26 views

Uploaded by rajain135

d

- Steganography 1584
- project 2
- User Guide Matlab
- UNIT-II ITC.ppt
- Mad
- Kale Ida Graph Tutorial
- DATE SHEET.pdf
- A Simple Algorithm for Data Compression in Wireless Sensor Networks
- Ch_4-6_pp INSY
- 07-communicationReliability
- 9781783988648_Learning_ArcGIS_Geodatabases_Sample_Chapter
- Code Size Compression
- Nr-410507-Digital Speech and Image Processing
- Pranav Steganography Report
- Introduction
- Chapter3 Contents
- srpwmc final
- 24 ComputerSC IT
- History Crifstools e
- DIP Syllabus

You are on page 1of 4

Minimizing the average code length of a source is not the only point in compression.

We have to decode it, too. There is a famous joke about the fundamental limits and

design strategies for compression algorithms:

Anonymous Engineer : "I have obtained an excellent compression algorithm

that encodes any source down to 1 bit!!! Well... I am still working on the

decoding algorithm."

This engineer will not be able to obtain the decoder for his/her compressor :-)

The radical solution of unique decodability is putting separations between symbols

in a message. For example, if we have a messagege like 01, 0, 1, 11, 1011, 110, 0,

11, ... then each symbol can easily be separated without any ambiguity. However,

This requires a new symbol (the coma symbol here) in out alphabet, and it adds

one extra symbol for each symbol in a message. Obviously, this is very inefficient,

therefore undesirable. However, we can do something else:

Consider a source alphabet consisting of 4 symbols :

occurence probabilities:

=0.5,

=0.25,

, each with

=0.125, and

=0.125. If

you use Eq. 3, you see that the entropy for this source is 1.75 bits/symbol. Let us

make some bit sequence assignments for these 4 letters, and see which assignment

sets are decodable, which ones are not, and why:

Letters

Probability

Code 1

Code 2

Code 3

Code4

0.5

0.25

10

01

0.125

00

110

011

0.125

10

11

111

0111

1.125

1.25

1.75

1.875

Average Length:

As you see, the first two codes have an average rate less than the entropy 1.75. It

is not possible that they are uniquely decodable. Just try to write down some of the

symbol representations one after another, and then try to read it. You will see that

you may read it in many different ways (similar to the Morse code example).

Code 3 is in one of the most desirable forms. It has exactly the minimum rate

(equal to the entropy) and it is uniquely decodable. Again, try to write a message

using the representations of these 4 letters, and then try to read it. You will come

up with only one way to decode the message. It is indeed uniquely decodable.

Code 4 is also uniquely decodable. However it has slight differences. First of all, its

rate is a little higher than Code 3, threfore it is not as efficient as Code 3. Secondly,

the decoding is slightly more involved. The decoder cannot determine the point

where the last bit of a symbol is encountered before the next symbol starts.

Whenever the next symbol starts, it starts with a 0, and according to the amount of

1's in the symbol, you can determine which symbol was sent.

The structure of Code 3 implies an important property for code generation. Notice

that Code 3 is not only uniquely decodable, but it is also "instantly" decodable. In

other words, we can decode the message while the symbols are arriving, without

any need of waiting for the end of message. These sorts of codes are called prefix

codes.

Test for unique decodability: It is easy do determine whether a code is uniquely

decodable, or not. Alternatively, we can test if it is a prefix code, or not. A uniquely

decodable code is a code which satisfies the following conditions for all of its binary

codewords. The prefix code test simpler.

Consider two binary codewords a and b, where a is k bits long, and b is n bits long.

Also assume that n>k (therefore b is a longer codeword). If the first k bits of b is

exactly equal to the codeword of a, we say that a is a prefix of b. The rest of the

bits of b (which have a length of n-k bits) are called the dangling suffix. As an

example, if a=01001 and b=010011001, then a is a prefix of b and 1001 is the

dangling suffix. The test for suffix code proceeds as follows:

Construct a list of all codewords in your alphabet.

Examine all the codewords and see if any code is a prefix of another code or

not.

If you find such a prefix case, obtain the dangling suffix, and add it to your list.

Continue examining your list and adding the dangling suffixes until one of the

following two things happen:

1. You come up with a dangling suffix which is a codeword in the list.

2. There are no more unique dangling suffixes

If you get the first outcome, then the code is not uniquely decodable.

Otherwise, the code is uniquely decodable.

