APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 1

The JPEG Image Compression Algorithm

John W. OBrien ([email protected])
AbstractOne of the most common methods of representing
graphical information in a digital form is by spatially sampling
the color components of an imagered, green, and blue for
example. These digitized samples can be placed in a long sequence
for the purpose of storage and transmission. For any reasonably
detailed image, this string of samples can become very long. As
of this writing, a typical digital camera will produce images in
excess of three million samples. Consequently, various methods
are employed to translate this representation into a compact
form. A particular approach to image data compressionknown
as the JPEG algorithmis examined herein.
The basis for the JPEG algorithm is the Discrete Cosine
Transform (DCT) which extracts spatial frequency information
from the spatial amplitude samples. These frequency components
are then quantized to eliminate the visual data from the image
that is least perceptually apparent, thereby reducing the amount
of information that must be stored. Finally, the redundant prop-
erties of the quantized frequency samples are exploited through
Huffman coding to produce the compressed representation. Each
of these steps is reversible to the extent that an acceptable
approximation of the original space-amplitude samples can be
reconstructed from the compressed form. This paper examines
each step in the compression and decompression sequence, with
special emphasis on the DCT.
Index Termscompression, jpeg, discrete cosine transform,
orthonormal basis transformation, vector projection
I. INTRODUCTION
A. Overview
The objective of JPEG image compression is simple: to
store the data necessary to reconstruct a digital image using as
little space as possible while maintaining enough visual detail
so that storing the image is actually worthwhile. It should
be noted early that, unlike compression techniques used with
other types of datalike nancial records, or word processing
documentsJPEG is a lossy algorithm, meaning that visual
information is selectively discarded in order to improve the
compression ratio. Once discarded, this information can not
be recovered, but it is selected to minimize the effect on the
perception of someone viewing the reconstructed image. In
order to determine which information can be removed, and to
perform the removal in a straight forward fashion, the image
data is converted from a space-intensity form to a space-
frequency form.
B. History
In the late 1980s and early 1990s, a joint com-
mittee, known as the Joint Photographic Experts Group
(JPEG), of the International Standards Organization (ISO)
and the Comit e Consultatif International T el ephonique et
T el egraphique (CCITT) developed and established the rst in-
ternational compression standard for continuous-tone images.
A good summary of their work can be found in [1]. The
standard they produced [2] has since become ubiquitous on
the Internet in the eld of digital photography. It has also been
extended and adapted for use in the encoding and compression
of digital videos.
C. Parallels
In general, people do not develop an intuition about spatial
frequency to the same degree as they do about temporal
frequency. The idea that a time signal has frequency com-
ponents, and that the signal can be analyzed and modied on
the basis of these components, should be familiar to anyone
who plays or listens to music. In that case the time signal
is the varying air pressure that manifests sound, and the
frequency components are the different pitches that we can
identify as notes. In fact, we characterize notes by describing
their relative fundamental frequency: low or high. Even on a
more subtle level, it is commonly understood that there is a
connection between the way something sounds and the relative
contributions of its different frequency components.
On the other hand, fewer people would be immediately
comfortable discussing the spatial frequency characteristics
of something, and might have difculty envisioning how an
image would look if you turned up the high frequency
elements of it. However, a basic, qualitative grasp of spatial
frequency in images is not hard. In an image, low frequency
elements represent smooth, gradual changes in color and
brightness, whereas higher frequency contributions are present
when an image contains sharp contrast and ne detail.
To get a general idea of the approach taken by the JPEG
algorithm, let us think briey about a simple audio signal. A
typical human ear can detect audio signals with frequencies as
low as 20Hz and as high as 20kHz; a factor of a thousand.
However, most human speech is conned within a range of
100Hz to 8kHz; less than a factor of a hundred. This is the
reason that the audio of a telephone conversation is encoded
at a rate of 64kbps while one channel of an audio CD takes
over 700kbps. Many of the frequencies unnecessary for the
transfer of voice information have been removed from the
telephone data stream, but the CD retains data representing
frequencies outside that range in order to facilitate richer, fuller
reproduction of tonal information.
D. Technical details
A digitized sample is known as a picture element or pixel
for short. Each pixel is most commonly represented by three
8-bit unsigned integers. The three values correspond to the
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 2
intensity of three color components
1
(e.g. red, green, and blue)
on a scale of 0 to 255. JPEG compression operates on each
color component separately, so the discussion that follows
will deal only with 8-bit pixels. It may be useful for the
reader to visualize grayscale images (like a black and white
photograph), wherein each pixel can be stored as an 8-bit
value indicating the amount of whiteness, instead of amounts
of redness, greenness, and blueness.
There are several variantsmodes of operationof the
JPEG algorithm included in the standard. We are principally
concerned with the Baseline DCT Sequential mode. This
compression method takes place in four stages, which will
be described in detail shortly.
1) The uncompressed source data is separated into 8 8
blocks of pixels. 128 is subtracted from the value of
each pixel so that the new effective range is from -128
to 127.
2) Each block is transformed into an 8 8 block of
frequency coefcients.
3) These coefcients are quantized.
4) An entropy encoder is applied to the quantized coef-
cients.
II. MODEL DEVELOPMENT AND MATHEMATICAL
FORMULATION
A. Denitions
An uncompressed image is stored as a two dimensional
array of pixels with width W and height H. We will use the
convention that pixels are indexed starting with (0, 0) in the
upper left corner of the image.
B. The Discrete Cosine Transform (DCT)
This transformation can be developed starting with the
Fourier series, extending to the Fourier transform, switching
from continuous to discrete, and nally using a cosine instead
of a complex exponential for the basis. The details of that
progression will not be reproduced here because they are
thoroughly covered elsewhere [3]. Instead, it will be most
useful to investigate the DCT in the context of vector spaces,
projections, and matrices.
1) Projections: We will begin by reviewing simple pro-
jections in R
3
, with the Euclidean dot product. Inspiration is
taken from Ch. 5 of [4]. Given an orthonormal basis we can
nd the projection of any vector v R
3
onto each of the basis
vectors, thereby giving a set of coefcients which can be used
to reproduce v as a linear combination of the basis vectors.
What does this mean? Well, a vector under the standard basis
of e
1
, e
2
, and e
3
(known to some people as

