Neural networks are characterized by massive parallelism, dense interconncection of processing elements(neurons), and information storage in distributed manners. Optical information processing using hologram inherently has such properties.
In this thesis, optical pattern recognition using holographic associative memory(HAM) is investigated, which is based on linear associative memory(LAM), quadratic associative memory(QAM), three-layer network, and adaptive learning rule.
As an application of LAM, rotation invariant pattern recognition is optically implemented. Patterns are stored in the HAM by using the scattered light holography, in which a memory pattern is associated holographically with its several(nine in this thesis) rotated patterns in different part of recording medium. For the input pattern rotated by any angle within -90°<θ<+90°, the associated memory pattern is reconstructed. Two memory patterns ㄱ and ㅅare stored in the HAM.
An optical system of QAM is proposed and implemented for the three memory patterns L, T, and X. The results of experiment and computer simulation show that the capacities of storage and error correction ability of QAM are greater than those of LAM.
In three-layer network, the SNR and storage capacity could be increased by introducing orthogonal intermediate patterns. Two holograms and a LCTV are used in optical system. In the experiment, four characters ㄹ,ㅅ,ㅇ, and ㅈ are taken as four memory patterns.
A holographic associative memory based on the adaptive learning which uses learning pattern method(LPM) is developed. The LPM utilizes the simple optical implementation of outer-product learning and the performance of adaptive learning. The learning patterns extracted from the interconnection matrix are stored in the HAM instead of the memory patterns. Results of the optical experiment and the computer simulation are represented for three memory patterns ㄱ,ㄷ, and ㅁ. This system will be a good solution for recognizing the characters or images which have similar or common parts in others such as human faces and fingerprints.