Class Notes for CMSC 426, Fall 2005
David Jacobs
Introduction
Correlation and Convolution are basic operations that we will perform to extract
information from images. They are in some sense the simplest operations that we can
perform on an image, but they are extremely useful. Moreover, because they are simple,
they can be analyzed and understood very well, and they are also easy to implement and
can be computed very efficiently. Our main goal is to understand exactly what
correlation and convolution do, and why they are useful. We will also touch on some of
their interesting theoretical properties; though developing a full understanding of them
would take more time than we have.
These operations have two key features: they are shiftinvariant, and they are linear.
Shiftinvariant means that we perform the same operation at every point in the image.
Linear means that this operation is linear, that is, we replace every pixel with a linear
combination of its neighbors. These two properties make these operations very simple;
it’s simpler if we do the same thing everywhere, and linear operations are always the
simplest ones.
We will first consider the easiest versions of these operations, and then generalize. We’ll
make things easier in a couple of ways. First, convolution and correlation are almost
identical operations, but students seem to find convolution more confusing. So we will
begin by only speaking of correlation, and then later describe convolution. Second, we
will start out by discussing 1D images. We can think of a 1D image as just a single row
of pixels. Sometimes things become much more complicated in 2D than 1D, but luckily,
correlation and convolution do not change much with the dimension of the image, so
understanding things in 1D will help a lot. Also, later we will find that in some cases it is
enlightening to think of an image as a continuous function, but we will begin by
considering an image as discrete, meaning as composed of a collection of pixels.
Notation
We will use uppercase letters such as I and J to denote an image. An image may be
either 2D (as it is in real life) or 1D. We will use lowercase letters, like i and j to denote
indices, or positions, in the image. When we index into an image, we will use the same
conventions as Matlab. First, that means that the first element of an image is indicated by
1 (not 0, as in Java, say). So if I is a 1D image, I(1) is its first element. Second, for 2D
images we give first the row, then the column. So I(3,6) is the pixel in the third row of
the image, and the sixth column.
An Example
One of the simplest operations that we can perform with correlation is local averaging.
As we will see, this is also an extremely useful operation. Let’s consider a simple
averaging operation, in which we replace every pixel in a 1D image by the average of
that pixel and its two neighbors. Suppose we have an image I equal to:
Averaging is an operation that takes an image as input, and produces a new image as
output. When we average the fourth pixel, for example, we replace the value 3 with the
average of 2, 3, and 7. That is, if we call the new image that we produce J we can write:
J(4) = (I(3)+I(4)+I(5))/3 = (2+3+7)/3 = 4. Or, for example, we also get: J(3) =
(I(2)+I(3)+I(4))/3 = (4+2+3)/3 = 3. Notice that every pixel in the new image depends
on the pixels in the old image. A possible error is to use J(3) when calculating J(4).
Don’t do this; J(4) should only depend on I(3), I(4) and I(5). Averaging like this is shift
invariant, because we perform the same operation at every pixel. Every new pixel is the
average of itself and its two neighbors. Averaging is linear because every new pixel is a
linear combination of the old pixels. This means that we scale the old pixels (in this
case, we multiply all the neighboring pixels by 1/3) and add them up. This example
illustrates another property of all correlation and convolution that we will consider. The
output image at a pixel is based on only a small neighborhood of pixels around it in the
input image. In this case the neighborhood contains only three pixels. Sometimes we
will use slightly larger neighborhoods, but generally they will not be too big.
Boundaries: We still haven’t fully described correlation, because we haven’t said what
to do at the boundaries of the image. What is J(1)? There is no pixel on its left to include
in the average, ie., I(0) is not defined. There are four common ways of dealing with this
issue.
In the first method of handling boundaries, the original image is padded with zeros
(in red italics).
The first way is to imagine that I is part of an infinitely long image which is zero
everywhere except where we have specified. In that case, we have I(0) = 0, and we can
say: J(1) = (I(0) + I(1) + I(2))/3 = (0 + 5 + 4)/3 = 3. Similarly, we have: J(10) =
(I(9)+I(10)+I(11))/3 = (3 + 6 + 0)/3 = 3.
In the second method of handling boundaries, the original image is padded with the
first and last values (in red italics).
The second way is to also imagine that I is part of an infinite image, but to extend it using
the first and last pixels in the image. In our example, any pixel to the left of the first pixel
in I would have the value 5, and any pixel to the right of the last pixel would have the
value 6. So we would say: J(1) = (I(0) + I(1) + I(2))/3 = (5 + 5 + 4)/3 = 4 2/3, and
J(10) = (I(9)+I(10)+I(11))/3 = (3 + 6 + 6)/3 = 5.
5 4 2 3 7 4 6 5 3 6
. . . 0 0 5 4 2 3 7 4 6 5 3 6 0 0 . . .
. . . 5 5 5 4 2 3 7 4 6 5 3 6 6 6 . . .
In the third method of handling boundaries, the original image is repeated cyclically
(in red italics).
Third, we can imagine the image as being like a circle, so that the pixel values repeat
over and over again. The pixel to the left of the first pixel, then, would be the last pixel
in the image. That is, in our example, we would define I(0) to be I(10). Then we would
have J(1) = (I(0) + I(1) + I(2))/3= (I(10) + I(1) + I(2))/3 = (6 + 5 + 4)/3 = 5, and J(10)
= (I(9)+I(10)+I(11))/3 = (I(9)+I(10)+I(1))/3 = (3 + 6 + 5)/3 = 4 2/3.
Finally, we can simply say that the image is undefined beyond the values that we have
been given. In that case, we cannot compute any average that uses these undefined
values, so J(1) and J(10) will be undefined, and J will be smaller than I.
These four methods have different advantages and disadvantages. If we imagine that the
image we are using is just a small window on the world, and we want to use values
outside the boundary that are most similar to the values that we would have obtained if
we’d taken a bigger picture, than the second approach probably makes the most sense.
That is, if we had to guess at the value of I(0), even though we can’t see it, the value we
can see in I(1) is probably a pretty good guess. In this class, unless we explicitly state
otherwise, you should use the second method for handling boundaries.
Correlation as a Sliding, Windowed Operation
We’re now going to look at the same averaging operation in a slightly different way
which is more graphical, and perhaps more intuitive to generalize. In averaging, for a
specific pixel we multiply it and its neighbors by 1/3 each, and then add up the three
resulting numbers. The numbers we multiply, (1/3, 1/3, 1/3) form a filter. This particular
filter is called a box filter. We can think of it as a 1x3 structure that we slide along the
image. At each position, we multiply each number of the filter by the image number that
lies underneath it, and add these all up. The result is a new number corresponding to the
pixel that is underneath the center of the filter. The figure below shows us producing J(1)
in this way.
* * *
1/3 1/3 1/3

5/3 5/3 4/3
Σ
J
. . . 3 6 5 4 2 3 7 4 6 5 3 6 5 4 . . .
I . . 5 5 4 2 3 7 4 6 5 3 6 6 6 . . .
14/3
To produce the next number in the filtered image, we slide the filter over a pixel, and
perform the same operation.
* * *
1/3 1/3 1/3

5/3 4/3 2/3
Σ
We continue doing this until we have produced every pixel in J. With this view of
correlation, we can define a new averaging procedure by just defining a new filter. For
example, suppose instead of averaging a pixel with its immediate neighbors, we want to
average each pixel with immediate neighbors and their immediate neighbors. We can
define a filter as (1/5, 1/5, 1/5, 1/5, 1/5). Then we perform the same operation as above,
but using a filter that is five pixels wide. The first pixel in the resulting image will then
be: J(1) = (I(1)/5 + I(0)/5 + I(1)/5 + I(2)/5 + I(3)/5) = 1+1+1+4/5 + 2/5 = 4 1/5.
A Mathematical Definition for Correlation
It’s helpful to write this all down more formally. Suppose F is a correlation filter. It will
be convenient notationally to suppose that F has an odd number of elements, so we can
suppose that as it shifts, its center is right on top of an element of I. So we say that F has
2N+1 elements, and that these are indexed from N to N, so that the center element of F is
F(0). Then we can write:
¿
− =
+ =
N
N i
i x I i F x I F ) ( ) ( ) (
where the circle denotes correlation. With this notation, we can define a simple box filter
as:
1 , 0 , 1 for 0
1 , 0 , 1 for
3
1
) (
− ≠
− =
=
i
i
i F
Constructing an Filter from a Continuous Function
It is pretty intuitive what a reasonable averaging filter should look like. Now we want to
start to consider more general strategies for constructing filters. It commonly occurs that
we have in mind a continuous function that would make a good filter, and we want to
come up with a discrete filter that approximates this continuous function. Some reasons
for thinking of filters first as continuous functions will be given when we talk about the
. . . 5 5 4 2 3 7 4 6 5 3 6 6 6 . . .
14/3 11/3
Fourier transform. But in the mean time we’ll give an example of an important
continuous function used for image smoothing, the Gaussian.
A onedimensional Gaussian is:

.

