【作者】 冯文毅；

【导师】 金国藩； 严瑛白；

【作者基本信息】 清华大学 ， 光学工程， 1999， 博士

【摘要】 本论文全面系统地阐述了光学子波并行处理技术的基本理论和构成策略。 深入研究了光学子波并行处理的实现技术及其在图象特征提取、纹理分割和模 式识别等方面的应用。 提出光学子波并行处理的基本思路和实现方法，包括阵列器件的组合、空 分复用、图象缩放旋转与滤波器多楔多环之间的均衡等，其关键是解决二维输 入图象子波变换输出的四维显示问题。在频域空分复用策略的指导下，构建了 微光学楔环子波检测系统，研制了集子波滤波、分束和会聚功能于一体的二元 光学全互连器件，通过仿真和实验证明了器件在特征提取上的有效性。针对器 件使用中出现的零级谱问题，提出了重新分布输出面滤波图象和增加Dammann 光栅减轻器件负担两种改进方案。 首次完整地计算和研究了体全息传统相关和子波相关衍射特性中sinc函数 对相关输出的调制规律和两个不同扫描记录方向的串扰噪声状况，表明子波变 换的引入除减小水平方向扫描记录的串扰噪声外，还明显减小了竖直方向扫描 记录的退化噪声，使并行相关输出由传统相关的一维扩充至二维。提出并比较 了子波变换与体全息存储的三种结合方式，选择具有最小串扰噪声的“子波特 征读取子波特征”方式构建了光折变晶体体全息子波相关处理系统。通过实验 验证了于波变换在改善相关输出质量和提高相关检测准确性方面的重要作用。 通过对晶体体全息子波相关处理系绞中几个重要问题（并行性能、多通道 拓展、子波参数选择和抗畸变识别能力）的专题研究和解决；全面提高了系统 的整体性能和实用效果。提出了系统并行性能的估算方法及相应的提高途径； 提出了系统的神经网络模型，利用该模型实现了子波滤波器参数的优化设计； 提出并实现了多输入通道和多子波通道并行处理的新构思和新系统，拓展了系 统的多通道特性；通过计算和实验对系统的抗畸变能力进行了研究和分析，得 出提高平移不变性的关键是增大变换透镜的焦距，还提出并实现了一种具有任 意角度旋转不变性的方案。 深入研究了光学子波并行处理的后续综合方法。提出模拟退火模糊聚类算 法实现了纹理图象的精确分割；通过子波展开特征的不同组合来提高识别的缩 放和旋转不变性。提出并实现了模糊综合评判和模糊均值聚类两种模糊识别方 ＞法，弥补了光学输出结果精度较低的不足，获得了很好的识别效果。 运用构建的光电混合型楔环子波检测系统和晶体体全息子波相关处理系统进行了纹理分割、关联检索、人身鉴别和车牌识别等方面的应用实验研究。更多还原

【Abstract】 The fundamental theory of optical wavelet parallel processing and its implementations are systematically described in this dissertation. The techniques for realization of optical wavelet parallel processing are analyzed in detail. The applications in feature extraction, texture segmentation and pattern recognition are explored.The basic ideas and methods to realize optical wavelet parallel processing, including combination of array elements, spatial division multiplexing and balance between input image’s scale-rotation and multiple wedge-ring filters, are proposed. The key problem of the implementations is to display the 4-D outputs of optical wavelet transform. A micro-optical wedge-ring wavelet detector is constructed by the means of spatial division multiplexing in the frequency domain. A binary optical element with functions of multiple wavelet-filtering, beam splitting and focusing is designed and fabricated. The simulated and experimental results testify the validity in feature extraction with the element. Two improved schemes, redistribution of the outputs and adding a Dammann grating to release the burden of the element, are brought forward to overcome the zero grade spectrum.Diffraction properties of conventional correlation and wavelet correlation based on the volume holographic storage are simulated perfectly for the first time. The modulation of sine function and the cross-talk noise on two different recording directions are studied. The cross-talk noise is decreased evidently with the join of wavelet transform in the volume holographic correlation system. The correlation outputs are extended from one dimension to two dimensions. Three combination forms of volume holographic storage and wavelet transform are proposed and compared. A volume holographic wavelet correlation system in a photorefractive crystal based on the combination form named "wavelet extracted features read out wavelet extracted features" is constructed, which has the minimal cross-talk noise. Experiments are performed to validate the importance of wavelet transform for the improvement of the correlation quality and the recognition accuracy.Several important issues of the volume holographic wavelet correlation system in a photorefractive crystal are discussed specially to improve the performance of the system. These issues are parallelism, multichannel development, choice of wavelet parameters and recognition invariance. An estimation method and some enhancementapproaches of the parallelism are proposed. A neural network is designed to optimize parameters of the wavelet filters. Two novel systems to implement the parallel processing of multiple input-image-channels and multiple wavelet-filter-channels are constructed. The invariance of the system is studied by simulation and experiments. It is concluded that the shift invariance can be improved with a bigger focal length of the transform lens, and a scheme with the rotation invariance at any rotation angle is proposed and realized.Post-processing methods of the optical outputs are studied thoroughly. A fuzzy c- means clustering algorithm combined with the simulated annealing mechanism is developed for texture segmentation. Different integration forms of the extracted features are used to improve the scale and rotation invariance. A fuzzy synthesis judgement algorithm and a fuzzy clustering recognition algorithm are proposed to improve the recognition performance of the system.The applications of the proposed systems in texture segmentation, associative searches, human identification and car plate recognition are performed by experiments.更多还原