies on simultaneously matching the foreground shape and the background pattern
surrounding it.
(Page 651, 652, 653)54…

Q.10 Describe the working of a minimum distance classifier.

Based on the provided text:
•
Pattern Classification: This involves automatically assigning patterns to their respective pattern
classes58. A pattern is a spatial arrangement of features, and a pattern class is a set of patterns
sharing common properties58. The process typically involves sensing, preprocessing, feature
extraction, and classification58.
•
Classification Methods: Basic approaches include prototype matching, statistical classifiers, and
neural networks59.
•
Prototype Matching: One approach is classification based on matching unknown patterns against
specified prototypes59.... This includes techniques like correlation (template matching)60.
•
Minimum Distance Classification Principle: The core principle of minimum distance classification is
described within the context of k-means clustering for region segmentation61 and distance measures
in RGB space segmentation32.
◦
In k-means clustering, which aims to partition a set of observations (pattern vectors) into k clusters,
each observation (vector) is assigned to the cluster with the nearest mean61. The mean of a cluster
is called its prototype61.
◦
For pattern classification, the concept extends to assigning an unknown pattern vector to the class
whose prototype is closest61.
◦
Prototypes can be specified (e.g., a known average color vector32) or learned (e.g., the means of
clusters in k-means clustering61).
◦
A distance measure is used to quantify "nearest" or "closest"32.... The Euclidean distance is a
common measure mentioned32.... Other measures like the Mahalanobis distance are also
mentioned62.
•
Working:
1.
Represent Patterns as Vectors: Unknown patterns are represented as feature vectors63.... These
vectors contain distinctive attributes of the pattern.
2.
Define Prototypes: Establish a prototype vector for each class. These prototypes represent the
typical pattern for each class.
3.
Calculate Distance: For an unknown pattern vector, calculate the distance between this vector and
the prototype vector of each class using a chosen distance measure (e.g., Euclidean distance)32....
4.
Assign Class: Assign the unknown pattern vector to the class whose prototype has the minimum
calculated distance to the unknown vector.
•
Role: Minimum distance classification is a simple yet fundamental approach within prototype
matching, where the decision boundary between classes is determined by the points equidistant
from the prototypes.
(Page 446, 759, 812, 904, 905, 906, 910)