\
 − −
=
2
2
2
) (
exp
2
1
) (
σ
µ
π σ
x
x G
This is also known as a Normal distribution. Here µ is the mean value, and σ is the
variance. Here’s a plot of a Gaussian:
The mean, µ, gives the location of the peak of the function. The parameter σ controls
how wide the peak is. As σ gets smaller the peak becomes narrower.
The Gaussian has a lot of very useful properties, some of which we’ll mention later.
When we use it as a filter, we can think of it as a nice way to average. The averaging
filter that we introduced above replaces each pixel with the average of its neighbors. This
means that nearby pixels all play an equal role in the average, and more distant pixels
play no role. It is more appealing to use the Gaussian to replace each pixel with a
weighted average of its neighbors. In this way, the nearest pixels influence the average
more, and more distant pixels play a smaller and smaller role. This is more elegant,
because we have a smooth and continuous dropoff in the influence of pixels on the
result, instead of a sudden, discontinuous change. We will show this more rigorously
when we discuss the Fourier transform. When we use a Gaussian for smoothing, we will
set µ =0, because we want each pixel to be the one that has the biggest effect on its new,
smoothed value. So our Gaussian has the form:
( )
2
2
0
2
exp
2
1
) (
σ
π σ
x
x G
−
=
σ will serve as a parameter that allows us to control how much we smooth the image
(that is, how big a neighborhood we use for averaging). The bigger σ is, the more we
smooth the image.
We now want to build a discrete filter that looks like a Gaussian. We can do this by just
evaluating the Gaussian at discrete locations. That is, although G is defined for any
continuous value of x, we will just use its values at discrete locations (… 3, 2, 1, 0, 1,
2, 3…). One way to think about this is that we are approximating our original,
continuous function with a piecewise constant function, where the pieces are each one
pixel wide.
However, we cannot evaluate G at every integer location, because this would give us a
filter that is infinitely long, which is impractical. Fortunately, as x gets away from 0, the
value of G(x) will approach 0 pretty quickly, so it will be safe to ignore values of x with a
high absolute value. There is no hard and fast rule about when it is safe to ignore some
values of G(x), but a reasonable rule of thumb is to make sure we capture 99% of the
function. What we mean by this is that the sum of values in our filter will contain at least
99% of what we would get in an infinitely long filter, that is, we choose N so that:
99 .
) (
) (
>
¿
¿
∞
∞ −
−
x G
x G
N
N
There is one final point. An averaging filter should have all its elements add up to 1.
This is what averaging means; if the filter elements don’t add up to one, then we are not
only smoothing the image, we are making it dimmer or brighter. The continuous
Gaussian integrates to one (after all, it’s a probability density function, and probabilities
must add up to one), but when we sample it and truncate it, there is no guarantee that the
values we get will still add up to one. So we must normalize our filter, meaning the
elements of our filter will be calculated as:
N N i for
x G
i G
i F
N
N
...
) (
) (
) ( − = =
¿
−
Taking Derivatives with Correlation
Averaging, or smoothing, is one of the most important things we can do with a filter.
However, another very useful thing is taking a derivative. This allows us to measure how
quickly an image is changing. Taking derivatives is something we usually think of doing
with a continuous function. Technically, derivatives aren’t even defined on a discrete
image, because a derivative measures how quickly the image is changing over an
interval, in the limit, as the interval becomes infinitesimally small. However, just as we
were able to produce a discrete filter that was an approximation to a continuous Gaussian,
we can discretely approximate a continuous derivative operator.
Intuitively, a derivative is found by taking the difference as we go from one part of a
function to another. That is, it’s the change in y divided by the change in x. So it’s
natural that a derivative filter will look something like: (1/2 0 1/2). When we apply this
filter, we get J(i) = (I(i+1)I(i1))/2. This is taking the change in y, that is, the image
intensity, and dividing it by the change in x, the image position.
Let’s consider an example. Suppose image intensity grows quadratically with position,
so that I(x) = x
2
. Then if we look at the intensities at positions 1, 2, 3, … they will look
like:
If we filter this with a filter of the form (1/2 0 1/2) we will get:
We know that the derivative of I(x), dI/dx = 2x. And sure enough, we have, for example,
that J(2) = 4, J(3) = 6, …Notice that J(1) is not equal to 2 because of the way we handle
the boundary, by setting I(0) = 1 instead of I(0) = 0. So at the boundary, our image
doesn’t really reflect I(x) = x
2
.
We could just as easily have used a filter like (0 1 1), which computes the expression:
J(i) = (I(i+1)I(i))/1, or a filter (1 1 0), which computes J(i) = (I(i) – I(i1))/1. These are
all reasonable approximations to a derivative. One advantage of the filter (1/2 0 1/2)
is that it treats the neighbors of a pixel symmetrically.
Matching with Correlation
Part of the power of correlation is that we can use it, and related methods, to find
locations in an image that are similar to a template. To do this, think of the filter as a
template; we are sliding it around the image looking for a location where the template
overlaps the image so that values in the template are aligned with similar values in the
image.
First, we need to decide how to measure the similarity between the template and the
region of the image with which it is aligned. A simple and natural way to do this is to
measure the sum of the square of the differences between values in the template and in
the image. This increases as the difference between the two increases. For the difference
between the filter and the portion of the image centered at x, we can write this as:
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( )
¿ ¿ ¿
¿ ¿
− = − = − =
− = − =
+ − + + =
+ − + + = + −
N
N i
N
N i
N
N i
N
N i
N
N i
i x I i F i x I i F
i x I i F i x I i F i x I i F
2
2 ) ( ) (
2 2
2 2 2
As shown, we can break the Euclidean distance into three parts. The first part depends
only on the filter. This will be the same for every pixel in the image. The second part is
the sum of squares of pixel values that overlap the filter. And the third part is twice the
negative value of the correlation between F and I. We can see that, all things being
equal, as the correlation between the filter and the image increases, the Euclidean
distance between them decreases. This provides an intuition for using correlation to
1 4 9 16 25 36 . . .
1 1/2 4 6 8 10 12 . . .
match a template with an image. Places where the correlation between the two is high
tend to be locations where the filter and image match well.
