• Image Processor: • Euclidean: The straight-

1. Define Sampling and Hardware that performs line distance (“as the crow
Quantization. high-speed processing flies”).
(like filtering). • City-block ( 𝐷4 ): Distance
Explain the effects of reducing
spatial resolution and gray- • Computer: The central measured by moving only
level resolution. unit (PC) that controls the horizontally and vertically
system and runs software. (like a taxi in a city grid).
• Sampling: The process of • Storage: Stores images. • Chessboard ( 𝐷8 ):
digitizing the coordinate Short-term: RAM; Long- Distance measured by the
values (𝑥, 𝑦) of an image. term: Hard drive. maximum number of steps
It determines the spatial • Display: A monitor to taken in any direction
resolution (number of view the output images. (like a King in chess).
pixels).
• Hard Copy: Printers for 4. Fundamental Steps in Digital
• Quantization: The physical output.
process of digitizing the Image Processing
• Networking: Transmits
amplitude (intensity) 1. Image Acquisition:
image data to remote
values. It determines the Capturing the image
locations.
gray-level resolution using a sensor.
(e.g., number of 3. Relationships Between 2. Image Enhancement:
colors/shades, such as Pixels Subjectively improving
256 levels).
Neighbors of a pixel 𝑝 at the image to look better
Effects of Reducing coordinates (𝑥, 𝑦) : to the human eye.
Resolution: 3. Image Restoration:
• 𝑵𝟒 (𝒑) The 4 pixels Objectively recovering
• Spatial Resolution: sharing an edge an image from
Reducing this causes a (horizontal/vertical) with damage/noise using
“checkerboard effect,” pixel 𝑝 . mathematical models.
where pixels become large • 𝑵𝑫(𝒑) The 4 diagonal 4. Color Image
and visible like blocks neighbors of pixel 𝑝 . Processing: Handling
(aliasing). • 𝑵𝟖 (𝒑) All 8 neighbors color models like RGB
• Gray-Level Resolution: (both 𝑁4 and 𝑁𝐷 ) or CMY.
Reducing this causes surrounding pixel 𝑝 . 5. Wavelets &
“false contouring,” where
Adjacency: Multiresolution:
smooth transitions look
Representing images at
like sharp ridges or bands
• 4-adjacency: Two pixels different
because there aren’t
are connected if they are sizes/resolutions.
enough colors to show the
𝑁4 neighbors and have 6. Compression:
gradient smoothly.
similar values. Reducing file size for
2. General-Purpose Image • 8-adjacency: Two pixels storage.
Processing System (Block are connected if they are 7. Morphological
Diagram Components) 𝑁8 neighbors. Processing: Extracting
• m-adjacency: A mix of 4 image components like
• Image Sensors: Captures shapes or boundaries.
and 8 adjacency used to
the physical scene and
eliminate “ambiguity” or 8. Segmentation:
converts it into an
multiple paths found in 8- Partitioning an image
electrical signal.
adjacency. into objects (e.g.,
• ADC (Analog-to-Digital separating a face from
Converter): Converts the Distance Measures:
the background).
electrical signal into
9. Representation &
digital data.
Description:
 Describing regions for 5. Round: Round the without blurring edges as
computer processing. result to the nearest much as a mean filter. It is
10. Object Recognition: integer to get the new non-linear because it
Identifying what the mapped gray level. relies on sorting/ranking
object is. values.
Example:
• Max Filter: Removes
5. Elements of Visual “pepper” (black) noise.
• Pixels: [1,5,1,5] . Max
Perception
Level: 7. • Min Filter: Removes “salt”
• Rods & Cones: Rods are • 𝑃(1) = 0.5, 𝑃 (5) = 0.5 . (white) noise.
for low light (night • 𝑆𝑘 for 1 is 0.5 . 𝑆𝑘 for 5 is Sharpening Filters
vision); Cones are for 0.5 + 0.5 = 1.0 . (Derivative Filters):
color and detail (day • New values: 1 → 0.5 × 7 =
vision). 3.5 ≈ 4 . 5 → 1.0 × 7 = 7 . • Laplacian: A second-
• Result: [4,7,4,7] . derivative operator that
• Brightness Adaptation: detects edges in all
The eye adjusts to the 2. Gray Level Transformations directions (isotropic).
average brightness level;
• Image Negatives: • Gradient
it cannot see all light
Reverses pixel values ( (Sobel/Prewitt): First-
levels simultaneously.