Ex. 5 : Consider a code list {0, 01, 11}. 0 is the prefix of 01, and the dangling

suffix is 1. 1 is not a codeword. There are no more dangling suffixes, so stop. The

code is uniquely decodable.

Ex. 6 : Consider a code list {0, 01, 10}. 0 is the prefix of 01, and the dangling

suffix is 1. 1 is not a codeword, so add to the list. The new entry 1 is a prefix of 10

and 0 is the dangling suffix which is a codeword. So the condition "1" is satisfied,

and the code is not uniquely decodable.

The way to test if the code is a prefix code is simple: No code in the list should be a

prefix of another.

There is an alternative way of testing prefix codes and unique decodability. Form a

tree using your codewords. For example, 0 may corespond to a tree branch to the

left, 1 may correspond to a branch to the right. In this way, as an example,

codewords 0110 and 0101 produce the following trees:

Form a tree that describes all codewords. A prefix code corresponds to having

codewords only on the leaves of the tree. Therefore, we should no have a codeword

corresponding to any internal node. For example, The following code is a prefix

code: (corresponding to {0, 10, 110, 111})

On the other hand, the following code is not a prefix code: (corresponding to {0,

00, 1, 11})

The Counting Argument: The compression methods are limited. No method can

compress all of the possible sources. The argument is: No lossless compression

algorithm can compress all possible sources of size N to some value less than N.

Proof:

There are

different sources of size N.

Assume that we can compress all of these sources down to some size less than

N.

Now let us calculate how many different sources there are which have size less

than N:

There are

different sources of size N-1, there are

different

sources of size N-2, ... , there are 2 different sources of size 2, there

is one different source of size 1.

different sources of size less than N.

As we see, the total number of different sources with size less than N is less than

the total number of different sources with size N. This means that if we can

compress all of the sources with size N, then at least two of the sources should

produce the same compressed source. When we try to decompress that source, we

cannot decide which original source (with size N) it should correspond to. Therefore,

we conclude that we cannot compress all of the sources. In fact, sometimes, we

expand the source size instead of compression.

Exercise: Compress a file in your computer using winzip or a similar compression

utility. Now try to compress the zip file once more. In some cases, you may observe

that the file size expands after the second compression. Repeat the exercise by

trying to compress MPEG movie files that you can download from the internet. You

may notice that you sometimes get data expansion instead of compression. This

phenomenon is more common if you try to compress a source which has already

been compressed (MPEG and JPEG files are compressed files).

Turning back to the topic of uniquely decodable codes; the uniquely decodable

codes (prefix codes are a subset of them) has the following important property.

Kraft-McMillan Ineqality: Let C be a code with N codewords with lengths

. If C is uniquely decodable, then we have:

This inequality gives a lower bound on the codeword lengths similar to the entropy

expression in Eq. 3. The proof is rather involved, and is beyond the scope of the

course.

Lemma: If the set of integers satisfy the Kraft McMillan inequality

, then

. This is also an

important observation wich indicates that, we can really come up with codeword

assignments to alphabet symbols as long as we obey the Kraft-McMillan inequality.

In the next section, we will develop strategies to assign codewords which are as

near to the entropy expression (Eq. 3) as possible, and at the same time satisfy the

Kraft-McMillan inequality.

- Steganography 1584Uploaded byproof14
- project 2Uploaded bysunee
- User Guide MatlabUploaded byLatrice Dennis
- UNIT-II ITC.pptUploaded byAbhishek Bose
- MadUploaded byVenu Gopal
- Kale Ida Graph TutorialUploaded byMassimiliano Sipala
- DATE SHEET.pdfUploaded byBrahmpal Bhardwaj
- A Simple Algorithm for Data Compression in Wireless Sensor NetworksUploaded byLe Ly Tuong
- Ch_4-6_pp INSYUploaded byindlaf85
- 07-communicationReliabilityUploaded bySam Shoukat
- 9781783988648_Learning_ArcGIS_Geodatabases_Sample_ChapterUploaded byPackt Publishing
- Code Size CompressionUploaded byBenny Goh
- Nr-410507-Digital Speech and Image ProcessingUploaded byvasuvlsi
- Pranav Steganography ReportUploaded byPranav Gupta
- IntroductionUploaded byfiqipraramadhan
- Chapter3 ContentsUploaded byghangale
- srpwmc finalUploaded byapi-340769184
- 24 ComputerSC ITUploaded byQazi Khubaib Alam
- History Crifstools eUploaded byHammam Al falah
- DIP SyllabusUploaded bylavanyavinayagam1985
- Inequalities & GraphingUploaded byDaisyListening
- A CHAOS ENCRYPTED VIDEO WATERMARKING SCHEME FOR THE ENFORCEMENT OF PLAYBACK CONTROLUploaded byIJAET Journal
- ijgi-04-02004.pdfUploaded byAdi Rakhmat Putranto
- 108T3_S05shortUploaded bysrikanth
- H_263_1Uploaded byAndreea Cristina Rusu
- Quiz DsaUploaded bylyminh194
- Course CalendarUploaded byEsenia Guerra
- bang ma asciiUploaded byvkhanhk22
- 38_hw_lpUploaded byNguyen Thi Hong Hanh
- ledservicec finalUploaded byapi-340769184

