Cluster Analysis based on Self-Organizing Maps in Fractal Image Compression

Abstract

Fractal image compression is a novel and attractive technique for still Image compression. The most significant advantages are high reconstruction quality at low coding rates, rapid decoding, and resolution independence. One of the major limitations is its high encoding time complexity. Many techniques have therefore been developed and are being developed to overcome this limitation. Different classification and clustering algorithms exist in the literature to classify the domain block pool accordingly so that the search space is reduced for each range block, which can reduce the encoding time complexity to a large extent. An algorithm based on Kohonen's Self-organizing maps is proposed in this thesis. for the cluster analysis of image data. This algorithm, having its roots in neural networks is different from the classical clustering algorithms and based on neighborhood relationships. Kohonen's SOM algorithm is useful as it rapidly generates high quality clustering and gives good results when using small number of training vectors with an arbitrary initial codebook. This scheme may be combined with other classical algorithms to overcome the limitations ofKohonen's algorithm. This thesis explains how Self-Organizing Maps clusters the image data and how this concept is used for reducing the encoding time complexity of fractal image compression. Implementation has been carried out using Matlab6.0 from Math Works Inc with SOM Toolbox. Experimental results obtained are presented and demonstrate the effectiveness oC the proposed algorithm.