Chapter 4

Digital Image Transforms

© Oxford University Press 2011

 NEED FOR IMAGE TRANSFORMS
• Image processing operations can be applied in both spatial domain and
frequency domain
• Image operations can be performed directly in the spatial domain by
manipulating the pixels array.
• Frequency domain deals with the frequency components of the image.
• Real time applications require faster algorithms as the requirements of
applications such as fingerprint attendance, airport baggage, screening
radar are speed and simplicity.
• The essential point in frequency domain is the transformation of a
problem to a convenient domain for performing calculations more easily.
• Similarly the spatial and frequency domains are precisely different
representations of the same information .
• Sometimes the frequency domain is simpler.
• For these conversions, mathematical transforms such as fourier
transform and cosine transform is used.
• Inverse transforms convert the results of the frequency domain
manipulations back to the original spatial domain.
© Oxford University Press 2011
NEED FOR IMAGE TRANSFORMS

© Oxford University Press 2011

Spatial Frequencies in Image Processing
• Frequency is used to refer the rate of repetition of signals in the
spatial domain with respect to time.
• ‘spatial frequency’ is used to describe the rate of change of pixel
intensity of an image in space. It can be visualized using the line
profile of the image in any row or column.
• The x-axis represents the distance of the pixels from the origin and
y-axis shows the grey value which ranges from 0 to 255.
• The intensity profiles are very useful in image processing.
• Intensity profile is a collection of either sinusoidal signals such as
sine, cosine and harmonic signals or non-sinusoidal signals such as
square waves.
• High frequency components in the images are rapidly changing
data.
• Edges are examples of high frequency components.
• The low-frequency components are contrast and other large scale
features of the image.
© Oxford University Press 2011
Image Profile

© Oxford University Press 2011

 Types of image transforms
• Essentially there are two categories of transforms available
1. Orthogonal and 2. Non-sinusoidal.
• Orthogonal transforms use basis functions that are sinusoidal in nature.
Example : Fourier and Cosine transforms.
• The non-sinusoidal basis functions that are not sinusoidal in nature.
Example : Haar, Walsh and Hadamard transforms.
Practical Applications of the Image Transforms :
• Transforms convert spatial information to frequency domain where
operations are easier to perform.
• Example convolution algorithm reduced matrix multiplication in the
frequency domain.
• It is useful in gaining insight into concepts such as sampling.
• The property of energy transformation states that signal energy is
preserved.
• The property of energy compaction states that a large fraction of the
average energy of the image is packed into a few components.

© Oxford University Press 2011

 Types of image transforms
• It also useful in image compression, as zeros need not be sent across the
transmission channel.
• It helps to design faster algorithms.
• Spatial frequency transforms help in assessing the image quality.
• Frequency components indicate repetitive noise patterns and also the
presence or absence of high/low/mid frequencies that exist in the image.
• Most of the popular transform is unitary . (A * A-1=I i.e Orthogonal matrix)
• If the transforms result in a matrix containing only real numbers such a
matrix is called orthogonal.
• If the matrix includes complex numbers is called hermittian.

© Oxford University Press 2011

 Basis of transforms
• Let f(x) be a vector size N, with input samples f=[f(0),f(1)…f(N-1)]T. Let T
be the transformation is applied to the input sequence and yield the
resultant vector of the form g(u).
• g(u)=T * f where g(u) is the transformation vector of f(x) and the matrix
T(u,x) is called the forward transformation kernel.
• The inverse transformation can be applied to get the original vector and is
given by
• f(x)=I * g(u) where I(x,u) is the inverse transformation kernel. Or it can be
written as f=T-1 g.
• The images are two-dimensional signals. For an N X N image f(x,y) the
forward and inverse transforms are as follows:

© Oxford University Press 2011

Basis of transforms

© Oxford University Press 2011

 Introduction to Fourier Transform
• The important feature of the fourier transform is the signal.
• The signal is the sum of a collection of sine and/or cosine waves of
different frequencies and amplitudes multiplied by the weighing function.
• It is a unitary transform and based on the concepts of basis vectors and
basis images.

© Oxford University Press 2011

• The inverse Fourier transform is expressed
mathematically as

© Oxford University Press 2011

 Discrete Fourier Transform
• Since the images are digitized, it is necessary to have a discrete
formulation of the fourier transform.
• So DFT takes the regular spaced data values and returns the value of the
Fourier transform by replacing the images by summation.

© Oxford University Press 2011

© Oxford University Press 2011
© Oxford University Press 2011
© Oxford University Press 2011
© Oxford University Press 2011
 Fast Fourier Transform
• The 2D DFT can be calculated by applying the 1D DFT for each column
and then performing a 1D DFT for each row on the resulting complex
data.
• This algorithm is computationally intensive. Hence the faster version is
required.
• The FFT algorithm is developed using a strategy called successive
doubling method is widely used.

© Oxford University Press 2011

FFT ALGORITHM

© Oxford University Press 2011

 FFT

© Oxford University Press 2011

INVERSE FFT

© Oxford University Press 2011

PROPERTIES OF FOURIER
TRANSFORM

© Oxford University Press 2011

© Oxford University Press 2011
 Discrete cosine Transform
• DCT is one of the most popular transforms in image processing.
• It is widely used in image compression because of its high energy
packing capabilities.
• DCT has many properties and involves only real components.
• Similarly FFT, a faster version of DCT is also available.

© Oxford University Press 2011

© Oxford University Press 2011
 Haar Transform
• The basis of the Haar transform consists of elements +1,-1 or 0.
• The kernel of the Haar transform is generated using the following
procedure.

© Oxford University Press 2011

Haar Transform

© Oxford University Press 2011

© Oxford University Press 2011
© Oxford University Press 2011