i,

j, and

k) gives
the number of units along the three principal axes; x, y, and z.
For instance, the vector (3, 5, 7)

means that you move three

units along the x-axis, ve units parallel to the y-axis, and
1
It turns out that the JPEG algorithm prefers a color space known as YCrCb.
These three components are called luminance (Y) and chrominance (Cr and
Cb). Luminance basically means brightness while chrominance has to do
with the difference between a color and a chosen reference color of the same
luminous intensity. The reference colors for YCrCb are red and blue as the
subscripts suggest.
nally seven units parallel to the z-axis. The standard basis is
implied. However, if it was agreed that we are to use a different
set of basis vectors, then that same vector would represent a
different point in space. For example, let us twist the standard
basis counterclockwise about the z-axis by 45

. The new basis

vectors are then given by
v
1
=
_
_
1

2
1

2
0
_
_
, v
2
=
_
_

2
1

2
0
_
_
, v
3
=
_
_
0
0
1
_
_
. (1)
Note that one of the vectors did not change. With this set
the same set of coefcients (3, 5, 7)

now means to move

three units along the line x = y, ve units parallel to the
line x = y, and seven units parallel to the z-axis. Under
the standard basis, this location would be represented as
(2/

2, 8/

2, 7)

.
In the preceding example, we observed how one set of coef-
cients can represent two different points in space depending
on the basis used. Next, and more relevant to the DCT, we
show how a single point can be represented by two different
sets of coefcients.
Let
v =
_
_

2
4

2
8
_
_
= c
1
v
1
+ c
2
v
2
+ c
3
v
3
.
(2)
Since the new basis dened in (1) is orthonormal, we can
nd the coefcients c
i
by taking the inner product of v with
each basis vector in turn. To show why this works, here is
the full computation for the rst coefcient, beginning with
a substitution from (2) and then using the properties of bi-
linearity (inner products in general) and of orthonormality
(these basis vectors in particular):
v, v
1
= c
1
v
1
+ c
2
v
2
+ c
3
v
3
, v
1