This also shows one of the weaknesses of using correlation to compare a template and
image. Correlation can also be high in locations where the image intensity is high, even
if it doesn’t match the template well. Here’s an example. Suppose we correlate the filter
3 7 5
with the image:
3 2 4 1 3 8 4 0 3 8 0 7 7 7 1 2
We get the following result:
40 43 39 34 64 85 52 27 61 65 59 84 105 75 38 27
Notice that we get a high result (85) when we center the filter on the sixth pixel, because
(3,7,5) matches very well with (3,8,4). But, the highest correlation occurs at the 13
th
pixel, where the filter is matched to (7,7,7). Even though this is not as similar to (3,7,5),
its magnitude is greater.
One way to overcome this is by just using the sum of square differences between the
signals, as given above. This produces the result:
25 26 26 41 29 2 59 54 34 26 78 13 20 32 61 38
We can see that using Euclidean distance, (3,7,5) best matches the part of the image with
pixel values (3,8,4). The next best match has values (0,7,7), but this is significantly
worse.
Another option is to perform normalized correlation by computing::
( ) ( ) ( )
( ) ( ) ( ) ( )
¿ ¿
¿
− = − =
− =
+
+
N
N i
N
N i
N
N i
i F i x I
i x I i F
2 2
.
This measure of similarity is similar to correlation, but it is invariant to scaling the
template or a corresponding region in the image. If we scale I(xN)…I(x+N), by a single,
constant factor, this will scale the numerator and denominator by the same amount.
Consequently, for example, (3,7,5) will have the same normalized correlation with (7,7,7)
that it would have with (1,1,1). When we perform normalized correlation between (3,7,5)
and the above image we get:
.946 .877 .934 .73 .81 .989 .64 .59 .78 .835 .61 .931 .95 .83 .57 .988
Now, the region of the image that best matches the filter is (3,8,4). However, we also get
a very good match between (3,7,5) and the end of the image, which, when we replicate
the last pixel at the boundary, is (1,2,2). These are actually quite similar, up to a scale
factor. Keep in mind that normalized correlation would say that (3,6,6) and (1,2,2) match
perfectly. Normalized correlation is particularly useful when some effect, such as
changes in lighting or the camera response, might scale the intensities in the image by an
unknown factor.
Correlation as an inner product
One way to get a better intuition for correlation and normalized correlation for matching
is to consider these comparisons as based on an inner product. To do this, we can
consider our filter as a vector, F = (F(N), F(N+1)…F(N)). We are comparing this to a
portion of an image, which we can also think of as a vector: I
x
=(I(xN), I(xN+1),
…I(x+N)). Then, when we compute the correlation between F and I at the position x, we
are computing <F, I
x
>. F and I
x
are vectors in a (2N+1)dimensional space, but still,
everything we’ve learned about inner products applies. In particular,
θ cos ,
x x
I F I F =
where θ is the angle between F and I
x
. This means that the correlation between the two
depends on the magnitude of each vector, and on the angle between them. For F and I
x
of
fixed magnitudes, the correlation between them is maximized when the angle between
them is zero, which occurs when F and I
x
are scaled versions of each other. The
correlation decreases as the angle between them increases. Normalized correlation is
correlation divided by F and I
x
, so it just computes cos θ. So normalized correlation
only computes the angle between F and I
x
. This serves as a good measure of their
similarity if we are not interested in whether their magnitudes are similar.
2D Correlation
Images are 2D, so we really want to perform correlation in 2D. The basic idea is the
same, except that the image and filter are now 2D. We can suppose that our filter is
square and has an odd number of elements, so it is represented by a (2N+1)x(2N+1)
matrix. We don’t lose anything with these assumptions, because we can take any filter
and pad it with zeros to make it square with an odd width. Padding a filter with zeros
does not change its behavior.
Given a square filter, we can compute the results of correlation by aligning the center of
the filter with a pixel. Then we multiply all overlapping values together, and add up the
result. We can write this as:
¿ ¿
− = − =
+ + =
N
N j
N
N i
j y i x I j i F y x I F ) , ( ) , ( ) , (
Example: We can perform averaging of a 2D image using a 2D box filter, which, for a
3x3 filter, will look like:
1/9 1/9 1/9
1/9 1/9 1/9
1/9 1/9 1/9
F
Below we show an image, I, and the result of applying the box filter above to I to produce
the resulting image, J.
8 3 4 5
7 6 4 5
4 5 7 8
6 5 5 6
I
I with padded boundaries J = FoI
Separable filters: Generally, 2D correlation is more expensive than 1D correlation
because the sizes of the filters we use are larger. For example, if our filter is
(N+1)x(N+1) in size, and our image contains MxM pixels, then the total number of
multiplications we must perform is (N+1)
2
M
2
. However, with an important class of
filters, we can save on this computation. These are separable filters, which can be written
as a combination of two smaller filters. For example, in the case of the box filter, we
have:
(
(
(
¸
(
¸
(
(
(
(
(
(
¸
(
¸
=
(
(
(
(
(
(
¸
(
¸
0 0 0
3
1
3
1
3
1
0 0 0
0
3
1
0
0
3
1
0
0
3
1
0
9
1
9
1
9
1
9
1
9
1
9
1
9
1
9
1
9
1
That is, the box filter is equal to the correlation of two filters that have a row and a
column of equal values, and all other values equal to zero (here we are handling
boundaries by using 0 values outside the boundary). When this is the case, correlating an
image with the box filter produces the same result as correlating it separately with the
two filters on the right side of the equals sign. Intuitively, this makes sense. Correlating
with the rightmost filter averages each pixel with its neighbors on the same row. The
second correlation then averages these resulting pixels with those neighbors in the same
8 8 3 4 5 5
8 8 3 4 5 5
7 7 6 4 5 5
4 4 5 7 8 8
6 6 5 5 6 6
6 6 5 5 6 6
6.44 5.22 4.33 4.67
5.78 5.33 5.22 5.67
5.56 5.44 5.67 6
5.22 5.33 5.78 6.33
column, which are already averages of neighbors on the same row. This produces an
average of all neighboring pixels.
The advantage of separating a filter in this way is that when we perform correlation with
the matrices on the right, we don’t need to actually perform any multiplications where
there are zeros. In fact, it is more convenient to simply write these filters as 3x1 and 1x3,
ignoring the zeros. This means that for each pixel in the image, we only need to perform
6 multiplications instead of 9. In general, if we can separate a (N+1)x(N+1) filter into
two filters that are (N+1)x1 and 1x(N+1), the total work we must do in correlation is
reduced from (N+1)
2
M
2
to 2(N+1)M
2
.
Gaussian filtering has the same separability. For a 2D Gaussian, we use a function that is
rotationally symmetric, where the height of the filter falls off as a Gaussian function of
the distance from the center. This has the form:
( )