𝑠 = 𝐿 − 1 − 𝑟 ). Used for derivative operators that
• Mach Bands: An optical medical images like X- detect edges in specific
illusion where the eye rays. directions (horizontal or
perceives exaggerated vertical).
• Log Transformations:
brightness changes at 𝑠 = 𝑐 ⋅ log(1 + 𝑟) . • High-Boost Filtering:
edges. Expands dark pixels. Good Subtracts a blurred
for brightening images version from the original
1. Histogram Equalization to get edges, multiplies
with a large dynamic
Definition: A technique used range (e.g., Fourier edges by a boost factor 𝐴 ,
to improve the contrast of an spectra). and adds them back to the
image by spreading out the original. Makes the image
• Power-Law (Gamma):
intensity values so they are “crispier.”
𝑠 = 𝑐 ⋅ 𝑟 𝛾 . If 𝛾 < 1 , it
uniform. brightens; if 𝛾 > 1 , it Note on “Point Processing”
darkens. vs “Spatial Filtering”:
Numerical Method (Step-by-
Step): 3. Spatial Filters (Smoothing vs. • Point Processing:
1. List gray levels ( 𝑟𝑘 ) Sharpening) Changes a pixel based only
and count the frequency on that single pixel (e.g.,
Linear Smoothing Filters: Log transform).
( 𝑛𝑘 ).
2. Calculate Probability ( • Mean Filter: Replaces a • Spatial Filtering:
𝑃(𝑟𝑘 ) ): 𝑃(𝑟𝑘 ) = pixel with the average of Changes a pixel based on
𝑛𝑘 /𝑡𝑜𝑡𝑎𝑙𝑝𝑖𝑥𝑒𝑙𝑠 . its neighbors. Blurs the its neighbors using a
3. Calculate Cumulative image to reduce noise but mask/kernel.
Probability ( sk) : Add also blurs edges.
the current 𝑃 (𝑟𝑘 ) to the Non-Linear Smoothing UNIT 3: Image Restoration
sum of all previous Filters (Order-Statistics):
ones. 1. Image
• Median Filter: Replaces a Degradation/Restoration
4. Map to new levels:
pixel with the median Model
Multiply 𝑆𝑘 by
(𝑀𝑎𝑥𝐺𝑟𝑎𝑦𝐿𝑒𝑣𝑒𝑙 − 1) value of its neighbors. Formula:
(usually 7 for 3-bit or Excellent for removing
255 for 8-bit). “Salt-and-Pepper” noise
 𝑔(𝑥, 𝑦) dividing small values of 𝐻 thermal imaging). Humans
= ℎ(𝑥, 𝑦) ∗ 𝑓 (𝑥, 𝑦) + 𝜂(𝑥, 𝑦) (near zero) by noise distinguish colors better
amplifies the noise to than shades of gray.
• 𝑔(𝑥, 𝑦) : The degraded infinity.
(observed) image. • Intensity Slicing:
• 𝑓 (𝑥, 𝑦) : The original true • Wiener Filtering: Assigning a specific color
image. Considers both to a range of gray levels.
• ℎ(𝑥, 𝑦) : The degradation degradation and statistical
noise. It minimizes the • Color Slicing: Isolating a
function (blur/PSF).
Mean Square Error specific color (e.g.,
• 𝜂(𝑥, 𝑦) : The additive “Strawberry Red”) by
noise. between the estimated
and original image. It keeping pixels within a
2. Enhancement vs. works much better in the defined range/sphere and
Restoration presence of noise. turning others gray.