= c
1
v
1
, v
1
+ c
2
v
2
, v
1
+ c
3
v
3
, v
1

= c
1
(1) + c
2
(0) + c
3
(0)
= c
1
(3)
We form a new vector w with these coefcients,
w =
_
_
c
1
c
2
c
3
_
_
=
_
_
0
4
8
_
_
. (4)
The reader might wish to apply (1) and (2) to (4) to verify
this result.
2) Extending and generalizing: It might not be immedi-
ately apparent why this concept is especially useful in the
application of image compression. To begin to uncover the
true utility of this approach, let us examine more closely the
above example. Notice that the original vector v has three,
non-zero coefcients, while the transformed vector w has
a coefcient that is zero. In this relatively simple situation
(vectors in R
3
) sending one of the coefcients to zero is no
great feat. What if, instead of a three-dimensional space, we
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 3
were working in a higher-dimensional space? It might be the
case that several, even most, of the coefcients in a vector are
transformed to zero. That would introduce the possibility of
representing the vector in a compact way. If you think about
it, we do something like this all the time: an expression like
k = {1, 2, 3, . . . , 99} really means that k is the rst ninety-
nine integers. The . . . ,99 is a six character shorthand for
ninety-six of the members of k, which would otherwise have
taken one hundred eighty six digits and ninety ve commas
(two hundred eighty one more characters). By adjusting the
representation we are able to hold a lot of information in a
small space. However, this alternate representation depends
heavily on special attributes of the information being repre-
sented.
3) Images as vectors: Now we take this idea and apply
it to the task of image compression. Despite the fact that
any given sub-pixel (e.g. the green component) in an image
can take on any integer on the interval [0, 255], the relation
to that pixel and its neighbors is typically not that arbitrary.
Pick up any photograph and look for patterns in the variation
of color and brightness across the image. Depending on the
particular image you may see large patches with very little, or
very gradual, variation. Other areas of the image will have
larger uctuations from one point to an adjacent point, or
uctuations that are spaced more closely, but they might form
some repeating pattern. There are many combinations as well,
but these patterns represent a property of the image that we call
spatial frequency content. Just as a passage of music with bass
notes has low temporal frequency components, an image with
slow changes in tone has low spatial frequency components.
The next step is to quantify these frequency components.
4) Intensity to frequency: Consider a line of eight adjacent
pixels. Think of it as some part of one of the rows or columns
from an image, or just think of it as an independent bunch of
pixels that decided to hang out together and form a line. Now
we treat this row of pixels like a vector; a list of coefcients
indicating how much of the appropriate color component is
present at that spatial location. Each basis vector in this model
would be just like the standard Euclidean basis in geometry,
and have the unit value in one of the eight positions with
zeros in the rest. The unit value, in this case, is equivalent to a
255th of the maximum possible intensity. A minor adjustment
at this stage, as mentioned at the end of (I-D), is to use
a signed integer representation of the pixels, subtracting 128
from each intensity sample so that they now fall on the interval
[-128,127]. Later on when we decompress a JPEG le, the last
step in the decoding sequence will be to add 127 back to every
sample.
Now we select an alternate basis that will exploit the
patterns that are generally found in groups of pixels like this.
Since we are interested in the spatial frequency content of the
group, we choose basis vectors that are parameterized in a
frequency variable: those produced by cosine functions.
The basis vectors of the DCT are given by the function
d

