Simulation, Design and Development of an Artificial Neural System (ANS) for Spatio-temporal Pattern Recognition

Abstract

An artificial neural system has been designed and developed for spatio-temporal pattern recognition. Bi-directional Associative Memory (BAM), Adaptive Resonance Theory (ART), Kohonen bit map and Fuzzy System have been utilized to develop the system. The system is gain controlled. Several approaches have been made to learn temporal and spatio- temporal patterns in the framework of Neural Network. ANS are advantageous in dealing with low-level high resolution of signal processing. Feed forward network for learning static patterns has been adapted. In designing the network and to process the information through it, both time-domain and frequency domain nature of spatio-temporal patterns has been considered and analyzed. Software has been developed to realize the system utilizing programming C++. Considering a variety of inputs, the system has been simulated, tested and finally outputs have been obtained. It has been found that the system exhibits the capacity to recognize & display more than one pattern at a time. Input vectors and weight matrix have been determined, which performed the convergence & similarity between the two patterns. A three-layer fully interconnected in feed forward neural network has been presented in this dissertation to recognize the patterns of Bangla alphabet, English alphabet, sound recognition, frequency measurement, any special pattern input, the territory map of Bangladesh, numerals, four geometric figures, pictures, any special symbol and spatio- temporal pattern for recognition. It has been realized that the neural network approach with BAM, ART utilizing Kohonen feature map and Fuzzy System have better in parallel patterns as compared to other methods. The system has been developed to recognize any pattern of any specific consideration