.

\
 + −
=
2
2 2
0
2
exp
2
1
) , (
σ
π σ
y x
y x G
This can be separated because it can be expressed as:
( )
( )

.

\
 −
−
=

.

\
 + −
=
2
2
2
2
2
2 2
0
2
exp
2
exp
2
1
2
exp
2
1
) , (
σ σ
π σ
σ
π σ
y
x
y x
y x G
It is the product of two components, one of which only depends on x, while the other only
depends on y. The first component can be expressed with a filter that has a single row,
and the second can be expressed with a filter that has a single column.
Convolution
Convolution is just like correlation, except that we flip over the filter before correlating.
For example, convolution of a 1D image with the filter (3,7,5) is exactly the same as
correlation with the filter (5,7,3). We can write the formula for this as:
¿
− =
− = ∗
N
N i
i x I i F x I F ) ( ) ( ) (
In the case of 2D convolution we flip the filter both horizontally and vertically. This can
be written as:
¿ ¿
− = − =
− − = ∗
N
N j
N
N i
j y i x I j i F y x I F ) , ( ) , ( ) , (
Notice that correlation and convolution are identical when the filter is symmetric.
The key difference between the two is that convolution is associative. That is, if F and G
are filters, then F*(G*I) = (F*G)*I. If you don’t believe this, try a simple example, using
F=G=(1 0 1), for example. It is very convenient to have convolution be associative.
Suppose, for example, we want to smooth an image and then take its derivative. We
could do this by convolving the image with a Gaussian filter, and then convolving it with
a derivative filter. But we could alternatively convolve the derivative filter with the
Gaussian to produce a filter called a Difference of Gaussian (DOG), and then convolve
this with our image. The nice thing about this is that the DOG filter can be precomputed,
and we only have to convolve one filter with our image.
In general, people use convolution for image processing operations such as smoothing,
and they use correlation to match a template to an image. Then, we don’t mind that
correlation isn’t associative, because it doesn’t really make sense to combine two
templates into one with correlation, whereas we might often want to combine two filter
together for convolution. When discussing the Fourier series as a way of understanding
filtering, we will use convolution, so that we can introduce the famous convolution
theorem.
The Fourier Series (this section is a bit more advanced)
The Fourier Series gives us a very important and useful way of representing an image.
While it is a very powerful representation for many reasons, we will mainly introduce it
as a way of understanding the effect of convolution. First, we will consider what
properties a good image representation should have by looking at familiar
representations. Then we will describe the Fourier Series, which has these properties.
Finally, we will describe the convolution theorem, which helps us better understand the
effect of convolution with different filters.
Orthonormal Basis for Vectors
How do we represent points in 2D? Using their x and y coordinates. What this really
means is that we write any point as a linear combination of two vectors (1,0) and (0,1).
(x,y) = x(1,0) + y(0,1).
These vectors form an orthonormal basis for the plane. That means they are orthogonal
(ie, perpendicular), <(1,0), (0,1)> = 0, and of unit magnitude. Why is an orthonormal
basis a good representation? There are many reasons, but two are:
1) Projection. To find the x coordinate, we take <(x,y), (1,0)>. In general, when we
have an orthonormal basis for a space, we can describe any point with
coordinates, and find each coordinate of a point by taking an inner product with
the basis vector for this coordinate.
2) Pythagorean theorem. (x,y)
2
= x
2
+ y
2
. When we have an orthonormal basis,
the squared length of a vector is the sum of the square of its coordinates in that
basis.
As an example of a less attractive basis, supposed we represented points using the two
basis vectors (1,0) and (1,1). These are not orthogonal, because <(1,0),(1,1)> = 1. Still,
given any point, (x,y), we could always represent it as:
(x,y) = u(1,0) + v(1,1)
for some choice of u and v. We can compute u and v by taking v = y, u = x – y. For
example, we would represent the point (7,3) as 4(1,0) + 3(1,1). So rather than represent
the point with the Euclidean coordinates (7,3), we can represent this same point in this
new basis with the coordinates (4,3). But note that the Pythogorean theorem wouldn’t
hold. And, if I give you two basis vectors that are not orthonormal, it is not so easy to
come up with a formula for determining the coordinates of a point using these basis
vectors (especially as we look at higher dimensions).
The Fourier Series
The Fourier series is a good representation because it provides an orthonormal basis for
images. We want to explore it in the simplest possible setting, to get the key intuitions.
This is one case in which it is easier to get the idea by working with continuous functions,
rather than discrete ones. So, instead of thinking of an image as a list of pixels, we’ll
think of it as a continuous function, I(x). To simplify, we’ll also only do things in 1D,
but all this extends to 2D images. We will also consider only periodic functions that go
from 0 to 2π, and then repeat over and over again. This is equivalent to the third way that
we describe above for handling pixels outside of the boundary of images. The main thing
I’ll do to make things easier, though, is to be very sloppy mathematically, skipping
precise conditions and giving intuitions instead of complete proofs.
The following functions provide an orthonormal basis for functions:
,... 3 , 2 , 1
) sin(
,
) cos(
,
2
1
= k for
kx kx
π π π
These functions form the Fourier Series. The first one is a constant function, and the
others are just sines and cosines of increasing frequency.
To define things like whether a set of functions are orthonormal, we need to define an
inner product between two functions. We do this with a continuous version of the inner
product. With discrete vectors, to compute the inner product we multiply together the
values of matching coordinates, and then add up the results. We do the same thing with
continuous functions; we multiply them together, and integrate (add) the result. So the
inner product between two functions, f(x) and g(x), defined in the domain from 0 to 2π,
is:
}
=
π 2
0
) ( ) ( , dx x g x f g f
When we say that a function, f, has unit magnitude, we mean that <f,f> = 1. When we
say that two functions are orthogonal, we mean that <f,g> = 0. One reason that these are
sensible definitions is that if we approximate a function by sampling its values, and
building them into a vector, then the definitions in the continuous case give us the same
result as the limit of what we get in the discrete case, as we sample the function more and
more densely.
We can use this definition to show that the functions in the Fourier Series have unit
magnitude, by taking the inner product of each function with itself, and showing that this
is equal to 1. We can show they are orthogonal by showing that their inner products with
each other are zero. Doing this requires solving some integrals of trigonometric
functions.
To show that the Fourier Series forms a good coordinate system for all functions we must
also show that we can express any function as a linear combination of the elements of the
Fourier Series. That is, we must show that for any function, f, we can write:
¿
∞
=


.

\

+ + =
1
0
) sin( ) cos(
2
1
) (
k
k k
kx
a
kx
b a x f
π π π
Then the values (a
0
, b
1
, a
1
, b
2
, a
2
, …) are the coordinates of the function in this new
coordinate system provided by the Fourier Series. This is a very important fact, but we
will skip the proof here. Keep in mind that I’m leaving out some necessary conditions,
such as that this only works when f is a continuous function.
This result means that any function can be broken down into the sum of sine waves of
different amplitudes and phases. We say that the part of the function that is due to longer
frequency sine waves is the low frequency part of the function. The part that is due to
high frequency sine waves is the high frequency component of the function. If a function
is very bumpy, with small rapid changes in its value, these rapid changes will be due to
the high frequency component of the function. If we can eliminate these highfrequency
components, we can make the function smoother.
Implications of the orthonormality of the Fourier Series.
As with an orthonormal basis for vectors, the orthonormality of the Fourier Series means
that we can use projection and (a generalization of) the Pythagorean theorem.
We can solve the for the values (a
0
, b
1
, a
1
, b
2
, a
2
, …) just by taking inner products. So:
}
} }
= =
= = = =
π
π π
π π
π π π π
2
0
2
0
2
0
0
) sin( ) ( ) sin(
,
) cos( ) ( ) cos(
,
2
) (
2
1
,
dx
kx x f kx
f a
dx
kx x f kx
f b dx
x f
f a
k
k
The analog to the Pythagorean theorem is called Parsevaal’s theorem. This is:
( )
¿
}
∞
+ + =
1
2 2 2
0
2
0
2
) (
i i
b a a dx x f
π
Convolution Theorem
Now we get to the reason why the Fourier Series is important in understanding
convolution. This is because of the convolution theorem. Let F, G, H be the fourier
series that represents the functions f, g, and h. That is, F,G, and H are infinite vectors.
The convolution theorem states that convolution in the spatial domain is equivalent to
componentwise multiplication in the transform domain, ie, f*g = h FG = H. Here,
by FG, we mean that we multiply each element of F by the corresponding element of G
and use this as an element of H.
This is a remarkable theorem. As an example of its significance, suppose we convolve a
filter, F, with an image I, and that F just consists of a single sine wave, F=sin(x). Then
the Fourier Series representation of F is:
0. components other all ,
1
= = π a
This means that no matter what I is, if we define J=F*I, then the components of the
Fourier series of J will all be zero except for a
1
, because all other components of F are 0,
and when we multiply these by the corresponding components of I we always get 0.
More generally, this means that convolution with a filter attenuates every frequency by
different amounts, and we can understand the effects of convolution by finding the
Fourier Series of the convolution kernel.
As another important example, we point out that the Fourier series representation of a
Gaussian is also a Gaussian. If g(t) is a Gaussian, the broader g is the narrower G is (and
vice versa). This means that smoothing with a Gaussian reduces high frequency
components of a signal. The more we smooth, the more we reduce the high frequency
components. In fact, attenuating high frequency components of a signal can be taken to
be the definition of smoothing. To some extent, this is why we learned the Fourier series.
It allows us to understand what smoothing really does.
We can also see why simple averaging isn’t as good a way of smoothing. Let f(x) be an
averaging filter, which is constant in the range –T to T. Then, we can compute its Fourier
Series as:
}
−
= =
T
T
k
k
kT
dx
kx
a
π π
) cos( 2 ) sin(
}
−
−
= =
T
T
k
k
kT
dx
kx
b
π π
) sin( 2 ) cos(
This is called the sinc function. If we graph it, we see that it looks like:
This does not decay as fast as a Gaussian. Also it has oscillations that produce odd
effects. As a result of these oscillations, some high frequency components of an image
will be wiped out by averaging, while some similar high frequency components remain.
Aliasing and the Sampling Theorem
Sampling means that we know the value of a function, f, at discrete intervals, f(nπ/T) for
n = 0, +1, +T. That is, we have 2T + 1 uniform samples of f(x), assuming f is a
periodic function. Suppose now that f is bandlimited to T. That means,
¿
=


.

