Competitive learning algorithms for vector quantization
Neural Networks
| Hi-index | 0.00 |
Summary form only given, as follows. Fractal based data compression has attracted a great deal of interest since Barnsley's introduction of iterated functions systems (IFS), a scheme for compactly representing intricate image structures. This paper discusses the incremental development of a block-oriented fractal coding technique for still images based on the work of Jacquin (1990). A brief overview of Jacquin's method is provided, and several of its features are discussed. In particular, the high order of computational complexity associated with the technique is addressed. This paper proposes that a neural network paradigm known as frequency sensitive competitive learning (FSCL) be employed to assist the encoder in locating fractal self-similarity within a source image. A judicious development of the proper neural network size for optimal time performance is provided. Such an optimally-chosen network has the effect of reducing the time complexity of Jacquin's original encoding algorithm from O(n/sup 4/) to O(n/sup 3/). In addition, an efficient distance measure for comparing two image segments independent of mean pixel brightness and variance is developed. This measure, not provided by Jacquin, is essential for determining the fractal block transformations. An implementation of fractal block coding employing FSCL is presented and coding results are compared with other popular image compression techniques. The paper also present the structure of the associated software simulator.