光折變全像術與圖形辨識

Abstract

本論文係研究光折變效應及其應用，包含全像光學資訊儲存與處理的技 術，以及智慧型之光學影像辨識。 在光學資訊儲存方面，以鈮酸鋰晶碟片為儲存介質，建立一體積全像光 學資訊儲存系統，分別以角度多工在體積光碟片上一個位置儲存了200張 影像， 以及用旋轉多工在一個位置儲存了90張影像，資訊密度可達10^8 Bytes/cm3。其 次，建立數位資料儲存技術，將電腦檔案轉成光學編碼 圖形，再以角度多工的 方式，將一圖案編碼所得之26頁之二位元式資 料存入晶片中，其檔案大小約為 15KBytes，而重建解碼後圖案之錯誤 率只有約0.02%。其次，為求解全像光學儲 存容量之極限，我們從理論 探討系統幾何結構及繞射效率限制對儲存容量之影 響。在系統應用上 ，相干辨識之技術被結合至儲存系統中，以角度多工的方式 儲存50個 相干器之匹配濾波片，平行進行50個頻道之光學圖像快速辨識。此 外，我們並從理論上分析此多頻道相干辨識器的信-噪比，得出系統儲存 頻道數 目的極限參數。 在智慧型光學圖形辨識方面，首先以光折變動態全像術，實現光學認知 學習神經網路系統，並由理論及實驗上探討其收斂的特性。然後，為克服 收斂 問題，以光電方法組成混合式之光學認知網路，與體積全像術配 合成為10個頻 道平行運算之手寫字形辨識系統，它可容忍輸入字形有 變形。 最後，探討儲存介質之物理特性對全像光學 儲存系統的限制。我們從理 論及實驗探討晶體吸收光所產生的熱 效應及光衰減，對儲存光柵的均勻性及布 拉格條件之影響，以及探討 晶體的單方向能量耦合效應的影響，找出記錄光折 變光柵的最佳化條 件。本論文的研究結果顯示，光折變晶體作為光學資訊的儲 存媒介十 分可行，而它在資訊處理方面有許多應用，深具潛力。 An investigations on photorefractive holographic memories and its applications are presented. At first, a holographic data storage system using a Fe:LiNbO3 crystal volume is studied. Two hundred images are successfully stored at one area in the crystal by using the angular multiplexing technique. The storage capacity is estimated 108 bytes/cm3. The technique for storing digital data in holographic memory is then studied. A computer file has been encoded as an optical pattern and then stored in the crystal volume. In the experiment, a file contains 26 Kbytes data is encoded and stored. The bit-error-rate of reconstructed image is less than 0.02%. By considering the geometric constraints and the dynamic range of the diffraction efficiency, the storage capacity of the system is theoretically derived. To explore the application of holographic memory for optical pattern recognition, a multi- channel correlator is constructed. Fifty holographic matched filters are stored at different reference angles in a crystal volume and a multi-channel image correlation are performed in parallel. The signal-to-noise ratio and the channel capacity of the multi-channel correlator system is theoretically derived. In the second part of the dissertation, the photorefractive memory is used as the dynamical recording medium for the learning mechanism in optical neural networks. An optical network for the implementation of perceptron-like learning algorithm is constructed. The convergence properties of the optical network is theoretically and experimentally investigated. In order to improve the convergence behavior of the learning network, a hybrid network is developed and realized. The optical learning network has been combined with the volume holographic memory to form a image classification system. For a demonstration, Chinese characters which are hand-written by ten persons are used as the data base. After training, the system performs parallel operation of image classification with distortion-tolerant capability. In the third part, we consider the effects of absorption-induced beam depletion and material heating on the index modulation profile, phase profile, diffracted beam profile, and Bragg selectivity of holographic grating. We have derived a "smart scheduling" technique for compensating the absorption-induced beam depletion effects. In addition, we also consider how the nonreciprocal beam coupling effect in photorefractive crystals alters the index modulation distribution of the grating. Equations that describe the dynamic behavior and diffraction efficiency of photorefractive gratings have been derived. The optimum conditions for recording a photorefractive holographic memory is obtained.