Image Compression using Singular Value Decomposition and
Discrete Cosine Transformation
Andrew Haas, Mark Van der Merwe, Ann Wilcox
Math 2270-002 Spring 2017
Introduction
Matrices and Linear Algebra are an integral part of many applications in computing such as
image representation and manipulation. Using the properties of matrices and their various properties, we
can manipulate images in many useful ways, such as the compression of image data. Here we explore two
ways of manipulating image matrices for compression. The first is Singular Value Decomposition (SVD)
and the second is Discrete Cosine Transformation (DCT). SVD, as will be shown, is less effective than
DCT, which is regularly used in JPEG compression.
Singular Value Decomposition
Background
Singular Value Decomposition is similar to Eigenvalue Decomposition in that the singular values
of a matrix A are the square roots of the eigenvalues of the matrix ATA. The same principle for
Eigenvalue Decomposition, A=PDP-1 is then applied to Singular Value Decomposition. The reason that
we cannot use Eigenvalue Decomposition for images is because it only works for square matrices, so we
use SVD because ATA is guaranteed to be square. The general formula for SVD is given as A=UΣVT,
where A is an mxn matrix of rank r, U is an mxm matrix, Σ is an mxn matrix, and V is an nxn orthogonal
matrix. The Σ matrix is obtained by D, which is a diagonal matrix containing the first r singular values of
A in decreasing order, and filling in zeros around D to get the correct mxn dimension. The U matrix is
obtained by finding the left singular vectors of A, which are the eigenvectors of AAT. The V matrix is
obtained by finding the right singular vectors of A, which are the eigenvectors of ATA. This can be used
in image compression by “throwing” out values; as the values are in decreasing order, the majority of the
image information is contained in the first and largest values. Once the “useless” values are changed to
zero, the matrices are multiplied together to get a new image matrix, which corresponds to a smaller file
size, thus a compressed image.
Process
To do the Singular Value Decomposition for our image, we first uploaded the image to Maple.
Then, we converted the image to grayscale to get a 2-dimensional image, and converted the image to an
image matrix. After we found our image matrix, we performed the Singular Value Decomposition on our
image matrix to find the singular values. We had to manipulate matrix and create our own Σ matrix as
maple only gave us a vector containing the values. Once we knew our values, we decided on our k, which
would determine the amount of values we would keep and which would change to zero. For our k we
chose to start at 5 and increase at increments of 20 until we ended at 205. We then did the matrix
multiplication for each k, along with running the error, which shows the image quality, and calculated the
compression ratio.
Pictures
Original Image:
Compressed Images: K = 5,25,45,65,86,105,125,145,165,186,205
Data Analysis
The following is the data that we got from our experiment with Singular Value Decomposition.
k Image size (kb) Error (mse) Compression Ratio
5 138 .01752945165 .6145251397
25 218 .009504274225 .3910614525
45 251 .007137542359 .2988826816
65 275 .005789235824 .2318435754
85 293 .004861137596 .1815642458
105 309 .004151782269 .1368715084
125 322 .003585364946 .1005586592
145 333 .003115425366 .06983240223
165 342 .0027716585017 .04469273743
185 349 .00237558426 .0251396648
205 355 .002081830939 .008379888268
These are the graphs we made to help visualize our data:
The first graph represents how the image file size changes as k (# of elements we keep) increases.
As one could expect, each increase in k changes the file size by less than the previous element. This is due
to each consecutive element influencing the image less than the previous one. The reason that we stop our
experiment at 205 k is that we found that increasing k beyond that will result in a bigger file size than the
original file. The second graph compares the error between the compressed file and the original file. The
error decreases as we increase k because we now have more terms that influence the image, and as
expected the curve bottoms out due to each k having less impact than the previous one. Mean Square
Error(mse) is the way we measured the difference between two images. To find the MSE you square the
the difference between a pixel in the original image and the corresponding pixel in the compressed image.