• Enhancement: Subjective. UNIT 4: Color Image UNIT 5: Image Compression

Manipulates image to be
Processing 1. Redundancy Types
more pleasing to the
human eye. 1. Color Models • Coding Redundancy:
• Restoration: Objective. Using more bits than
Tries to recover the • RGB (Red, Green, Blue):
necessary to represent
original image based on a Additive model. Used for
symbols (e.g., fixed-length
mathematical model of the monitors/cameras (light
coding). Solved by
damage. is added).
Huffman Coding.
• CMY/CMYK (Cyan,
3. Noise Models • Interpixel Redundancy:
Magenta, Yellow, Black):
Neighboring pixels are
Subtractive model. Used
• Gaussian Noise: highly correlated (similar
for printing (ink absorbs
Statistical noise with a values). Solved by
light).
PDF equal to the normal Predictive Coding (DPCM).
distribution. Common in • HSI (Hue, Saturation,
• Psychovisual
circuits. Intensity): Closer to
Redundancy: Information
human perception.
• Rayleigh Noise: the human eye cannot see
Asymmetrical PDF, used – Hue: The pure color. (high-frequency details).
in range imaging. – Saturation: The Solved by Quantization
• Impulse (Salt-and- purity/richness of the (Lossy compression).
Pepper) Noise: Random color.
– Intensity: Brightness. 2. Lossy vs. Lossless
occurrences of pure black
Compression
and pure white pixels.
Why HSI over RGB? RGB
• Periodic Noise: mixes color and brightness. To • Lossless: No data lost.
Repeating patterns caused change brightness in RGB, you Reconstructed image is
by electrical interference. must change all three values. In identical (ZIP, PNG,
Removed by Frequency HSI, you can change ‘I’ without Huffman, RLC). Used for
Domain filters (Band affecting ‘H’ or ‘S’, making it medical/legal documents.
Reject). better for image processing. • Lossy: Data is discarded
for higher compression.
4. Inverse vs. Wiener Filtering 2. Pseudo-Color and Color Reconstructed image is an
• Inverse Filtering: Divides Slicing approximation (JPEG,
the degraded image by the • Pseudo-Color: Assigning MP3).
degradation function ( “false” colors to a
1/𝐻 ). It fails in the 3. Compression Techniques
grayscale image to make • Huffman Coding:
presence of noise because details visible (e.g., Lossless. Assigns shorter
 binary codes to more captures a single tiny Spatial
frequent symbols. point of light. Resolutio Gray-Level
Feature n Resolution
– Example: A(0.6), B(0.3), Comparison Tables Visual Aliasing: False
C(0.1). Combine B+C → • Continuous vs Discrete Artifacts Blocky, Contouring
0.4. Sort: A(0.6), BC(0.4). A pixelated : Banding
Image: Continuous has
edges effects in
gets ‘0’, BC gets ‘1’. B gets infinite smooth values; (“jaggies smooth
‘10’, C gets ‘11’. Discrete has finite integer ”). gradients.
• Run Length Coding pixels.
(RLC): Replaces identical • Mean vs Median Filter: 2. Continuous Image vs.
sequences with a count. Mean blurs edges; Median Discrete Image
Example: preserves edges while Featur Continuous Discrete
“WWWWWBBBB” → removing impulse noise. e Image Image
“5W4B”. • Image Processing vs Natur Analog signal; Digital
• Bit-Plane Coding: Splits Analysis: Processing e of values are signal;
outputs an image; Data continuous values are
image into 8 binary functions integer
planes. Compresses each Analysis outputs
𝑓 (𝑥, 𝑦) . arrays
separately. data/description. 𝐴[𝑚, 𝑛] .
• Adaptive vs Non- Doma Infinite Finite
• JPEG Standard: Lossy. Adaptive Filters: Non- in precision in precision;
Uses DCT (Discrete Cosine adaptive treats the whole both limited by
Transform) to convert image the same. Adaptive coordinates grid size
spatial data to frequency. changes behavior based and amplitude. and bit
depth.