\

+ + =
T
k
k k
kx
a
kx
b a x f
1
0
) sin( ) cos(
2
1
) (
π π π
}
−
= =
T
T
T
dx a
π π 2
2
2
1
0
We know the value of f(x) at 2T + 1 positions. If we plug these values into the above
equation, we get 2T + 1 equations. The a
i
and b
i
give us 2T + 1 unknowns. These are
linear equations. So we get a unique solution. This means that we can reconstruct f from
its samples. However, if we have fewer samples, reconstruction is not possible. If there
is a higher frequency component, then this can play havoc with our reconstruction of all
low frequency components, and be interpreted as odd low frequency components.
This means that if we want to sample a signal, the best thing we can do is to lowpass
filter it and then sample it. Then, we get a signal that matches our original signal
perfectly in the lowfrequency components, and this is the best we can do. To shrink
images, in fact, we lowpass filter them and then sample them.
we multiply all the neighboring pixels by 1/3) and add them up. In this case the neighborhood contains only three pixels. and J(10) = (I(9)+I(10)+I(11))/3 = (3 + 6 + 6)/3 = 5. This example illustrates another property of all correlation and convolution that we will consider. we have: J(10) = (I(9)+I(10)+I(11))/3 = (3 + 6 + 0)/3 = 3. Or. . 5 5 5 4 2 3 7 4 6 5 3 6 6 6 . This means that we scale the old pixels (in this case. I(0) is not defined. but to extend it using the first and last pixels in the image. In the second method of handling boundaries. . That is. and produces a new image as output. Similarly. we replace the value 3 with the average of 2. Boundaries: We still haven’t fully described correlation. Averaging like this is shiftinvariant. we also get: J(3) = (I(2)+I(3)+I(4))/3 = (4+2+3)/3 = 3. Averaging is linear because every new pixel is a linear combination of the old pixels. There are four common ways of dealing with this issue. . I(4) and I(5). . and 7. In the first method of handling boundaries. but generally they will not be too big. and we can say: J(1) = (I(0) + I(1) + I(2))/3 = (0 + 5 + 4)/3 = 3. A possible error is to use J(3) when calculating J(4). because we haven’t said what to do at the boundaries of the image. we have I(0) = 0. 0 0 5 4 2 3 7 4 6 5 3 6 0 0 . . if we call the new image that we produce J we can write: J(4) = (I(3)+I(4)+I(5))/3 = (2+3+7)/3 = 4. Don’t do this. The second way is to also imagine that I is part of an infinite image. The first way is to imagine that I is part of an infinitely long image which is zero everywhere except where we have specified. So we would say: J(1) = (I(0) + I(1) + I(2))/3 = (5 + 5 + 4)/3 = 4 2/3. Sometimes we will use slightly larger neighborhoods. Let’s consider a simple averaging operation. . because we perform the same operation at every pixel. . ie. J(4) should only depend on I(3). Suppose we have an image I equal to: 5 4 2 3 7 4 6 5 3 6 Averaging is an operation that takes an image as input. In our example. the original image is padded with the first and last values (in red italics). When we average the fourth pixel. In that case. What is J(1)? There is no pixel on its left to include in the average. . . this is also an extremely useful operation. the original image is padded with zeros (in red italics). and any pixel to the right of the last pixel would have the value 6. . any pixel to the left of the first pixel in I would have the value 5. Notice that every pixel in the new image depends on the pixels in the old image. for example. 3. in which we replace every pixel in a 1D image by the average of that pixel and its two neighbors.. As we will see. .One of the simplest operations that we can perform with correlation is local averaging. The output image at a pixel is based on only a small neighborhood of pixels around it in the input image. Every new pixel is the average of itself and its two neighbors. for example.
if we had to guess at the value of I(0). we can imagine the image as being like a circle. 5 * 1/3 5 * 1/3 4 * 1/3 2 3 7 4 6 5 3 6 6 6 . If we imagine that the image we are using is just a small window on the world. . then. and J(10) = (I(9)+I(10)+I(11))/3 = (I(9)+I(10)+I(1))/3 = (3 + 6 + 5)/3 = 4 2/3. 1/3) form a filter. I . . and we want to use values outside the boundary that are most similar to the values that we would have obtained if we’d taken a bigger picture. you should use the second method for handling boundaries. These four methods have different advantages and disadvantages. In averaging. . the value we can see in I(1) is probably a pretty good guess. In that case. the original image is repeated cyclically (in red italics). At each position. In the third method of handling boundaries. would be the last pixel in the image. and add these all up. and perhaps more intuitive to generalize. and then add up the three resulting numbers. so J(1) and J(10) will be undefined. In this class. unless we explicitly state otherwise. . in our example. 1/3. we multiply each number of the filter by the image number that lies underneath it. The numbers we multiply. Then we would have J(1) = (I(0) + I(1) + I(2))/3= (I(10) + I(1) + I(2))/3 = (6 + 5 + 4)/3 = 5. Windowed Operation We’re now going to look at the same averaging operation in a slightly different way which is more graphical. The pixel to the left of the first pixel. That is.  5/3 5/3 4/3 Σ J 14/3 . for a specific pixel we multiply it and its neighbors by 1/3 each. 3 6 5 4 2 3 7 4 6 5 3 6 5 4 . and J will be smaller than I. Correlation as a Sliding. We can think of it as a 1x3 structure that we slide along the image. . we can simply say that the image is undefined beyond the values that we have been given. Finally. we would define I(0) to be I(10). Third. even though we can’t see it. This particular filter is called a box filter. . so that the pixel values repeat over and over again. That is. . (1/3. we cannot compute any average that uses these undefined values. The result is a new number corresponding to the pixel that is underneath the center of the filter. than the second approach probably makes the most sense. The figure below shows us producing J(1) in this way..
. so that the center element of F is F(0). So we say that F has 2N+1 elements. With this view of correlation. Suppose F is a correlation filter. we slide the filter over a pixel. Now we want to start to consider more general strategies for constructing filters. and that these are indexed from N to N. With this notation.  5/3 4/3 2/3 Σ 14/3 11/3 We continue doing this until we have produced every pixel in J. For example. we want to average each pixel with immediate neighbors and their immediate neighbors. 1/5. its center is right on top of an element of I. but using a filter that is five pixels wide. . we can define a simple box filter as: 1 for i = −1. 5 5 * 1/3 4 * 1/3 2 * 1/3 3 7 4 6 5 3 6 6 6 . Then we perform the same operation as above. and we want to come up with a discrete filter that approximates this continuous function. . 1/5). We can define a filter as (1/5.To produce the next number in the filtered image. and perform the same operation. 1/5. suppose instead of averaging a pixel with its immediate neighbors. Then we can write: F I ( x) = N i =− N F (i ) I ( x + i ) where the circle denotes correlation. .0. we can define a new averaging procedure by just defining a new filter. It will be convenient notationally to suppose that F has an odd number of elements. . Some reasons for thinking of filters first as continuous functions will be given when we talk about the . 1/5.1 F (i ) = 3 0 for i ≠ −1. The first pixel in the resulting image will then be: J(1) = (I(1)/5 + I(0)/5 + I(1)/5 + I(2)/5 + I(3)/5) = 1+1+1+4/5 + 2/5 = 4 1/5. A Mathematical Definition for Correlation It’s helpful to write this all down more formally.0. It commonly occurs that we have in mind a continuous function that would make a good filter.1 Constructing an Filter from a Continuous Function It is pretty intuitive what a reasonable averaging filter should look like. so we can suppose that as it shifts.