By doing this for every pixel, summing the differences, and dividing by the number of pixels we get an
approximate error. This number represents the average squared difference for each pixel in the image like
the name suggests. Lastly the third graph represents the compression ratio of SVD as k changes. We
defined the compression ratio as 1 - the compressed file size over the original file size. This graph is the
opposite of the image size graph because as k increases the file size increases which causes the ratio to 0.
Discrete Cosine Transformation
Background
In order to perform a Discrete Cosine Transformation on we take advantage of the addition of
sinusoidal waves. In applying DCT, we express images as a series of data points as a sum of cosine
functions at different frequencies. The trick is to represent an image using a sinusoidal wave. We simply
let 1 be white and -1 be black and everything else be some gradient between the two. We can then
increase or decrease the frequency to change the image. Below is what is called the 8 by 8 DCT matrix
which shows how the frequencies of each of the sinusoidal waves and their combinations change.
Here we can see how the changes in the frequencies result in different images. The secret to DCT
is that we split a given image into 8 by 8 pixel groups, or submatrices, and then determine how we can
combine the above 8 by 8 images from our DCT matrix to form each of our original 8 by 8 matrices. By
combining the sinusoidal waves represented in the DCT, it will return to our original image
(approximately).
How much each square from the DCT matrix contributes is stored in a corresponding contribution
matrix. In order to determine the actual values in our contribution matrix we apply what is called DCT II:
The actual compression comes when we apply a mask upon the resulting contribution matrix. For
our purposes we applied a simple mask that removed all but some of the values in the top left corner. The
reason we keep the top left is that those simpler matrices almost always contribute more to the 8 by 8
square we are dealing with. So, we remove most of the values and recombine the 8 by 8 matrices with
more zeros (thus taking up less space).
Process
We wrote a simple script using Matlab that uploads the image, then grayscales it (for simplicities
sake). We then created our 8 by 8 DCT matrix (like the one above) and determined the contribution
matrix for each 8 by 8 sub matrix in our image. We applied our mask to remove the smaller DCT values
and recombined. To test different levels of compression, we increased the number of values being kept in
each contribution matrix starting at just 1 and increasing by diagonals in our matrix to 34. We then, as in
our SVD work, calculated error using Mean Square Error and calculated compression ratio.
Pictures
Original Image is the same as for SVD.
Compressed Images: k=1,3,6,10,15,21,27,34
Data Analysis
The following is the data we got from our DCT experiment:
k Image size (kb) Error (mse) Compression Ratio
1 37 .0104 .8966480447
3 79 .0074 .7793296089
6 130 .0057 .6368715084
10 183 .0042 .4888268156
15 231 .0029 .3547486034
21 276 .0019 .2290502793
27 306 .0012 .1452513966
34 332 .00057695 .07262569832
These are the graph we made to help visualize our data:
The first graph represents how the image file size changes as k (# of elements we keep in each
8x8 sub matrix) increases. As one could expect, each increase in k changes the file size by less than the
previous element. This is due to each consecutive element influencing the image less than the previous
one. The second graph compares the error between the compressed file and the original file. The error
decreases as we increase k because we now have more terms that influence the image, and as expected the
curve bottoms out due to each k having less impact than the previous one. Mean Square Error(mse) is the
way we measured the difference between two images. The third graph represents the compression ratio of
DCT as k changes. We defined the compression ratio as 1 - the compressed file size over the original file
size. This graph is the opposite of the image size graph because as k increases the file size increases
which causes the ratio to 0.
Comparison of SVD and DCT
After experimenting with both SVD and DCT on the same image, they both gave us different
data. Although k represented different things in both SVD and DCT, we compared the two by graphing
the error vs. compression ratio of each.
Based on the graph, DCT is better at compressing images because it consistently has a lower error
for each compression ratios. What this means is that the image quality of a DCT compressed image is
much better than the image quality of a SVD compressed image of the same file size. This is because
DCT is much more efficient at removing unnecessary values. In conclusion, DCT is a much better way to
compress images.
