Image Restoration

Image restoration refers to the process the average of its neighbors; reduces
of recovering an original image that has Gaussian noise but can blur edges. 6. Homomorphic Filtering
been degraded by known distortions - Median Filter: Replaces each pixel with Homomorphic filtering is used for
such as noise, blur, or motion. The main the median of the surrounding pixels; contrast enhancement and illumination
goal is to reconstruct a clean image from effective for salt-and-pepper noise. correction. It is based on separating
a noisy or blurred version using - Gaussian Filter: Uses a Gaussian illumination and reflectance:
mathematical models and filtering function for smoothing; preserves edges f(x, y) = i(x, y) * r(x, y)
techniques. better than mean filtering.
- Adaptive Filters: Adjust filter behavior Taking the logarithm:
based on local statistics. ln f(x, y) = ln i(x, y) + ln r(x, y)
1. Image Degradation and Restoration
Model This transforms multiplicative
The degradation process can be 4. Degradation Function components into additive ones. The
mathematically modeled as: The degradation function h(x, y) defines image is then processed in the
g(x, y) = h(x, y) * f(x, y) + η(x, y) the distortion applied to the image. frequency domain to reduce low-
Where: Common causes include: frequency illumination and enhance
- g(x, y): Degraded image - Motion blur: When the camera or high-frequency reflectance (details),
- f(x, y): Original image object moves during image capture. followed by exponential transformation
- h(x, y): Degradation function (e.g., blur - Defocus blur: Due to out-of-focus to get the final result.
function) lenses.
- *: Convolution operation .
- η(x, y): Additive noise In the frequency domain, degradation is
represented as:
The aim of image restoration is to G(u, v) = H(u, v) * F(u, v) + N(u, v)
estimate the original image f(x, y) from
the observed g(x, y), given knowledge of Where:
h(x, y) and the noise. - G(u, v): Fourier transform of degraded
image
- H(u, v): Degradation function in
2. Noise Models frequency domain
Noise is unwanted random variation in - F(u, v): Original image in frequency
pixel values, often introduced during domain
image acquisition or transmission. - N(u, v): Noise component
Common noise models include:
- Gaussian Noise: Caused by sensor
noise; follows a normal distribution. 5. Inverse Filtering
- Salt-and-Pepper Noise: Caused by Inverse filtering attempts to restore the
transmission errors; appears as random image by reversing the effect of the
black and white dots. degradation:
- Poisson Noise: Seen in low-light or F̂ (u, v) = G(u, v) / H(u, v)
photon-limited images.
- Speckle Noise: Multiplicative noise However, it has limitations:
common in radar and ultrasound - Highly sensitive to noise.
imaging. - If H(u, v) is near zero, division
amplifies the noise, making the result
unstable.
3. Noise Filters
To remove noise from images, various
spatial domain filters are used:
- Mean Filter: Replaces each pixel with
Image Compression and Redundancy 2. Assign shorter codes to
more frequent symbols.
Image compression reduces the amount of
 Efficient and simple but requires
knowledge of the symbol
data required to represent a digital image. probabilities.
This is achieved by removing different types
of redundancy from the image data.

5. Arithmetic Coding

1. Coding Redundancy
 A more advanced lossless
compression method.
 Refers to inefficient representation  Encodes the entire message into a
of pixel values. single number (a fraction between
 Occurs when fixed-length codes 0 and 1).
are used regardless of how  Advantage: Provides better
frequently a pixel value appears. compression than Huffman coding,
 Example: Assigning the same especially when symbol
number of bits to frequent and probabilities are skewed.
rare symbols.  Drawback: More complex and
 Solution: Use variable-length slower than Huffman coding.
coding (like Huffman coding),
where more frequent values get
shorter codes.
6. Lossy Compression Techniques

2. Interpixel Redundancy  Discards some image data

permanently for higher
compression.
 Refers to the correlation between  Exploits psychovisual redundancy.
neighboring pixels.
 Common techniques:
 Since adjacent pixels in an image o Transform Coding
often have similar values, this (e.g., DCT)
similarity can be exploited to o Quantization
reduce data.
o Run-length encoding
 Solution: Predict one pixel from after quantization
its neighbors and encode only the
 Used in formats like JPEG, MPEG,
difference (DPCM – Differential
and WebP.
Pulse Code Modulation).

7. JPEG Compression
3. Psychovisual Redundancy

 A widely used lossy image

 Based on the human visual system
compression standard.
(HVS).
 The eye is less sensitive to small
 Based on Discrete Cosine
Transform (DCT).
color or brightness changes in
certain areas.  Steps:
1. Divide image into 8×8
 Solution: Remove information
blocks.
that is perceptually insignificant to
2. Apply DCT to each
humans.
block.
 Used in: Lossy compression 3. Quantize DCT
techniques like JPEG. coefficients (introduces
loss).
4. Zigzag scan the
quantized matrix.
5. Apply Run-length
4. Huffman Coding
encoding + Huffman
coding.
 A lossless compression technique  Achieves high compression with
based on symbol probabilities. acceptable image quality for
human perception.
 Creates an optimal prefix code (no
code is a prefix of another).
 Steps:
1. Build a binary tree using
frequency of symbols.
Image segmentation is the process of Accumulator array stores the votes; peaks
partitioning an image into meaningful represent detected lines.
regions or objects to simplify its analysis.
Representation involves describing the 5. Region-Based Segmentation
segmented regions using mathematical
features for further processing or analysis.
 Divides image into regions with
similar properties (intensity,
texture, color).
1. Point, Line, and Edge Detection  Techniques include:
Point Detection: Identifies isolated o Region Growing: Starts
pixels with significantly different with seed points and
intensity than surroundings. adds similar
neighboring pixels.
Line Detection: Detects straight lines o Region Splitting and
using masks oriented at different Merging: Recursively
angles. splits image or merges
regions based on a
similarity criterion.
Edge Detection: Locates boundaries
between different regions by detecting
abrupt intensity changes. 6. Boundary Representation
 Focuses on representing the shape
or contour of regions.
Common edge detectors:
 Examples:
o Chain Code: Represents
Sobel Operator boundary as a sequence
of directional moves.
Prewitt Operator o Polygonal
Approximation:
Approximates boundary
Canny Edge Detector – widely used due to
with line segments.
its accuracy.
o B-splines / Fourier
Descriptors: Smooth
2. Thresholding curve representations.
 Simplest method for segmentation.
 Converts grayscale image into 7. Boundary Descriptors
binary:  Quantify boundary shape for
 T is a threshold value. recognition or classification.
 Can be:  Types:
o Global Thresholding: o Simple Descriptors:
Same TTT for whole Perimeter, area,
image. curvature.
o Adaptive o Fourier Descriptors:
Thresholding: Represent shape in
Different TTT for frequency domain.
different regions. o Moment Invariants:
Statistical measures
3. Edge and Boundary Linking invariant to scale,
rotation, translation.
 After detecting edges, linking is
used to form continuous object
boundaries.
 Methods:
o Local Analysis: Connect
nearby pixels with
similar edge
orientation/magnitude.
o Global Analysis: Uses
graph-based or chain-
code techniques to trace
complete boundaries.

4. Hough Transform
 A technique for detecting
geometric shapes like lines and
circles.
 Converts edge points in image
space to parameter space.
 Example (line detection):

Ρ = x coso + y sin o