[t] =
C()
2
cos
_
(2t + 1)
16
_
C() =
_
1

2
= 0
1 else
for t = {0, 1, . . . , 7}
= {0, 1, . . . , 7},
(5)
where t is the coefcient index, and identies one of
the eight basis vectors by the frequency used to generate it.
This denition is chosen to give the basis set the property
of orthonormality
2
. Fig. 1 shows what each of these vectors
would look like on an intensity versus position plot. The
highest values would correspond to the brightest pixels while
the lowest (largest negative) values are the darkest pixels.
= 0 = 1 = 2 = 3
= 4 = 5 = 6 = 7
Fig. 1. A largely qualitative view of the DCT basis vectors.
If we now apply the method described above in (II-B.1),
we can express the pixel vector as a linear combination of the
sinusoidal basis vectors dened in (5). The remarkable feature
of the transformed vector is that many of the high frequency
coefcients are often zero or close to zero. This happens
because continuous tone digital images typically have little or
no high frequency spatial variation. Look again at the = 7
basis vector in Fig. 1. The corresponding pixel values would
alternate between very bright and very dark pixels across the
whole row. Given that computer monitors have pixels that are
about 0.3mm wide, one may readily imagine that an image
with a drastically different color every fraction of an inch
would be unpleasant to look at.
In a group of eight frequency coefcients it would not be
out of the question to expect three or four of them to be zero,
and certainly four or ve could be very close to zero. If a
coefcient is nearly zero, it means that the contribution of
the corresponding frequency component is small, perhaps even
sufciently insignicant that it would not be missed if it were
to be discarded.
2
Orthonormality can be proved for this set of vectors by simply testing every
pair under the dot product. There are 24 pairs, each requiring 8 multiplications
and 7 additions for the dot product. A mere 360 arithmetic operations.
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 4
5) Spatial frequency in two directions: There is one more
extension of the DCT that is important for the JPEG algorithm.
Since images are represented as 2-D arrays of pixels, it is
advantageous to incorporate a quantitative measure of both
vertical and horizontal spatial frequency, not just one or the
other. The discussion in (II-B.4) dealt only with a line
of pixels. While an image can certainly be separated into
many such lines, and transformed accordingly, it is better to
transform blocks (sub images, so to speak).
6) Denition of the DCT: The 2-D DCT used for JPEG
image compression is dened as follows
3
. We use the general
form of the DCT basis vectors from (5).
F(v, u) =
7

x=0
7

y=0
p(y, x)d
u
[x]d
v
[y] (6)
where F(v, u) is the frequency coefcient with vertical
frequency v and horizontal frequency u, and p(y, x) gives the
value of pixel in row y and column x of the block.
This denition can be framed in terms of matrix operations.
Let
F =
_

_
f
00
f
07
.
.
.
.
.
.
f
70
f
77
_

_ f
vu
= F(v, u), (7)
P =
_

_
p
00
p
07
.
.
.
.
.
.
p
70
p
77
_

_ p
yx
= p(y, x), (8)
and
D =
_

_
d
00
d
07
.
.
.
.
.
.
d
70
d
77
_

_ d
t
= d

[t]. (9)
That is, the matrices F and P are direct analogs to the digital
storage of frequency coefcients and pixels respectively, while
D has the DCT basis vectors as its rows. By using the element-
wise denitions of (7), (8), and (9) to substitute in (6) we see
a slightly different form of the DCT equation.
f
vu
=
7

x=0
7

y=0
p
yx
d
ux
d
vy
(10)
By rearranging factors and grouping appropriately
4
we are
able to recognize (10) as the element-wise denition of matrix
multiplication:
F = DPD

(11)
3
The term two-dimensional (2-D) in this context refers to the shape of
the 8 8 pixel block, not the number of basis vectors of the DCT.
4
Recall also that the subscripts of a matrix element can be reversed by
taking the transpose of the matrix.
7) The inverse DCT (IDCT): Until now, this discussion has
been exclusively concerned with the compression of an image.
Let us not overlook the fact that a compression scheme is of
no use unless the process can be reversed and the information
decompressed. We now have all we need to prove that the
DCT is reversible. The proof is very nice: The matrix D has
full rank because its rows are a basis. Therefore D and its
transpose are invertible. Since the rows of D are orthonormal,
D is an orthogonal matrix, so D

= D
1
. The conclusion of
the proof gives us an equation for the IDCT.
D
1
F(D

)
1
= D
1
DPD

(D

)
1
= P
P = D

FD
(12)
If we take the transpose of (12) we nd that we can perform
the IDCT by using the DCT to operate on F

