Unit - II
Image transformation
DFT is the most used
in DSP/DIP because
real-world
signals/images are
finite and stored
digitally
General Signal Representation
A signal can be represented by the following
fundamental components :
• Amplitude – how strong the signal is (height of
the wave).
• Frequency – how fast the signal oscillates
(number of cycles per second).
• Phase – the shift or alignment of the signal in
time/space.
• Duration (time or space support) – over what
interval the signal exists.
Important components in signal
In Frequency Domain (after DFT)
• When we apply DFT to a signal/image, the
important components become:
• Magnitude spectrum → tells how much of each
frequency is present.
• Phase spectrum → tells the alignment/structure
of those frequencies.
• Together, magnitude + phase fully describe the
signal.
If either is missing, we lose essential information.
In Image Processing
For a 2D image signal:
• Low-frequency components →
smooth/average intensity (background).
• High-frequency components → edges, details,
sharpness.
• Phase information → shape & structure of
objects.
Fourier transform
• The Fourier transform is a fundamental
mathematical tool that breaks complex
waveforms into their basic frequency
components. Widely used in science and
engineering, it supports technologies from JPEG
image compression to MRI scans.
Discrete Fourier Transform (DFT)
• The forward transform converts an image into its
frequency components.
• The inverse brings it back to the spatial domain.
• Both are double sums because images are 2D
signals.
• Practical implementations use 2D FFT (Fast
Fourier Transform) for speed.
Properties & Use in DIP
• Converts image to frequency domain.
• Both magnitude and phase are important.
– Magnitude → contains information about
textures, patterns, and how much of each
frequency is present.
– Phase → contains information about object
shapes, edges, and positions.
• Used in filtering, image restoration, image
analysis.
• 2D DFT
For an image f(x,y) of size M×N:
• f(x,y): Input image in spatial domain
• x,y: spatial coordinates (rows & columns)
• F(u,v): frequency-domain representation
• u,v: frequency coordinates
• M,N: image dimensions
Basis function
• A basis function is a building block used to
represent signals or images in a transformed
domain.
• In image processing, any image can be written
as a combination of basis functions.
• Each transform (DFT, DCT, KLT) has its own
type of basis functions.
Example in Transforms
• DFT Basis Functions
– Complex exponentials (sine + cosine waves).
– Each basis function corresponds to a certain frequency.
– The image is expressed as a weighted sum of sinusoids.
• DCT Basis Functions
– Cosine waves of different frequencies.
– Used in JPEG compression (the 8×8 cosine basis patterns).
• KLT Basis Functions
– Eigenvectors of the covariance matrix.
– Adapted to the data, not fixed.
What is a Kernel in DFT?
• In mathematics, a kernel is the “building
block” or basis function used in a transform.
• For the 2D DFT, the kernel is:
• So the DFT equation can be written as:
• Here, WM,N is the DFT kernel.
Feature DFT DCT KLT (KL Transform)
Complex
Eigenvectors of
Basis Functions exponentials Cosine only
covariance
(sin+cos)
Complex
Coefficients (magnitude + Real numbers Real numbers
phase)
Energy Compaction Good Very good Optimal
Expensive (eigen
Computation Fast via FFT Fast via Fast DCT
decomposition)
Compression
Filtering, spectral Image/video
Applications (theoretical),
analysis compression (JPEG)
pattern recognition
Adaptive (depends
Fixed / Adaptive Fixed basis Fixed basis
on data)
In Time/Spatial In Frequency
Component Example (Audio) Example (Image)
Domain Domain
Instantaneous Magnitude of
value of the signal Fourier coefficients Loudness of sound Brightness value of
Amplitude
(height of wave / (how strong each at a moment a pixel
pixel intensity) frequency is)
Smooth
Peaks in the Pitch of sound
Not directly visible background = low
Frequency spectrum show (e.g., 440 Hz for
in raw signal freq, Edges/details
frequency content "A" note)
= high freq
Determines
waveform shape Determines
Phase spectrum
Not obvious in time (two same position &
Phase (angle of Fourier
domain frequencies can structure of
coefficients)
sound different if edges/shapes
phase differs)
Resolution in
frequency domain Image size
Signal length in Song length (5 sec
Duration / Extent (longer signal → (256×256 vs
time / size in space vs 5 min)
finer frequency 1024×1024)
resolution)