A Proficient Lossless Compression Using Predictor Nerual Network

In this paper, a new approach based on Artificial Neural Network for data compression is presented. ANN (Artificial Neural Network) is playing a vital role to do complex works like face recognition, pattern recognition, currency recognition, compression, decompression etc. Artificial Neural networks have the potential to extend data compression algorithms beyond the character level n-gram models now in use, but have usually been avoided because they are too slow and occupy more space. The new approach presented here produces better compression ratio and competitive in time than conventional algorithms such as Huffman, Limpel-Ziv (pkzip, gzip, Winzip), Prediction by Partial Match (PPM), Burrows-Wheeler. The compression technique used in this thesis is based on the neural network predictor consisting of two layers. The input layer has 4 million neurons and the output layer has only one neuron. Since the entire file of data is used as an input to the network, here 4 million neurons are used. The repeated occurrence of various characters’ information is stored in 22 bits hash function lookup array. The input to the hash function is the context, treated as a number and the output is the index of the corresponding neuron. To assign a code to each character using the predictor network algorithm needs more bits. In order to overcome this difficulty a new approach is employed to assign a code to the entire string instead of each character. This requires less number of bits for a particular character. The weights associated with the links are updated using one-pass training with variable learning rate. The compression ratio achieved by this proposed neural network approach is better than conventional algorithms such as Huffman, Limpel-Ziv, PPM and Burrow wheeler.