and then taking

the transpose of the result.
As (12) shows, the DCT transform and its inverse will
theoretically result in no loss of image detail. In practice, all
values are stored with some nite precision, and the cosine
functions are not ideal, so it is possible to introduce some
small error in the image reconstruction just by transforming
data back and forth from intensity to frequency. However, this
loss is insignicant compared to the loss that is intentionally
introduced during the next stage of the algorithm.
C. Quantization
After a block of pixels has been transformed to frequency
coefcients, it is quantized. This is the stage of the algorithm
when information is discarded. Quantization means that a
larger unit step size is selected for each element in the
8 8 block and the sample at that block is forced to the
nearest multiple of that step size. For example, if a frequency
coefcient after the DCT was 47.98, and the corresponding
quantization value was 28, the quantized frequency coefcient
would be 2. Note that any frequency coefcient with a value
less than half of the corresponding quantization value will be
forced to zero.
The quantization step can be formally dened in this way:
g
vu
= N
_
f
vu
q
vu
_
(13)
Where N(x) chooses the nearest integer value, and q
vu
is
the 8-bit unsigned integer quantization value.
The quantization table is the primary input parameter to
a JPEG compressor. There are approximately 1.3 10
1
54
valid quantization tables. Many are of little use because high
quantization values cause too much visual information to be
lost, ruining the image, and because it is inconsistent with the
common properties of images to keep information about high
frequency components while removing information about low
frequency components. The point is that the quality scale of
1 to 10 available through the user interface of most JPEG
compressors comes nowhere near the level of control that is
supported by the JPEG algorithm.
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 5
More than one quantization table can be used for a single
image (up to four). This is useful, for example, because our
eyes are less sensitive to variations in blue chrominance than to
luminance, so certain color components can be more heavily
quantized with less effect on our perception of the image.
Each quantization table that is used must be stored in the
compressed le.
D. Intermediate coding
After the transform and the quantization, an intermediate
coding strategy is applied to the blocks of frequency co-
efcients. Each 8 8 block is mapped into a 1-D array
64 elements long using a zig-zag pattern. This choice of
mapping is designed to place coefcients with similar overall
frequency (the combined horizontal and vertical frequency,
approximately speaking) in adjacent locations in the linear
array. Because the non-zero coefcients tend to cluster in
the upper-left corner of the block, the zig-zag pattern tends
to result in a linear array terminated by a long string of
consecutive zeros.
The second part of the intermediate coding is a mixture of
run-length encoding (RLE) and variable length integer (VLI)
encoding. The RLE technique is applied only to runs of zeros,
and the VLU coding is applied to all non-zero coefcients.
Each element that is produced by this combined coding has
this form: (run,size,value). Run indicates the number of zeros
that precede this non-zero entry and is represented by a 4-
bit unsigned integer. Size is also a 4-bit integer and indicates
the number of bits required by the value eld. Value itself is
the value of the quantized frequency coefcient. There is a
small difference in how the lowest frequency terms (DC) are
handled, but it is not necessary to discuss it for the purpose
of this paper.
E. Entropy coding
The most common entropy coding method used by the
JPEG algorithm is Huffman coding. Arithmetic coding is also
used, and produces slight smaller les, but is more complex
to implement. In essence, entropy coding attempts to further
reduce the number of bits necessary for storing each element
from the intermediate coded array based on the probability that
a certain element will appear. A discussion of the probabilistic
interpretation of information is worthy of its own paper, but a
brief example is included below which should give the reader
a good idea of how the technique works. A stream of Huffman
coded bits must be accompanied by a code table, or key, for
decoding the stream. The JPEG le format supports using
multiple tables (up to 2 for the Baseline DCT mode), which
means that the compressor can select two distinct probability
density functions and choose then match the distribution of a
given block to one or the other to give the best compression.
In essence, the DCT puts the data in a form that can take
best advantage of the compaction afforded by the intermediate
and entropy coding. The latter steps are where the data
compression actually happens.
F. Decompression
To decompress an encoded JPEG image each of the above
steps are reversed in reverse order.
1) Use the stored Huffman table to decode strings of bits
into a 64-element array;
2) reform the array into an 8 8 block;
3) multiply each element in the block by a corresponding
quantization value;
4) perform the IDCT on the dequantized block;
5) and add 128 to each of the de-transformed pixels to
recover an approximation of the original image.
III. NUMERICAL WORK AND EXAMPLES
To exercise the theory that the rst sections of this paper
developed, we will now choose an example image, select a
representative block of pixels, and perform all of the steps
outlined above.
See Fig. 2 for the source image we will be working with
5
,
where the selected block is shown in a magnied region.
Fig. 2. A representative block of pixels.
The MatLab code used to perform most of the following
computations can be found in Appendix I.
A. Intensity samples
The selected block of pixels has these values:
5
Photograph courtesy of The Give & Take Jugglers.
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 6
P
0
=
_

