Optical memory disks in optical pattern recognition systems

Citation

Neifeld, Mark A. (1991) Optical memory disks in optical pattern recognition systems. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/5bbq-0q27. https://resolver.caltech.edu/CaltechETD:etd-07122007-073524

Abstract

We describe the use of optical memory disks in optical pattern recognition systems. Algorithmic and architectural issues associated with the realization of such systems are discussed. Experimental demonstrations of several optical disk-based architectures are included to aid in the understanding of system limitations and performance issues. First we discuss correlation-based pattern recognition and describe the relationship between this approach and the neural paradigm. The need for invariances in image recognition leads to the notion of the reference image library. This approach is shown to be attractive in the case of limited processor and spatial light modulator dynamic range. We characterize the optical disk as a parallel readout device. An overview of optical storage media is included. Parallel readout of data from Sony sampled format media is characterized. We identify a match between the characteristics of the optical disk and the requirements for pattern recognition systems. Four optical disk-based image correlators which may serve as building blocks in disk-based pattern recognition systems are introduced. These image correlators are experimentally demonstrated and compared in terms of speed, efficiency, and sensitivity to noise sources and disk imperfections. We discuss advantages and limitations of these systems.

We include a discussion of learning and generalization in neural networks. We present a new learning algorithm and discuss its generalization characteristics. Three disk-based systems for pattern recognition are proposed. The first is a correlation-based architecture. The performance of this system as compared with theoretical expectations is encouraging; however, data rate constraints suggest the investigation of an alternate approach. The next two systems are more neurally inspired and realize the k-nearest neighbor and radial basis function algorithms. An evaluation of the performance of these two systems is presented with respect to the handwritten digit recognition problem.

Lastly, we present two candidates for future optoelectronic computing and pattern recognition systems. We detail the operation of these architectures and discuss the need for a better understanding of the relationship between mass memory and a general parallel processing environment.

Thesis Files