Cellular Associative Neural Networks for Pattern Recognition

A common factor of many of the problems in shape recognition and, in extension, in image interpretation is the large dimensionality of the search space. One way to overcome this situation is to partition the problem into smaller ones and combine the local solutions towards global interpretations. Using this approach, the system presented in this thesis provides a novel combination of the descriptional power of symbolic representations of image data, the parallel and distributed processing model of cellular automata and the speed and robustness of connectionist symbolic processing. The aim of the system is to transform initial symbolic descriptions of patterns to the corresponding object level descriptions in order to identify patterns in complex and noisy scenes. The scene is represented by the configuration of a cellular array. At the initial level, the states of the cells in the array represent local and elementary features of the objects. At every iteration, these local features are ‘connected’ together forming higher level features, ultimately forming the object level description. An associative symbolic processing element is placed in each cell of the array while the exchange of information and the state transitions that take place are controlled by the rules of a global pattern description grammar. These rules are produced using a learning algorithm which is based on a hierarchical structural analysis of the patterns. Efficient management of these rules in terms of speed and storage capacity is provided by the underlying neural associative symbolic processing engine of the system (AURA) which also facilitates its operation with increased tolerance in order to overcome problems caused by noise and uncertainty in the data. In order to present the basic characteristics of the architecture the system is tested in the task of recognising simple geometric shapes. The behaviour of the learning algorithm and the influence of various parameters defining the operation of the system are examined in these experimental sessions and a prominent characteristic is shown to be the robustness to noise. Yet from this initial stage, the current architecture demonstrates the advantages arising from the combination of cellular, neural and symbolic processing and also shows how a simple principle can provide an efficient learning algorithm.