Dr.

Qadri Hamarsheh

Fundamentals of spatial Filtering

Outline of the Lecture
 Introduction.
 Spatial Correlation and convolution.
 Vector Representation of linear filtering

Introduction
Filters in frequency domain:
 Lowpass filter that passes low frequencies: used for smoothing (blurring) on the
image.
 Highpass filter that passes high frequencies: used for sharpening the image.
 Bandpass filter.
Filters in spatial domain:
 Spatial filters used different masks (kemels, templates or windows).
 There is a one-to-one correspondence between linear spatial filters and filters in
frequency domains.
 Spatial filters can be used for linear and nonlinear filtering. (Frequency domain filters
just for linear filtering).
 The mechanics of spatial filtering spatial filters consists of:
1. Neighbourhood (small rectangle).
2. Predefined operation that is performed on the image pixel.

Origin
(x, y)

3*3 neighbourhoods of
(x,y)
Image f spatial domain

Figure 1
 Filtering creates new pixel with coordinates equal to the coordinates of the centre of
the neighbourhood, and whose value is the result of the filtering operation.
o If the operation performed on the image pixel is linear, then the filter is called
a linear spatial filter, otherwise, the filter is nonlinear.
o Figure 1 presents the mechanics of linear spatial filtering using a 3*3
neighborhood.
o the response (output) ( , ) of the filter at any point ( , ) in the image is the
sum of products of the filter coefficients and the image pixels values:
 Dr. Qadri Hamarsheh
( , )= (− , − ) ( − , − ) + (− , ) ( − , ) +
… + ( , ) ( , ) +. . . + ( , ) ( + , + )
Observe that the center coefficient of the filter, ( , ) aligns with the pixel at
location( , ).
General mask of size ∗ :
Assume that
= +
and
= +
(where , are positive integers).
(Odd filters)
In general, linear spatial filtering of an image of size ∗ with a filter of size ∗ is
given by the expression:

( , ) = ( , ). ( + , + )

Where and are varied so that each pixel in visits every pixel in .

Image origin Image pixels

W(-1,-1) W(-1,0) W(-1,1)
W(0,-1) W(0,0) W(0,1)
W(1,-1) W(1,0) W(1,1)
Filter coefficients
Filter mask F(x-1, y-1) F(x-1, y) F(x-1, y+1)

image
F(x, y-1) F(x, y) F(x, y+1)

F(x+1, y-1) F(x+1, y) F(x+1, y+1)

Pixels of image
Figure 1
‘‘Section under filter''

Spatial Correlation and convolution

 Correlation: the process of moving a filter mask over the image and computing the
sum of products at each location.
 Convolution: the same process as correlation, except that the filter is first rotated
by
 Dr. Qadri Hamarsheh
Example: 1-D illustration: (figure 2)
Assume that f is a 1-D function, and w is a filter

Figure 2
 Dr. Qadri Hamarsheh

Notes:
 There are parts of the functions (images) that do not overlap (the solution of this
problem is pad f with enough 0s on either side to allow each pixel in w to visit every
pixel in f.
 If the filter is of size m, we need (m-1) 0s on either side of f.
 The first value of correlation is the sum of products of f and w for the initial position
(Figure 2.c).
(The sum of product =0) this corresponds to a displacement x =0
 To obtain the second value of correlation, we shift w are pixel location to the right
(displacement x=1) and compute the sum of products (result =0).
 The first nonzero is when x=3, in this case the 8 in w overlaps the 1 in f and the
result of correlation is 8.
 The full correlation result (figure 2.g) -12 values of x
 To work with correlation arrays that are the same size as f, in this case, we can crop
the full correlation to the size of the original function. (Figure 2.h).
 The result of correlation is a copy of w, but rotated by
 The correlation with a function with a discrete unit impulse yields a rotated version of
the function at the location of the impulse.
 The convolution with a function with a discrete unit impulse yields a copy of that
function at the location of the impulse.
Correlation and convolution with images
 With a filter of size m*n, we pad the image with a minimum of m-1 rows of 0s at
the top and the bottom, and n-1 columns of 0s on the left and right.
 If the filter mask is symmetric, correlation and convolution yield the same result.
Summary:
 Correlation of a filter w(x, y) of size m ∗ n with an image f(x, y) denoted as
a b
W(x, y) °f(x, y) = w(s, t)f(x + s, y + t)
s=−a t=−b
 In similar manner, the convolution of w(x, y) and f(x, y) denoted by
w(x, y) ∗ f(x, y) is given by:

W(x, y) ∗ f(x, y) = w(s, t)f(x − s, y − t)

Where the minus sign on the right flip (rotate by )

(We can flip and shift either f or w)

4
 Dr. Qadri Hamarsheh

Vector Representation of linear filtering:

 Correlation

= + + ⋯+ = = Z

 R- the response of a mask

 – the coefficients of an m ∗ n filter
 – the corresponding image intensities encompassed by the filter
 Convolution
We simply rotate the mask by
Example: The general 3*3 mask equation:

= + + ⋯+ = = Z

Where:
and are g-dimensional vectors (mask and image)
 Dr. Qadri Hamarsheh
Another representation of 3*3 filter mask

Example of filters masks