This article will outline the basic workings behind the Huffman Encoding technique. All materials were written by me but inspired from multiple web resources. Check the bibliography at the end of this article for some useful links.
The Hufman encoding technique was named for it's inventor, David A. Huffman, a professor at the University of California, Santa Cruz. The technique involves creating a mapping of literals into key codes based on their frequency. Most archivers use some form of Huffman encoding for their compression.
The following terms will be widely used throughout this article, so it's best if we define them in one spot.
code - a set of bits that will represent a literal
literal - a single unit of data, usually a byte from the source
source - the file or data stream we are compressing
The process of compressing data using Huffman encoding is rather simple once it's broken down. First, you calculate the frequency of each literal in the source, listing them in increasing order. Then you construct a binary tree using the frequencies, with the less frequent literals being lower leaves in the tree. Using this tree, you can construct bit codes for the literals.
Since the masses generally accept learning by example, we will do just that. Here is our test string that we will compress.
I am what I am and that's all that I am
Hopefully you can see why I chose this phrase. There is a healthy mix of common characters as well as some who appear only once.
Counting the occurence of each character, and sorting them by increasing frequency, we get the following (the last node represents the ASCII space):
As the diagram shows, each node pair contains the character literal and the number of times it appears in the source stream. The pairs are ordered sequentially based on their frequency, and lexicographically in the case where the frequencies are equal. These pairs will become the leaves of the tree we will build.
To build the tree, we take the first two frequencies and make a new node that holds the sum of the frequencies. We remove the first two nodes and place the new node in the list based on it's value (2 in this case).
We continue this process again, taking the first two nodes, adding the frequencies, and making a new node with the sum.
Again, we repeat the process.
At this point we are finally combining an original node with one of our new ones. Luckily, the process remains the same.
For completeness, I will include each step hereafter without descriptions, but if you've followed thus far, you should be able to see how each step is done.
In the above figure, see that the root node of the tree, 40, is the total count of all the literals in the source stream. At this point we can generate the codes for our literals.
The codes for our literals will be sets of bits of varying length. The code lengths for the more frequent literals will be shorter than lengths for rare literals. To generate a code for a literal, you traverse the tree, appending a '0' for a left branch, and a '1' for a right branch. So the code for the literal 'a' would be '00' and the code for 't' will be '011'. Below I have made a table containing each literal and it's Huffman code.
As you can see, the length of a literals code is inversely proportional to it's frequency in the source stream.
At this point we can represent our source stream using our newly generated codes.
Counting these bits we get 128, or 16 bytes, as opposed to the original 40. We've compressed it by 60%! But, this is only half of the battle. Somewhere, sometime you'll want the original stream back, so we move onto decompression.
If you've followed along this far, you should easily see that decompression is a snap. Given the alphabet of literal/code pairs in the above table, we can convert the stream of bits back into literals. Using the tree makes it even easier. Looking at the long line of bits above, the first bit is '1', so from the root of the tree, we go down the right branch. Since we're not in a "leaf" yet, we get the next bit from the stream, a '1'. We go down the right branch and we're still not in a leaf. The next bit '0' takes us down the left branch. Again, we must keep looking. The next bit '0' puts us in a leaf for the literal 'I'. Here we would write an I to the output. You continue this process until the entire stream is decompressed.
One issue you may have noticed has not been mentioned. What happens when one process compresses and another process needs to decompress? How will the decompressor get the alphabet? This is how all the many Huffman derivatives differ, each packing their alphabet into some type of block header before the compressed data. This method is truly dependent on the application. The resources listed at the end of this article will lead you to some more verbose methods.
Hopefully you now have a general idea behind the process of how to generate Huffman codes and compress/decompress a stream of characters. Here I have listed some valuable links related to this topic.
An excellent Java animated huffman example, it shows how the tree is built and traversed
definition of Huffman encoding
another good resource