Efficient modification of LZSS compression algorithm

This paper presents a new method of lossless data compression called LZPP, being an advanced modification of the well-known algorithm LZSS [1]. It introduces improvements of the LZ family algorithms [2, 3], such as the use of a special coding of two and three byte matches, use of an auxiliary entropy coder and new criteria of symbol exclusions. Minimization of the data compression ratio (bpc) has been chosen as the primary goal of the proposed modifications of LZSS algorithm. 1. Analysis of LZSS algorithm and proposed modifications 1.1. Index coding The first thing which should be considered, when one wants to improve the quality of compression, are indexes generated by every LZ method. Indexes can be coded with one of the statistical adaptive compression methods, like the arithmetic or the Huffman coding. Confirmation of this is the sample probability distribution of indexes (calculated as the distance from the end of the window) shown in Figure 1, which LZSS method generates during compression of the “xplik” file (merged all files from Calgary Corpus). Graphs for the most of the files for which experiments were conducted had similar shapes. Therefore it can be assumed that generally indexes have such distribution. Moreover, symbols having such distribution should be compressed well with a statistical coder, because the entropy (see [4]) equals H(S) = 8,954 bits in this case. This means that the maximal static entropy coding would reduce the number of bits necessary to store indexes even by 44%. Using this fact in the LZPP method, the fast variant of arithmetic coding (RangeCoder [5]) was applied for index coding. The range coder was additionally equipped with a rotating buffer, handling of the escape marker and the exclusions mechanism. Additionally, an independent coding of index bytes ∗ E-mail address: [email protected] PDF created with FinePrint pdfFactory Pro trial version www.pdffactory.com Pobrane z czasopisma Annales AIInformatica http://ai.annales.umcs.pl Data: 06/09/2017 07:08:20