Quantization removes on local content (e.g.,
high-frequency details stops smoothing at edges). Noise Electronic/An Quantizati
Type alog noise is on noise
(loss step). Zig-Zag continuous. (rounding
scanning orders DIFFERENTIAIONS
errors) is
coefficients for coding. 1. Spatial Resolution vs. introduce
d.
Gray-Level Resolution
Key Definitions
• Pixel: Smallest element of Spatial 3. Under-sampling vs. Over-
a digital image. Resolutio Gray-Level sampling
Feature n Resolution
• Sampling Theorem Under- Over-
(Nyquist Rate): Samples Primary Number Number of
Feature sampling sampling
Definitio of pixels bits per
must be taken at a rate at n (density) pixel (color Relations Sampling Sampling
least twice the highest in the depth). hip to frequency frequency
frequency to avoid image Nyquist < 2 × Max ≫ 2 × Max
aliasing. grid. Frequenc Frequency.
y.
• Aliasing: Error caused by Determin The The Analog-
ed By physical to-Digital Reconstru Impossib Easy:
under-sampling,
sensor Converter ct ability le: Signal is
appearing as “jaggies” or size and (ADC) Original easily
moiré patterns. density sensitivity signal reconstruc
• Compression Ratio: (samplin (quantizatio cannot be ted with
g). n). recovered simple
Ratio of uncompressed
perfectly. filters.
size to compressed size ( Impact Increase Increases
on File s with linearly Frequenc High High
𝐶𝑅 = 𝑛1 /𝑛2 ).
Size the with bit y Effect frequenci frequencie
• PSF (Point Spread square of depth. es s are
Function): The shape of the masquera captured
the blur blob formed dimensio de as low accurately;
n( frequenci noise floor
when an imaging system
𝑁 × 𝑀 ). es is lowered.
(Aliasing)
 Under- Over- Linear Non-
Feature sampling sampling Filters Non-Linear Adaptive Adaptive
. (e.g., Filters (e.g., Feature Filters Filters
Feature Mean) Median) Behavio Static; Dynamic;
4. Image Processing vs. Mathema Linear Logical/Ord r applies the kernel
Image Analysis tical Ops arithmeti ering: same parameters
c: Sorting, kernel change
Image Addition Ranking, everywher based on
Processi Image and Max/Min. e. local image
Feature ng Analysis Multiplic content.
ation Complex Simple Complex
Input / Input: Input: Image
(Convolut ity convolutio logic
Output Image → → Output:
ion). n; fast (calculating
Output: Data/Descrip
Image. tion. Frequenc Predictab Difficult to computati local
y le; can be model or on. variance/st
Level of Low-to- High-level
Analysis analyzed predict in ats);
Abstract Mid processing
easily in the slower.
ion level (closer to
Fourier Frequency Assumpt Assumes Handles
processi AI/Computer
domain. domain. ion noise/feat non-
ng. Vision).
Reversib Often Generally ures are stationary
Goal To To extract
ility reversibl irreversible uniform images
improve meaning,
e (via (informatio across the (varying
visual quantify
inverse n is image. noise
quality objects, or
filtering). discarded levels/text
or recognize
during ure).
prepare patterns.
sorting).
for
analysis.
7. Mean Filter vs. Median 9. Image Enhancement vs.
5. Point Processing vs. Filter Image Restoration
Spatial Filtering Mean Median Image
Feature Filter Filter Enhanceme Image
Point
Noise Spreads Complete Feature nt Restoration
Processin Spatial
Feature g Filtering Handling noise out ly Criteria Subjective: Objective:
(blurs it); eliminate “Does it “Is the
Scope Operates Operates look good mathematical
poor for s “salt-
on 1 pixel on a to the error
“salt-and- and-
at a time ( neighbor human minimized?”
pepper”. pepper”
1×1 hood of eye?”