_
188 145 88 58 67 110 134 134
187 193 152 125 115 130 137 139
166 184 201 194 198 195 151 139
152 168 188 214 229 225 172 143
156 159 165 181 201 199 169 144
168 163 164 158 167 174 156 145
174 171 169 158 155 163 160 144
170 172 167 161 159 167 169 147
_

_
(14)
The visual-numerical analog can be seen, for instance, by
noting the association between the lowest value in the matrix
58, or 22.7% white, at location (0,3)and the dark pixel in
the top row of the block, and between the highest value229,
or 89.8% white, at location (3,4)and the nearly white pixel
near the center of the block.
B. Level shift
The sample range is shifted to be zero-centered by subtract-
ing 128:
P

0
= P
0
(128)I =
_

_
60 17 40 70 61 18 6 6
59 65 24 3 13 2 9 11
38 56 73 66 70 67 23 11
24 40 60 86 101 97 44 15
28 31 37 53 73 71 41 16
40 35 36 30 39 46 28 17
46 43 41 30 27 35 32 16
42 44 39 33 31 39 41 19
_

_
(15)
C. Forward DCT
Switching to a oating point representation we apply (11)
to P

0
to nd the frequency coefcients.
F
0
= DCT[P

0
] =
_

_
263.00 46.69 10.84 45.62
67.22 24.13 56.18 6.96
119.71 32.57 129.33 7.86
87.58 7.34 71.10 16.89
11.75 33.87 3.76 23.69
3.79 10.86 13.13 8.78
1.56 6.20 7.08 4.27
2.73 0.94 1.93 1.05


28.00 4.91 5.84 5.57
2.32 8.65 0.87 5.86
8.71 10.05 8.83 1.19
4.75 4.86 1.32 2.37
4.75 5.40 3.85 0.44
1.96 4.01 6.98 0.26
1.17 0.42 8.42 2.77
0.31 0.95 2.29 3.48
_

_
(16)
D. Quantization
The quantization table we will use is chosen from an
example in [1]. Note that the values in the upper-left corner
of the matrix, which correspond to DC and low-frequency
components, are fairly low (10-20), whereas the values in the
lower-right region of the matrix, which are used to quantize
the high frequency components, are much higher (80-120).
Q
0
=
_

_
16 11 10 16 24 40 51 61
12 12 14 19 26 58 60 55
14 13 16 24 40 57 69 56
14 17 22 29 51 87 80 62
18 22 37 56 68 109 103 77
24 35 55 64 81 104 113 92
49 64 78 87 103 121 120 101
72 92 95 98 112 100 103 99
_

_
(17)
Element-wise division F
0
[v, u]/Q
0
[v, u] and rounding gives
the block of quantized coefcients.
G
0
=
_

_
16 4 1 3 1 0 0 0
6 2 4 0 0 0 0 0
9 3 8 0 0 0 0 0
6 0 3 1 0 0 0 0
1 2 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
_

