【作者】 霍义萍；

【导师】 王应宗；

【作者基本信息】 陕西师范大学 ， 光学， 2003， 硕士

【摘要】 光学信息处理是上世纪七十年代发展起来的一门新兴学科。1948年全息术的发明，1955年光学传递函数概念的建立，以及1960年新型的强相干光源——激光的诞生，是近代光学发展史上的三件大事，也是光学信息处理的基础，而透镜的傅里叶变换效应则构成了光学信息处理的理论框架。 光学图像识别是光学信息处理重要研究领域之一，它可以广泛应用于空中和地面军事目标的识别和跟踪以及工业自动化、机器人视觉等方面。目前光学图像识别主要有光学相关和光学神经网络两类实现方法。由于光学相关4-f系统的结构简单、输出结果明了，因而成为光学图像识别的重要方法。光学相关在光学自适应、目标跟踪、模式识别、机器人视觉等领域的重要应用前景越来越受到人们的高度重视与关注。早期的研究工作主要集中在利用胶片等手段进行非实时匹配滤波光学相关研究，近年来的一个重要发展方向是进行实时相关研究。这一工作可以大致地分为两个方面：一方面是利用电寻址空间光调制器作为接口器件，利用联合变换(joint transform)实现实时相关；另一方面是利用匹配滤波器的方法进行实时相关研究。利用光电混合联合变换可实现实时相关的方案，把光学和电子学有机地结合起来，使系统具有优化的性能。 光学相关的基本任务是对输入的信号是否与参考信号相同做出判断。传统的相关器把输入信号和参考信号相似性的度量转化为一个点的强度的测量，相关峰的形态是输入信号与参考信号一致性的标志。对于同一个参考信号，若输入信号与之相同，相关器给出一个强而锐的相关峰；若输入信号与之不同，相关器的输出则较为平坦。但是，在军事、医学、生物等领域常常遇到这样的情况，即需要识别的信号与参考信号仅有微小的差异，信号本身的结构也较复杂，这时传统的方法就显得无能为力。本论文针对以传统相关为基础的图像识别难以识别相似图像的缺陷，提出了差异相关的概念。差异相关对于相似图像具有很强的识别功能。 本论文主要做了以下四个部分的工作： 本文第一部分详尽地介绍了光学图像识别的基本理论。内容涉及非相干光相关器、Vander Lugt相关器(VLC)和联合傅里叶变换相关器(JTC)。Vander Lugt相关和联合变换相关都是通过对两个函数傅里叶变换的积进行傅里叶逆变换的方案实现相关。非相干相关器输出的反差太小，因而还不实用。JTC与VLC相比，具有容易进行实时识别以及自适应的特点。如果输入中包含多个物体时JTC的性能迅速变差，然而当匹配空间滤波器研制成功后，输入中包含物体的个数对VLC的性能不产生影响。本章还介绍了分数傅里叶变换和分数相关，分数相关亦可用于光学图像识别。 本论文第二部分针对以传统相关为基础的光学图像识别难以识别相似图像的缺陷，提出差异相关的概念。给出了差异相关的理论推导、构建了实现这一方案的光电混合处理系统，并给出了数值模拟结果。若两物体完全相同，则没有相关峰，只要有差异存在，就会给出明显的相关峰。差异相关系统对相似图像具有很强的识别功能，但相关峰的分布仅指示出差异单元的相对偏移量。 论文第三部分提出基于分数相关的差异相关。基于常规傅里叶变换的差异相关的方法虽然对相似图像具有很强的识别能力，但无法提供差异的位置信息。将分数傅里叶变换和分数相关引入差异相关系统，形成基于分数相关的差异相关。数值模拟结果表明，通过调整分数相关的级次，该系统不仅对结构复杂的相似物体具有很强的识别能力，并且能够指示出差异的位置所在。 论文的最后一部分提出了一种基于频谱相减的改善联合变换相关输出的方法。在联合变换相关输出平面上互相关峰的位置依赖于输入平面上参考图像与输入图像的之间位置关系，若输入面上二者距离较小，输出面上互相关峰与中间项将相互重叠，影响到观察者的识别。这一部分利用频谱相减的方法去除了输出平面上的中间项。这一方法将扩大联合变换相关的应用范围，有助于识别结果的判读。 本论文提出的识别微小差异图像的差异相关的光电混合系统具有结构简单。准实时的特点，为光学图像识别提供了一条新的思路。但如何快速地实现两个图像的完全重合，是一个尚待解决的问题。同时，考虑到实际情况下常容易得到的图像是非相干图像，还需要通过空间光调制器来实现向相干图像的转换。更多还原

【Abstract】 Optical information processing is a newly subject developed 1870’s. The invention of holography in 1948, the establishment of optical transfer function in 1955 and the nativity of laser-a new style strong coherent lamp in 1960 are not only three great invents in optics’ developing history, but also the foundation of optical information processing. Moreover the Fourier transform effect of lens constitutes the theoretical frame of optical information processing.Optical image recognition is an important field of optical information processing. It can be widely used in the field such as the recognition and tracking of military object in the sky or on the ground, industry auto-adaptation, robot vision, and so on. At present there are two realization ways of optical image recognition: optical correlation and optical nerve network. Optical correlation has become a most important way for its 4-f system structure is simple and its output is apparent. Because of the great potential application of it in optical self-accommodation, target tracking, pattern recognition and robot vision, optical correlation was regarded more and more important. Early researches centered in the study of non-realtime matching-filter optical correlation using films. Recently, a magnitude development aspect is the study of real-time correlation which can be divided into two parts: one is making use of EA-SLM as joint apparatus and realizing real-time correlation using joint transform; the other is the study of real-time correlation using MSF. It is proposed that optics and electrics should be combined to optimize the system.The basic task of optical correlation is to achieve a judge that the reference signal and the input one are same or not. Traditional correlator transfers the measurement of comparability between reference signal and input signal into the measurement of intensity of a point. Correlation peak is the sign of consistency of them. To a certain reference signal, correlator present a strong and sharp peak when input signal is the same as the reference and present a relative flat peak when input image is different from the reference. When there is little difference between complex images in such field as military, medicine and biology, traditional method is inefficient. Traditional correlator is unable to recognize similar images. To overcome the defect oftraditional correlator, difference correlation that can recognize similar images efficiently was proposed in this paper.The work of this thesis has four parts:In the first part, we will describe the basic theory of optical image recognition, incoherent optical correlator, Vander Lugt correlator(VLC) and joint Fourier transform correlator(JTC). Both VLC and JTC realized optical correlation via processing inverse Fourier transform on the product of two Fourier transforms. The contrast of incoherent optical correlator output is too small to put into practical use. Compared with VLC, JTC is characterized by easily processing real-time recognition and self-adaptation. The capability of JTC declined rapidly if there are many signals on the input plane. After MSF got great success, the number of input object did not affect the output. Fractional Fourier transform and fractional correlation are also introduced in this part. Fractional correlation is also a way of optical pattern recognition.In the second part of this thesis, difference correlation is proposed to overcome the defect of traditional correlator. We give the academic deduce of difference correlation, construct the photoelectric processing system of this scheme and present the numerical simulative results. If input image is the same as the reference, we will get zero output. When they are not identical with each other, the output will present obvious correlation peak. This system can recognize similar images efficiently and the distribution of the output-peak can indicate the relative offset quantity of diversity unit.In the third part of this thesis, we put forward the difference correlation based on the frac更多还原