(impulse
neighborh pixels Prior Not Required
) noise.
ood). (e.g., 3 × 3 Knowled required (needs
, 5 × 5 ). Edge Poor: Excellen ge (trial and knowledge of
Preservati Blurs t: error). the
Computat Very low; Higher;
on edges Preserve Degradation
ional Cost easily requires Function 𝐻 ).
significant s sharp
parallelize convolutio
ly. edges Techniqu Heuristic Inverse
d per n
while e approaches modeling
pixel. operation (Contrast (Deconvoluti
removing
s and stretch, on, Wiener
noise.
memory Unsharp Filter).
buffers. Outlier Sensitive: Robust: mask).
Sensitivity One bright Extreme
Primary Contrast Blurring,
bad pixel outliers 10. Random Noise vs.
Use adjustmen sharpenin
pulls the are
t, g, edge
average ignored Periodic Noise
thresholdi detection.
up. in the
ng, Random Periodic
sorting
brightness Feature Noise Noise
process.
. Frequen Broad Narrow
8. Adaptive vs. Non-Adaptive cy spectrum spectrum
6. Linear vs. Non-Linear Spectru (spread (appears
Filters
Filters across all as specific
 Random Periodic Spatial Frequenc RGB Human
Feature Noise Noise Feature Domain y Domain Feature Model Perception
m frequencies spikes/dot Best For Local Global Sensitivit Linearly Non-linear:
). s). defects defects y equal Eyes are
Origin Thermal Electrical (salt-and- (periodic sensitivit much more
sensors, interferen pepper interferen y to R, G, sensitive to
low light, ce, motor noise). ce, and B Green than
transmissio vibrations, motion sensors. Blue.
n errors. 60𝐻𝑧 blur). Luminan Brightne Luminance
hum. ce ss is (rods) and
Remova Spatial Frequency mixed Chrominanc
l smoothing domain 13. RGB vs. CMY/CMYK into all e (cones)
Method (Mean/Med Notch or three are
ian filters). Band- CMY/CMYK color processed
Reject RGB (Red, (Cyan, channels. separately.
filters. Green, Magenta, Correlati High Channels
Feature Blue) Yellow, Key) on correlati are
11. Inverse Filtering vs. Model Additive: Subtractive: on perceived
Wiener Filtering Type Starts Starts with between independent
with white paper, channels ly.
Inverse Wiener black, ink subtracts (e.g., a
Feature Filtering Filtering adds light light. shadow
to get darkens
Method Direct Statistical
white. R, G, and
inversion: inversion:
B).
𝐹(𝑢, 𝑣) = Considers Primary Digital Printers,
𝐺 (𝑢, 𝑣)/ Noise-to- Usage displays, Plotters,
𝐻 (𝑢, 𝑣) . Signal Cameras, Physical 16. Lossy vs. Lossless
ratio. Scanners. media. Compression
Handling Catastroph Robust: Gamut Generally Narrower
Lossless Lossy
Noise ic: Suppress (Range) wider gamut
Compressi Compressi
Amplifies es color (difficult to
Feature on on
noise where frequenci gamut. print bright
𝐻 is small es where neon RGB Data 100% Irreversib
(dividing by noise colors). Integrity Reversibl le:
zero). dominate e: Original Original ≠
s. 14. HSI vs. HSV = Reconstru
Reconstru cted.
Optimali Not optimal Optimal
HSI HSV cted.
ty in the (minimiz
presence of es Mean Feature (Intensity) (Value) Compres Low High
noise. Square Lightne 𝐼= 𝑉= sion (typically (typically
Error). ss Def. 𝐴𝑣𝑒𝑟𝑎𝑔𝑒(𝑅, 𝐺, 𝑀𝑎𝑥𝑖𝑚𝑢𝑚
𝐵) (𝑅, 𝐺, 𝐵) Ratio 2: 1 or 3: 1 10: 1 to
. . ). 50: 1 ).