_
(18)
E. Zig-zag sequence
Most computers store 2-D arrays in memory in a row-wise
sequence. That is, all elements from the top row are stored
consecutively, followed by all elements from the next row,
and so forth. For the purpose of JPEG compress, on the other
hand, it is desirable to have as many of the zero coefcients
next to each other as possible to maximize the compression
available with the RLE and Huffman coding. Consequently,
the 1-D representation chosen for this 2-D array starts in the
upper left corner and sweeps back and forth diagonally. The
linear representation of G
0
would then be
G
0
= {16, 4, 6, 9, 2, 1, 3, 4, 3, 6, 1, 0, 8,
0, 1, 0, 0, 0, 3, 2, 0, 0, 0, 0, 1, 0, . . . , 0} (19)
F. Intermediate coding
Though the principal thrust of this paper is not entropy
coding, it is worthwhile to glance at an example. A general
sense of what is happening can be gained without extensive
analysis of the theoretical background of the approach.
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 7
1) Representing values: Commonly, the elements of G
0
would be stored in the computer as xed-length binary in-
tegers. Since they are so small, this choice would waste
many bits since a small number can be represented with a
small number of bits. The biggest number in this array can be
specied with ve bits (10000). To exploit this fact, a set of
variable length integer (VLI) codes are specied by the JPEG
standard. The ones that will be used for this example are given
in Table I.
value code size
-9 0110 4
-6 001 3
-2 01 2
-1 0 1
1 1 1
2 10 2
3 11 2
4 100 3
8 1000 4
16 10000 5
TABLE I
VARIABLE LENGTH INTEGERS USED IN THE EXAMPLE OF (III-F.1).
2) Intermediate and entropy coding: The block G
0
will
be entropy encoded starting at the rst element and stepping
through it to the last element.
6
Each of the corresponding
sequence of coded pieces of information will have the form
(run,size,value). As per the brief discussion in (II-D), the
intermediate code of G
0
would be
G
0
= {(0, 5), 10000, (0, 3), 100, (0, 3), 001, (0, 4), 0110,
(0, 2), 10, (0, 1), 0, (0, 2), 11, (0, 3), 100, (0, 2), 11,
(0, 3), 001, (0, 1), 0, (1, 4), 1000, (1, 1), 0, (3, 2), 11,
(0, 2), 01, (4, 1), 1, (0, 0)} (20)
Each (run,size) pair are the zero-run and value size, while
the numbers that follow are the VLI values from Table I. The
block is terminated with the pair (0,0) indicating that there are
no more non-zero elements.
It is the (run,size) pairs that receive the Huffman coding.
Table II shows all of the pairs that appear in 20, as well as
how many times they appear in this block.
Huffman codes have the following properties:
1) No code is a prex for any other code (e.g. if 010 is
valid, then 010... is not)
2) Codes that are more likely to occur have shorter path
lengths typically
Since we are only concerned for the moment with data from
a single block, it is possible to develop a Huffman table that
is beautifully suited to this block. The choice is shown in
Table II. Substituting from that table into (20) gives this string
of bits:
6
DC coefcients are actually treated differently than AC coefcients. The
values are stored as differences relative to previous blocks, and a different
coding table is used. However, this detail will be neglected for the purpose
of this example.
# pair code
4 (0,2) 00
4 (0,3) 01
2 (0,1) 100
1 (0,0) 101
1 (0,4) 11000
1 (0,5) 11001
1 (1,1) 11010
1 (1,4) 11011
1 (3,2) 11100
1 (4,1) 11101
TABLE II
A LIST OF (RUN,SIZE) PAIRS FROM (20), NUMBER OF OCCURRENCES, AND
HUFFMAN CODES.
G
0
= {
1100110000011000100
1110000110001010000
0110110000110100110
0011011100011010011
100110001111011101
} (21)
That is 94 bits. The original block of intensity samples
would have been 4096 bits long in memory. Of course, this
is not counting the storage space required to store the quanti-
zation table(s) and the Huffman coding table(s), but they are
shared among many blocks so the storage expense per block
is negligible. Assuming that we have already taken care of
storing the Huffman and quantization tables, the compression
ratio for this block is 2.3%, or about 0.18 bits per pixel.
G. Decompression
If the inverse of this process is applied to the compressed
data, a block of pixels P
1
will be generated. This block is
intended to be a reconstruction of the original P
0
, acknowl-
edging that there will be some difference due to the loss and
nite precision of the system. The
P
1
=
_

_
187 146 88 53 63 100 127 134
190 175 148 126 125 135 137 131
171 185 197 198 194 183 160 139
142 165 193 212 220 210 180 150
147 155 167 184 202 203 177 146
176 168 158 158 172 179 160 134
183 175 162 154 159 165 155 138
164 167 164 158 159 166 163 154
_

_
(22)
To see how much these samples vary from the original we
will also compute a percent error block. Note that most pixels
have changed by 1-5% and none more than 10%. There are
even some that have been reconstructed exactly.
APPM-3310 FINAL PROJECT, DECEMBER 2, 2005 8
E =
_

