Abstract

A multiresolution pyramidal image/data compression technique via the perfect convergent unipolar terminal attractor based associative memory is described. By properly choosing adaptive thresholds for dynamic iteration of the unipolar binary neuron states with an inner-product terminalattractor based neural associative memory,¹ it has been demonstrated via computer simulation that perfect convergence and correct retrieval can be achieved.² The perfect convergence feature of the neural associative memory can be used to reduce a 2D image/data into an image of much smaller dimensions, i.e., if n × n neurons are used as one unit cell, an image of N × N unipolar binary pixels can first be divided into (N/n) × (N/n) cells. For example, each cell can be associatively retrieved into two pre-determined memory vectors that may be considered as belonging to two bit planes. If one of the memory states is assigned a value of 1 and the other is assigned 0, then the combination of the two bit planes yields an image of a dimension being that of the original reduced by n × n. Continuation of the process will yield a pyramidal architecture of reduced image/data of different resolutions. Since in each layer the data features and format are dependent on the memorized vectors chosen, an abundance of choices can be made to tailor for specific needs of image processing applications.

© 1992 Optical Society of America