A onedimensional Gaussian is: G ( x) = 1 2 exp − ( x − µ ) σ 2π 2σ 2 This is also known as a Normal distribution. we will just use its values at discrete locations (… 3.Fourier transform. because we want each pixel to be the one that has the biggest effect on its new. 1. some of which we’ll mention later. the more we smooth the image. Here’s a plot of a Gaussian: The mean. how big a neighborhood we use for averaging). We now want to build a discrete filter that looks like a Gaussian. smoothed value. instead of a sudden. and more distant pixels play no role. 2. When we use a Gaussian for smoothing. although G is defined for any continuous value of x. In this way. As σ gets smaller the peak becomes narrower. 0. This is more elegant. we can think of it as a nice way to average. Here µ is the mean value. It is more appealing to use the Gaussian to replace each pixel with a weighted average of its neighbors. The bigger σ is. The Gaussian has a lot of very useful properties. 2. the nearest pixels influence the average more. 1. But in the mean time we’ll give an example of an important continuous function used for image smoothing. gives the location of the peak of the function. the Gaussian. The parameter σ controls how wide the peak is. and σ is the variance. we will set µ =0. . We will show this more rigorously when we discuss the Fourier transform. The averaging filter that we introduced above replaces each pixel with the average of its neighbors. because we have a smooth and continuous dropoff in the influence of pixels on the result. µ. We can do this by just evaluating the Gaussian at discrete locations. That is. This means that nearby pixels all play an equal role in the average. So our Gaussian has the form: 2 1 G0 ( x ) = exp − x 2σ 2 σ 2π ( ) σ will serve as a parameter that allows us to control how much we smooth the image (that is. and more distant pixels play a smaller and smaller role. When we use it as a filter. 3…). One way to think about this is that we are approximating our original. discontinuous change.
it’s the change in y divided by the change in x. where the pieces are each one pixel wide. So we must normalize our filter. as x gets away from 0. Fortunately. the image intensity. so it will be safe to ignore values of x with a high absolute value.99 G ( x) −∞ There is one final point. There is no hard and fast rule about when it is safe to ignore some values of G(x). if the filter elements don’t add up to one. but when we sample it and truncate it. we can discretely approximate a continuous derivative operator. the value of G(x) will approach 0 pretty quickly. When we apply this filter.continuous function with a piecewise constant function.. Intuitively. Technically. derivatives aren’t even defined on a discrete image. This is taking the change in y. the image position. What we mean by this is that the sum of values in our filter will contain at least 99% of what we would get in an infinitely long filter. Taking derivatives is something we usually think of doing with a continuous function. just as we were able to produce a discrete filter that was an approximation to a continuous Gaussian. that is. and probabilities must add up to one). we are making it dimmer or brighter. So it’s natural that a derivative filter will look something like: (1/2 0 1/2). meaning the elements of our filter will be calculated as: G (i ) F (i ) = N for i = − N . However. However. This is what averaging means. in the limit. we cannot evaluate G at every integer location. but a reasonable rule of thumb is to make sure we capture 99% of the function. then we are not only smoothing the image. This allows us to measure how quickly an image is changing. However. An averaging filter should have all its elements add up to 1. we choose N so that: N G ( x) −N ∞ > . a derivative is found by taking the difference as we go from one part of a function to another. . is one of the most important things we can do with a filter. there is no guarantee that the values we get will still add up to one. because this would give us a filter that is infinitely long.. That is.N G ( x) −N Taking Derivatives with Correlation Averaging. which is impractical. because a derivative measures how quickly the image is changing over an interval. we get J(i) = (I(i+1)I(i1))/2. as the interval becomes infinitesimally small. that is. it’s a probability density function. and dividing it by the change in x. The continuous Gaussian integrates to one (after all. another very useful thing is taking a derivative. or smoothing.
First. to find locations in an image that are similar to a template. The first part depends only on the filter. If we filter this with a filter of the form (1/2 0 1/2) we will get: 1 1/2 4 6 8 10 12 . dI/dx = 2x. Suppose image intensity grows quadratically with position. Then if we look at the intensities at positions 1. We could just as easily have used a filter like (0 1 1). . This will be the same for every pixel in the image. our image doesn’t really reflect I(x) = x2. by setting I(0) = 1 instead of I(0) = 0. One advantage of the filter (1/2 0 1/2) is that it treats the neighbors of a pixel symmetrically. This provides an intuition for using correlation to . the Euclidean distance between them decreases. and related methods. And sure enough. we need to decide how to measure the similarity between the template and the region of the image with which it is aligned. …Notice that J(1) is not equal to 2 because of the way we handle the boundary. For the difference between the filter and the portion of the image centered at x. all things being equal. which computes the expression: J(i) = (I(i+1)I(i))/1. J(3) = 6. A simple and natural way to do this is to measure the sum of the square of the differences between values in the template and in the image. that J(2) = 4. . we can write this as: N i =− N (F (i) − I ( x + i))2 = = N i=− N N (F (i ) + I (x + i ) − 2 F (i )I (x + i )) 2 2 i=− N (F (i )) + (I (x + i )) − 2 (F (i )I (x + i )) N N 2 2 i =− N i =− N As shown. We can see that. Matching with Correlation Part of the power of correlation is that we can use it. 3. we have. we are sliding it around the image looking for a location where the template overlaps the image so that values in the template are aligned with similar values in the image. . We know that the derivative of I(x). And the third part is twice the negative value of the correlation between F and I. So at the boundary. for example. we can break the Euclidean distance into three parts. think of the filter as a template. so that I(x) = x2. 2.Let’s consider an example. … they will look like: . as the correlation between the filter and the image increases. which computes J(i) = (I(i) – I(i1))/1. The second part is the sum of squares of pixel values that overlap the filter. To do this. This increases as the difference between the two increases. or a filter (1 1 0). These are all reasonable approximations to a derivative. 1 4 9 16 25 36 .
for example.7. If we scale I(xN)…I(x+N).989 . as given above. even if it doesn’t match the template well.match a template with an image. its magnitude is greater.59 .8.5) matches very well with (3.83 .7).7.931 . where the filter is matched to (7. this will scale the numerator and denominator by the same amount.57 . (3. by a single. Places where the correlation between the two is high tend to be locations where the filter and image match well.7.5) and the above image we get: .946 . but it is invariant to scaling the template or a corresponding region in the image. constant factor.5) will have the same normalized correlation with (7.78 . Even though this is not as similar to (3. but this is significantly worse. This produces the result: 25 26 26 41 29 2 59 54 34 26 78 13 20 32 61 38 We can see that using Euclidean distance. Consequently. This also shows one of the weaknesses of using correlation to compare a template and image. But. the highest correlation occurs at the 13th pixel.64 .934 .4).5).73 . One way to overcome this is by just using the sum of square differences between the signals. (3. Correlation can also be high in locations where the image intensity is high.61 .7) that it would have with (1.7.7. The next best match has values (0. Here’s an example.7. Suppose we correlate the filter 3 7 5 with the image: 3 2 4 1 3 8 4 0 3 8 0 7 7 7 1 2 We get the following result: 40 43 39 34 64 85 52 27 61 65 59 84 105 75 38 27 Notice that we get a high result (85) when we center the filter on the sixth pixel.5) best matches the part of the image with pixel values (3.81 .835 .95 .4). (I (x + i ))2 (F (i ))2 This measure of similarity is similar to correlation.988 .1).8.7.1. Another option is to perform normalized correlation by computing:: N i =− N N i =− N (F (i )I (x + i )) N i =− N .877 .7). When we perform normalized correlation between (3.7. because (3.
