| This thesis records a continuous effort during the past several years in the exploitation of using moment invariants in neurocomputing for pattern recognition. In the introductory section of this thesis, the neuron as a computational unit and the concept of algebraic invariants are discussed in the common sense.;As the underlying body of the thesis, Hu's invariants of visual patterns are reviewed. A unique explanation of the significance of moments and moment invariants of different order is proposed. This explanation forms the basis of a new method of character recognition, which may provide additional information about feature selection or discrimination between characters. Image descriptors for character recognition are also considered, they are circular harmonic expansions for rotation invariant pattern recognition, Mellin transform for scale invariant pattern recognition and the combination the two, namely the Fourier-Mellin image descriptors (FMDs), for rotation and scale invariant pattern recognition. A method for accurately calculating the FMDs is proposed and is applied to the calculations of all the alphabetic-numeric characters. These 36 characters are designed as the reference patterns for pattern recognition, for which the geometrical parameters, Hu's invariants and the FMDs have been calculated and listed in the appendix.;Attention is then turned to three neural network models (Hopfield, Fukushima and Inter-Pattern Association) which are described in terms of the correspondence between these models and the biological nerve systems and the effectiveness of applying these models to pattern recognition. The information storage capacity of these models is also estimated.;Application of the moment invariants with neurocomputing begins with an investigation of the feasibility of using the image irradiance moments to replace the Hamming distance which is generally used in the criterion that shows the convergence in neurocomputing. Moreover, invariant pattern recognition is obtained by introducing the binary codes of moment invariants to neurocomputing. Combining moment invariants with neural network processing allows us to recognize patterns which have been subjected to various distortions, such as noise, translation, rotation and scale variation.;Finally, a brief discussion about a future study of the invariant pattern recognition and a summary conclude this thesis. |