12. Spatial vs. Frequency Indepen Ideally Color and Artifacts None. Blocking,
Domain Restoration dence decouples brightness Perfect blurring,
color (H/S) are quality. or ringing
Spatial Frequenc from somewhat artifacts
Feature Domain y Domain brightness correlated appear at
(I). in V. high
Operation Convoluti Multiplica
ratios.
on with a tion of Applicat Best for Best for
kernel transform ion Image Human Use Cases Medical Web
mask. arrays. Processing/ Computer imaging, streaming,
Analysis Interaction Legal docs, Social
Computati Faster for Faster for
algorithms. (Color Zip files. media,
onal small large
Pickers). Consumer
Speed kernels/fil filters/glo
photograp
ters. bal
15. RGB vs. Human hy.
operation
s ( 𝑁log𝑁 Perception
complexit 17. JPEG vs. MPEG
y).
Feature JPEG MPEG • Power: 𝑠 = 𝑐𝑟 𝛾 • Neighbors: 𝑁4 (Edges),
Redunda Exploits Exploits • Entropy: 𝐻 = −∑𝑝log 2 𝑝 𝑁𝐷 (Corners), 𝑁8 (All).
ncy Type Spatial Spatial • Adjacency:
redundan AND 1. Overview
cy Temporal
– 4-adj: Connected if values
(within redundanc • Image Processing: adjacent in 𝑁4 .
one y (between Manipulation of digital – 8-adj: Connected if values
image). frames). images by computer. adjacent in 𝑁8 .
Motion None Uses • Types: Binary (1-bit), (Ambiguous).
Handling (treats Motion
Grayscale (8-bit), Color – m-adj: Fixes 8-adj
every Vectors to
image as predict (24-bit RGB). ambiguity (removes
static). pixel • System: Sensor → ADC → closed loops).
movement. Processor → Memory → • Distance:
Key DCT DCT + Display. – Euclidean: Real distance
Compone (Discrete Motion
(√…).
nt Cosine Compensat 2. Visual Perception
Transfor ion + Audio – D4 (City): |𝑥1 − 𝑥2 | +
m) on compressio • Rods: Scoptic (Night), |𝑦1 − 𝑦2 |.
blocks. n. B/W, Low Res. – D8 (Chess):
• Cones: Photopic (Day), max(|𝑥𝑑𝑖𝑓𝑓 |, |𝑦𝑑𝑖𝑓𝑓 |).
18. Types of Redundancy
Color, High Res.
Interpi Psycho • Weber Ratio: 𝑘 = 𝛥𝐼/𝐼. 5. Transforms
Coding xel visual Measures eye’s sensitivity
Redun Redun Redun • DFT: Spatial → Freq.
to brightness change. (Complex numbers).
Feature dancy dancy dancy
Lower 𝑘 = Better
Cause Ineffici Correl Percep • DCT: Real numbers. Packs
sensitivity.
ent ation tion energy into fewer
coding (neigh (eye • Mach Bands: Optical coefficients. Used in JPEG.
(using bors cannot illusion where edges look
8 bits usually see brighter/darker than
for a look certain 🟡 Unit 2: Enhancement
reality (Human visual
value like fine
that neighb details) system overshoot). 1. Histogram
appear ors). .
s 90% 3. Sampling & Quantization • Def: Graph of frequency of
of the gray levels 𝑟𝑘 .
time).
• Sampling: Digitizing
• Equalization: Flattens
Coordinates (𝑥, 𝑦).
Solutio Variabl Predict Quanti histogram → Uniform Dist
n e ion / zation Determines Spatial
→ Max Contrast.
Length Run (discar Resolution (No. of pixels).
Coding Length ding – Error: Under-sampling → – Step: 𝑠𝑘 = (𝐿 − 1)∑𝑃(𝑟𝑗 ).
(Huffm Encodi high Aliasing (Moiré patterns,
an). ng. freque 2. Point Transforms
ncies).
Jaggies).
– Nyquist: Sample at > • Negative: 𝑠 = 𝐿 − 1 − 𝑟.
Revers Rever Rever Irreve
ibility sible sible rsible 2𝑓𝑚𝑎𝑥 . (X-ray).
(Lossle (Lossle (Lossy) • Quantization: Digitizing • Log: 𝑠 = 𝑐log(1 + 𝑟).
ss). ss). . Amplitude (Intensity). (Expands dark/Fourier).
Determines Gray • Power/Gamma: 𝑠 = 𝑐𝑟 𝛾 .
