Artificial neural system architectures for image processing
Utilizing an Artificial Neural Network System (ANNS) for Image Processing and Pattern Recognition is an advanced processing technique. The parallel processing function of the multi-Processing Element (PE) in the ANNS provides sufficient capability for increasing the speed of image processing. An optimized ANNS architecture can give full play to the potentiality of an ANNS. In this thesis a two-layer ANNS architecture is established for image convolution. This is the most popular image processing technique and can be employed with various algorithms, such as image enhancement, image segmentation, and image representation. In this thesis, I analyzed this architecture and concluded that the mathematical property of the convolution masks that can be applied to this architecture is that the rank of the mask matrix must be one. I also deduced an optimized PE arrangement of an ANNS with which one can approach the highest output-to-input ratio, or fewest input data, for a certain image. Since the data input and output time is one of the key points that affect processing speed, the advantage that comes with optimizing the PE arrangement is obvious. The thesis also describes a few other ANNS architectures that are most efficient for some special algorithms. In this thesis, I analyzed the difference between the processing speed of the ANNS with the optimized architecture and the traditional Single Instruction stream, Single Data stream (SISD) computer. The gains of processing speed as well as data input and output speed of the ANNS is significant. In addition, I also established corresponding architectures for some non-convolution image processing methods. The idea of using a multi-layer ANNS architecture to increase processing speed was suggested by Professor Sing T. Bow. The architecture developed in this thesis has been simulated in a personal computer. Several images were used to evaluate the affectivities of this application. Results obtained are satisfactory.