and on the angle between them. is (1. So normalized correlation only computes the angle between F and Ix.5) and the end of the image. Given a square filter.6) and (1. This means that the correlation between the two depends on the magnitude of each vector. when we compute the correlation between F and I at the position x. Ix >.2). the region of the image that best matches the filter is (3. the correlation between them is maximized when the angle between them is zero. everything we’ve learned about inner products applies. y + j ) j =− N i =− N .Now. I x = F I x cos θ where θ is the angle between F and Ix. Correlation as an inner product One way to get a better intuition for correlation and normalized correlation for matching is to consider these comparisons as based on an inner product. These are actually quite similar. we also get a very good match between (3. such as changes in lighting or the camera response. might scale the intensities in the image by an unknown factor. Padding a filter with zeros does not change its behavior. We are comparing this to a portion of an image. except that the image and filter are now 2D. F = (F(N). which. when we replicate the last pixel at the boundary. Keep in mind that normalized correlation would say that (3. However. up to a scale factor. Then. y ) = N N F (i. and add up the result. Normalized correlation is particularly useful when some effect. Then we multiply all overlapping values together. so it is represented by a (2N+1)x(2N+1) matrix.8. F(N+1)…F(N)). To do this.2. I(xN+1).2. which occurs when F and Ix are scaled versions of each other.2) match perfectly. we can consider our filter as a vector. so we really want to perform correlation in 2D. …I(x+N)). but still. j ) I ( x + i.7. 2D Correlation Images are 2D. We can suppose that our filter is square and has an odd number of elements. F and Ix are vectors in a (2N+1)dimensional space. so it just computes cos θ. F . We can write this as: F I ( x. We don’t lose anything with these assumptions. In particular. For F and Ix of fixed magnitudes. which we can also think of as a vector: Ix =(I(xN). The basic idea is the same.6.4). we can compute the results of correlation by aligning the center of the filter with a pixel. The correlation decreases as the angle between them increases. This serves as a good measure of their similarity if we are not interested in whether their magnitudes are similar. we are computing <F. Normalized correlation is correlation divided by F and Ix. because we can take any filter and pad it with zeros to make it square with an odd width.
Intuitively.67 5.22 5.78 4. When this is the case. if our filter is (N+1)x(N+1) in size. and all other values equal to zero (here we are handling boundaries by using 0 values outside the boundary). with an important class of filters. These are separable filters. in the case of the box filter. For example. for a 3x3 filter. The second correlation then averages these resulting pixels with those neighbors in the same . J. 8 7 4 6 3 6 5 5 I 4 4 7 5 5 5 8 6 8 8 7 4 6 6 8 8 7 4 6 6 3 3 6 5 5 5 4 4 4 7 5 5 5 5 5 8 6 6 5 5 5 8 6 6 6.44 5.Example: We can perform averaging of a 2D image using a 2D box filter.44 5.22 5.67 5. the box filter is equal to the correlation of two filters that have a row and a column of equal values. I.67 6 6. we have: 1 9 1 9 1 9 1 9 1 9 1 9 1 1 0 0 9 3 1 1 = 0 0 9 3 1 1 0 0 9 3 0 1 3 0 0 1 3 0 0 1 3 0 That is.22 5. we can save on this computation. which. then the total number of multiplications we must perform is (N+1)2M2. correlating an image with the box filter produces the same result as correlating it separately with the two filters on the right side of the equals sign.33 4. However.56 5.33 5. this makes sense.33 5. Correlating with the rightmost filter averages each pixel with its neighbors on the same row. 2D correlation is more expensive than 1D correlation because the sizes of the filters we use are larger. For example. will look like: 1/9 1/9 1/9 1/9 1/9 1/9 F 1/9 1/9 1/9 Below we show an image. and the result of applying the box filter above to I to produce the resulting image. and our image contains MxM pixels. which can be written as a combination of two smaller filters.78 5.33 I with padded boundaries J = FoI Separable filters: Generally.
and the second can be expressed with a filter that has a single column.7. where the height of the filter falls off as a Gaussian function of the distance from the center. y ) = exp − (x + y ) 2 = exp − x 2 exp 2σ 2σ 2σ 2 σ 2π σ 2π It is the product of two components. convolution of a 1D image with the filter (3. This produces an average of all neighboring pixels. we only need to perform 6 multiplications instead of 9. . j ) I ( x − i.7. y ) = 2σ σ 2π This can be separated because it can be expressed as: 2 2 2 1 1 − y2 G 0 ( x. if we can separate a (N+1)x(N+1) filter into two filters that are (N+1)x1 and 1x(N+1). except that we flip over the filter before correlating. In general. y ) = N N F (i. ( ) Convolution Convolution is just like correlation. For example. we don’t need to actually perform any multiplications where there are zeros. For a 2D Gaussian. while the other only depends on y.column. the total work we must do in correlation is reduced from (N+1)2M2 to 2(N+1)M2. we use a function that is rotationally symmetric. The advantage of separating a filter in this way is that when we perform correlation with the matrices on the right. This means that for each pixel in the image. In fact. it is more convenient to simply write these filters as 3x1 and 1x3.5) is exactly the same as correlation with the filter (5. This has the form: 2 2 1 exp − (x + y ) 2 G 0 ( x. We can write the formula for this as: F ∗ I ( x) = N i=− N F (i ) I ( x − i ) In the case of 2D convolution we flip the filter both horizontally and vertically. y − j ) j =− N i =− N Notice that correlation and convolution are identical when the filter is symmetric. This can be written as: F ∗ I ( x. ignoring the zeros. one of which only depends on x. The first component can be expressed with a filter that has a single row. which are already averages of neighbors on the same row. Gaussian filtering has the same separability.3).
for example. we don’t mind that correlation isn’t associative.0)>. We could do this by convolving the image with a Gaussian filter. try a simple example. What this really means is that we write any point as a linear combination of two vectors (1. Then we will describe the Fourier Series. Suppose. we want to smooth an image and then take its derivative. because it doesn’t really make sense to combine two templates into one with correlation. we can describe any point with coordinates. To find the x coordinate. First. Then. people use convolution for image processing operations such as smoothing. and of unit magnitude.1). <(1.0) and (0. Finally. we will consider what properties a good image representation should have by looking at familiar representations. Why is an orthonormal basis a good representation? There are many reasons. The nice thing about this is that the DOG filter can be precomputed.y). .y) = x(1. and find each coordinate of a point by taking an inner product with the basis vector for this coordinate. (1. and we only have to convolve one filter with our image. if F and G are filters. (0. whereas we might often want to combine two filter together for convolution. we will mainly introduce it as a way of understanding the effect of convolution.y)2 = x2 + y2. Orthonormal Basis for Vectors How do we represent points in 2D? Using their x and y coordinates. we take <(x.0). That means they are orthogonal (ie. perpendicular). When we have an orthonormal basis. which helps us better understand the effect of convolution with different filters.1)> = 0. when we have an orthonormal basis for a space. but two are: 1) Projection. the squared length of a vector is the sum of the square of its coordinates in that basis. But we could alternatively convolve the derivative filter with the Gaussian to produce a filter called a Difference of Gaussian (DOG). which has these properties. we will describe the convolution theorem. 2) Pythagorean theorem. That is. and then convolving it with a derivative filter. When discussing the Fourier series as a way of understanding filtering. (x. using F=G=(1 0 1). then F*(G*I) = (F*G)*I. The Fourier Series (this section is a bit more advanced) The Fourier Series gives us a very important and useful way of representing an image. It is very convenient to have convolution be associative. for example. and then convolve this with our image.The key difference between the two is that convolution is associative. If you don’t believe this.1). we will use convolution. These vectors form an orthonormal basis for the plane. In general. so that we can introduce the famous convolution theorem. and they use correlation to match a template to an image. While it is a very powerful representation for many reasons.0) + y(0. In general. (x.