📐 Formulas Resolution. (Monitor correction).
• Degradation: 𝑔 = ℎ ∗ 𝑓 + – Error: Low levels → False
3. Operations
𝜂 Contouring (Bands).
• Geometric Mean: 𝑓 = • Subtraction: 𝑔 = 𝑓 (𝑡2 ) −
1 4. Pixel Relationships
[𝛱𝑔] 𝑚𝑛 𝑓 (𝑡1 ). Detects Motion.
• Log: 𝑠 = 𝑐log(1 + 𝑟)
•
1
Averaging: 𝑔 = 𝑁 ∑𝑓𝑖 . 4. Frequency Filters • Redundancy: Coding
(Bad codes), Interpixel
Reduces Random Noise. • Band Reject: Removes (Neighbors), Psychovisual
• Logic: AND (Masking), OR ring of freqs (Periodic (Invisible detail).
(Merging). Noise).
• Notch: Removes specific 2. Lossless
4. Spatial Filtering
points (Interference
• Huffman: Variable length.
• Smoothing (Low Pass): spikes).
Frequent → Short code.
– Mean: Linear. Averages • Inverse: 𝐹 = 𝐺/𝐻. Fails (Bottom-up tree).
neighbors. Blurs edges. with noise (𝑁/𝐻 → ∞).
• Shannon-Fano: Split
– Median: Non-linear. Sorts • Wiener: Min MSE. probs top-down.
& picks middle. Best for Accounts for Noise-to-
• Run Length (RLC): Count
Salt/Pepper noise. Signal ratio. Optimal.
runs (AAAA → 4A). Binary.
Preserves edges.
5. Geometric Transforms • Bit-Plane: Split 8 planes.
• Sharpening (High Pass):
Compress each with RLC.
– Laplacian: ∇2 𝑓. Isotropic • Rubber Sheet: Distorting
(Rotation invariant). coordinate system. 3. Lossy
– Sobel: 1st Deriv. • Affine: Scales, Rotates,
• Predictive: DPCM. Encode
Directional edges. Translates, Shears.
error (difference).
• Interpolation: Nearest
🟠 Unit 3: Restoration • Transform: DCT. Drop
Neighbor (Blocky) vs
high freq coefficients.
Bilinear (Smooth).
1. Model • JPEG: 8x8 Blocks → DCT
🔵 Unit 4: Color → Quantize (Loss) →
• 𝑔(𝑥, 𝑦) = ℎ(𝑥, 𝑦) ∗ ZigZag → Huffman.
𝑓 (𝑥, 𝑦) + 𝜂(𝑥, 𝑦). 1. Models • MPEG: Video. Uses JPEG +
• Goal: Estimate 𝑓 (Original) Inter-frame (Motion)
using 𝑔 (Degraded) and ℎ • RGB: Additive (𝑅 + 𝐺 +
compression.
(Blur). 𝐵 = 𝑊). Monitors.
• CMY(K): Subtractive (𝐶 +
2. Noise Types 𝑀 + 𝑌 = 𝐾). Print.
• Gaussian: Normal bell • HSI: Hue (Color), Sat
curve (Electronic/Sensor). (Purity), Int (Bright).
• Impulse: Salt (White), Decouples intensity.
Pepper (Black).
2. Processing
• Periodic: Interference
bars. • Pseudocolor: Assign
colors to Gray levels
3. Spatial Filters (Thermal/X-ray).
• Arithmetic Mean: 𝑚𝑛 ∑𝑔.
1 • Full-Color: Vector
processing (RGB).
Averages noise (Blurs).
• Transform: RGB → HSI →
• Geometric Mean: Process I → RGB.
1
[∏𝑔(𝑠, 𝑡)]𝑚𝑛 . Less blur
than Arith. 🟣 Unit 5: Compression
• Median: Order statistic.
1. Basics
Removes Impulse.
• Adaptive: Adapts to local • CR: 𝑛1 /𝑛2
variance (smooths noise, (Uncompressed/Compres
keeps edges). sed).