_
0.5 0.7 0.0 8.6
1.6 9.3 2.6 0.8
3.0 0.5 2.0 2.1
6.6 1.8 2.7 0.9
5.8 2.5 1.2 1.7
4.8 3.1 3.7 0.0
5.2 2.3 4.1 2.5
3.5 2.9 1.8 1.9


6.0 9.1 5.2 0.0
8.7 3.8 0.0 5.8
2.0 6.2 6.0 0.0
3.9 6.7 4.7 4.9
0.5 2.0 4.7 1.4
3.0 2.9 2.6 7.6
2.6 1.2 3.1 4.2
0.0 0.6 3.6 4.8
_

_
(23)
IV. CONCLUSION
This examination of the JPEG compression algorithm, in
particular the DCT operation, has demonstrated both the qual-
itative concepts behind the technique as well as the quantitative
processes that are used to apply them.
The important results that can be drawn from this work in-
clude an understanding of the scope of the technique. Perform-
ing a DCT on blocks in an image gives easily compressible
data because of the content of most images, not because of an
inherent constraint of the DCT. The other important result is
the framing of the DCT in terms of matrix operations. It is the
rst step in developing computational methods for computing
the transform quickly.
Study that could logically follow from this work would be:
1) an exploration of Huffman coding in the context of proba-
bility and information theory; 2) a review of the other modes of
operation of the JPEG algorithm (Progressive, Hierarchical);
and 3) applications of the DCT or similar transforms to the
compression and manipulation of other kinds of data (like
audio).
V. ACKNOWLEDGMENT
The author wishes to recognize and thank Dr. Anne
Dougherty for her thoughtfulness and high-standards.
REFERENCES
[1] G. K. Wallace, The jpeg still picture compression standard, Communi-
cations of the ACM, Apr. 1991.
[2] Recommendation T.81, International Telecommunication Union (ITU)
Std., Sept. 1992, joint Photographic Expert Group (JPEG). [Online].
Available: http://www.w3.org/Graphics/JPEG/itu-t81.pdf
[3] C. L. Phillips, J. M. Parr, and E. A. Riskin, Signals, Systems, and
Transforms, 3rd ed. Upper Saddle River, NJ: Prentice Hall, 2003.
[4] P. J. Olver and C. Shakiban, Applied Linear Algebra. Upper Saddle
River, NJ: Prentice Hall, 2006.
APPENDIX I
MATLAB CODE
% FILE: jpeg.m
% DATE: 2005-12-01
% DESCRIPTION: Example code to demonstrate
% the JPEG compression algorithm
% Block of pixel values
P0 = [188 145 88 58 67 110 134 134;
187 193 152 125 115 130 137 139;
166 184 201 194 198 195 151 139;
152 168 188 214 229 225 172 143;
156 159 165 181 201 199 169 144;
168 163 164 158 167 174 156 145;
174 171 169 158 155 163 160 144;
170 172 167 161 159 167 169 147];
% Quantization table
quant = [
16 11 10 16 24 40 51 61;
12 12 14 19 26 58 60 55;
14 13 16 24 40 57 69 56;
14 17 22 29 51 87 80 62;
18 22 37 56 68 109 103 77;
24 35 55 64 81 104 113 92;
49 64 78 87 103 121 120 101;
72 92 95 98 112 100 103 99];
% Perform level shift
P = P0 - 128;
% Construct elementary row operation matrix
% to normalize the DC basis vector
S = eye(8)/2;
S(1,1) = 1/2/sqrt(2);
% DCT basis vectors
D = zeros(8,8);
for t = [0:7]
for w = [0:7]
D(w+1,t+1) = cos((2
*
t+1)
*
w
*
pi/16);
end
end
D = S
*
D;
% Perform forward DCT
F = D
*
P
*
D;
% Perform quantization of coefficients
Q = round(F./quant);
% Dequantize
F1 = quant.
*
Q;
% Perform IDCT
P1 = D
*
F1
*
D;
% Level shift up
P1 = P1 + 128;
% Integer round
P1 = round(P1);
% Compute precent error
% to one decimal place
E = round((P1 - P0)./P0
*
1000)/10;