skipping precise conditions and giving intuitions instead of complete proofs. I(x). we multiply them together.3) as 4(1. To simplify. we could always represent it as: (x. π π These functions form the Fourier Series.. we need to define an inner product between two functions. to get the key intuitions. But note that the Pythogorean theorem wouldn’t hold. we’ll also only do things in 1D. cos(kx) sin(kx) .. This is one case in which it is easier to get the idea by working with continuous functions. So. it is not so easy to come up with a formula for determining the coordinates of a point using these basis vectors (especially as we look at higher dimensions). So the inner product between two functions. The Fourier Series The Fourier series is a good representation because it provides an orthonormal basis for images. These are not orthogonal. To define things like whether a set of functions are orthonormal. is to be very sloppy mathematically. With discrete vectors. we’ll think of it as a continuous function.y).1)> = 1. instead of thinking of an image as a list of pixels. The first one is a constant function.0) + 3(1. because <(1.1). We want to explore it in the simplest possible setting. u = x – y. we would represent the point (7. The following functions provide an orthonormal basis for functions: 1 2π . This is equivalent to the third way that we describe above for handling pixels outside of the boundary of images. We do this with a continuous version of the inner product. So rather than represent the point with the Euclidean coordinates (7. (x. We do the same thing with continuous functions.(1. For example. we can represent this same point in this new basis with the coordinates (4.1) for some choice of u and v.3).y) = u(1. if I give you two basis vectors that are not orthonormal.0) and (1.2.As an example of a less attractive basis.0) + v(1.. We will also consider only periodic functions that go from 0 to 2π.1). though. f(x) and g(x). And. for k = 1. and then repeat over and over again. We can compute u and v by taking v = y. The main thing I’ll do to make things easier. and integrate (add) the result. rather than discrete ones. is: . and the others are just sines and cosines of increasing frequency. and then add up the results. to compute the inner product we multiply together the values of matching coordinates. but all this extends to 2D images.3. given any point.3).0). defined in the domain from 0 to 2π. supposed we represented points using the two basis vectors (1. Still.
and building them into a vector. Doing this requires solving some integrals of trigonometric functions. f. with small rapid changes in its value. When we say that two functions are orthogonal. That is.f> = 1. a2. a1. If a function is very bumpy.g = 2π f ( x) g ( x)dx 0 When we say that a function. We can use this definition to show that the functions in the Fourier Series have unit magnitude. a1. If we can eliminate these highfrequency components. as we sample the function more and more densely. To show that the Fourier Series forms a good coordinate system for all functions we must also show that we can express any function as a linear combination of the elements of the Fourier Series. and showing that this is equal to 1. then the definitions in the continuous case give us the same result as the limit of what we get in the discrete case. by taking the inner product of each function with itself. …) just by taking inner products. these rapid changes will be due to the high frequency component of the function. a2. we can make the function smoother. such as that this only works when f is a continuous function. So: . b1. we can write: f ( x) = a0 1 2π + ∞ k =1 bk cos(kx) π + ak sin(kx) π Then the values (a0. has unit magnitude. We say that the part of the function that is due to longer frequency sine waves is the low frequency part of the function. One reason that these are sensible definitions is that if we approximate a function by sampling its values. we mean that <f. As with an orthonormal basis for vectors.g> = 0. but we will skip the proof here. b1. This is a very important fact. b2. We can show they are orthogonal by showing that their inner products with each other are zero. f.f . we must show that for any function. b2. The part that is due to high frequency sine waves is the high frequency component of the function. This result means that any function can be broken down into the sum of sine waves of different amplitudes and phases. Keep in mind that I’m leaving out some necessary conditions. Implications of the orthonormality of the Fourier Series. the orthonormality of the Fourier Series means that we can use projection and (a generalization of) the Pythagorean theorem. We can solve the for the values (a0. …) are the coordinates of the function in this new coordinate system provided by the Fourier Series. we mean that <f.
As an example of its significance. This means that smoothing with a Gaussian reduces high frequency components of a signal. this means that convolution with a filter attenuates every frequency by different amounts. F. and that F just consists of a single sine wave. Then the Fourier Series representation of F is: a1 = π . To some extent. ie. attenuating high frequency components of a signal can be taken to be the definition of smoothing. and when we multiply these by the corresponding components of I we always get 0. The more we smooth. the broader g is the narrower G is (and vice versa). 1 2π = 2π f ( x) 2π 2π dx bk = f .G. This is because of the convolution theorem. As another important example. F=sin(x).a0 = f . That is. This means that no matter what I is. g. This is: 2π 2 f 2 ( x)dx = a0 + ∞ 1 (a 2 i + bi2 ) 0 Convolution Theorem Now we get to the reason why the Fourier Series is important in understanding convolution. Let f(x) be an averaging filter. F. This is a remarkable theorem. with an image I. f*g = h FG = H. if we define J=F*I. we can compute its Fourier Series as: . all other components = 0. which is constant in the range –T to T. dx cos(kx) 0 π = 2π f ( x) cos(kx) 0 π dx sin(kx) π = f ( x) sin(kx) 0 π The analog to the Pythagorean theorem is called Parsevaal’s theorem. and H are infinite vectors. Then. G. and we can understand the effects of convolution by finding the Fourier Series of the convolution kernel. Let F. If g(t) is a Gaussian. ak = f . More generally. the more we reduce the high frequency components. by FG. H be the fourier series that represents the functions f. We can also see why simple averaging isn’t as good a way of smoothing. this is why we learned the Fourier series. The convolution theorem states that convolution in the spatial domain is equivalent to componentwise multiplication in the transform domain. suppose we convolve a filter. and h. we point out that the Fourier series representation of a Gaussian is also a Gaussian. It allows us to understand what smoothing really does. Here. then the components of the Fourier series of J will all be zero except for a1. we mean that we multiply each element of F by the corresponding element of G and use this as an element of H. because all other components of F are 0. In fact.
That means. assuming f is a periodic function. while some similar high frequency components remain. at discrete intervals. f ( x) = a0 1 2π + T k =1 bk cos(kx) π + ak sin(kx) π . That is. f(nπ/T) for n = 0. Suppose now that f is bandlimited to T. If we graph it.T a0 = −T T 1 2π dx = 2T 2π 2 cos(kT ) k π − 2 sin(kT ) k π ak = −T T sin(kx) π cos(kx) dx = bk = −T π dx = This is called the sinc function. +1. +T. Also it has oscillations that produce odd effects. we have 2T + 1 uniform samples of f(x). Aliasing and the Sampling Theorem Sampling means that we know the value of a function. f. we see that it looks like: This does not decay as fast as a Gaussian. some high frequency components of an image will be wiped out by averaging. As a result of these oscillations.
So we get a unique solution. . To shrink images. If there is a higher frequency component. This means that we can reconstruct f from its samples. then this can play havoc with our reconstruction of all low frequency components.We know the value of f(x) at 2T + 1 positions. If we plug these values into the above equation. and this is the best we can do. These are linear equations. However. and be interpreted as odd low frequency components. Then. reconstruction is not possible. we get 2T + 1 equations. in fact. if we have fewer samples. the best thing we can do is to lowpass filter it and then sample it. The ai and bi give us 2T + 1 unknowns. we lowpass filter them and then sample them. This means that if we want to sample a signal. we get a signal that matches our original signal perfectly in the lowfrequency components.