- KesariUploaded byrajain135
- 181903Uploaded byrajain135
- GPSC_201617_32Uploaded byrajain135
- Stress Concentration Factors and Notch SensitivityUploaded byapi-3710585
- Schedule of Work Paul Day3!17!12 15Uploaded byrajain135
- MoiUploaded byrajain135
- 181903(1)Uploaded byrajain135
- lec2xUploaded byAnonymous 6iFFjEpzYj
- IndexUploaded byBhargav Ravat
- 14072016Uploaded byrajain135
- De-2A_5th Sem - Course Abstract(1)Uploaded byrajain135
- comp6-6Uploaded byrajain135
- 123Uploaded byrajain135
- PIG SetupUploaded byAkshat Singh
- CircularofIDPevaluationBEUploaded byJanak Mandowara
- Prof Member Application FormUploaded byrajain135
- Welcome to Institute of ...Ng Personnel Selection!Uploaded byrajain135
- Steam Stop Valve ModelUploaded byrajain135
- Feed Check Valve ModelUploaded byrajain135
- Process for Arriving at TWSS Specialist OfficersUploaded byVaibhav Chitra Mohan Muddebihal
- MSW A4 FormatUploaded byricardus25
- Important-Notice-80-99-65-90-75-96-68-78-2013-14-97-2013-14Uploaded byrajain135
- 455 1 CutOff Degree APUploaded byrajain135
- So Handout Eng(1)Uploaded byrikumohan
- GSRTC_201213_3Uploaded byrajain135
- GIDC_201314Uploaded bysandeep198989
- sdarticle(1)Uploaded byrajain135
- Bourden Guage ModelUploaded byrajain135

- 1y0-301Uploaded byqabbas
- elb-ag-1Uploaded byimsreejith
- DocumentUploaded byAnonymous 9DrJUf0mz
- 70-464 certification Guide and How to Crack Exam on Developing Microsoft SQL Server DatabasesUploaded byPalak Mazumdar
- Inoerp Project FundingUploaded byendy
- Memory Management in C++Uploaded byTommaso Matticchio
- HC110310001 HCNA-Security-CBSN Chapter 1 Network Security Overview V2.0Uploaded byajaykumar988
- Control Units TALAMUploaded byDavid Talam
- Ccna2-1 WANs and RoutersUploaded byreimarie12
- Amazon Interview QuesUploaded byJyapriya
- A2L FileUploaded byRam Krishan Sharma
- Nested VlansUploaded byericlmarxx
- Data Structures and Algorithms Lab7SectionAUploaded byFaizan Butt
- SAP TablesUploaded bySiva Prasad Nadikoti
- Final Sem 1 v1 PLSQLUploaded bycrispy_joy
- Cs9228 Web Technology Lab ManualUploaded bykeerthi_sm18
- Rfc 3284Uploaded bychetan666123
- Hitachi h SeriesUploaded byIsaac Santos
- Batch Script ArraysUploaded byognemalfirt
- Lab 01 RS_232 SolvedUploaded byCarolina
- mcUploaded byManase Romeo
- PIX-ASA - How to Upgrade a Software Image Using ASDMUploaded byprincesly
- Service Bus 12c - A Hands-on Tutorial.pdfUploaded byeduflormir
- Entity Framework Learning GuideUploaded byAndy00
- SAP Security MaterialUploaded byNaveenkrishna Mohanasundaram
- An Effective Approach to Extract Information from web pagesUploaded byIRJET Journal
- cisco 2 lab 1Uploaded byAdrian Bagayan
- RMAN Backup and RecoveryUploaded byprassu1
- Oracle ASMUploaded byVarun Krishna
- Data MaskingUploaded byanon_758644381