【作者】 许廷发；

【导师】 宋建中；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 光学工程， 2004， 博士

【摘要】 本文主要研究的是Gabor小波神经网络算法及其在灰度图像目标识别中的应用。研究涉及神经网络理论、小波分析理论和小波神经网络理论及其它们在目标识别中的应用技术。论文以基于多CPU神经网络目标识别系统的设计与实现为宗旨，在理论和实践两个方面研究和探讨了系统实现过程中所涉及的关键算法和关键技术。论文包括以下诸方面内容： 对目前国际、国内BP神经网络算法、小波变换理论、Gabor小波神经网络算法以及在目标识别中的应用现状进行了综述，对存在的问题进行了剖析。同时以灰度图像目标识别为主线，阐述与论文直接相关的基本理论和关键技术。 详细分析了多层前馈型神经网络描述及训练算法机理，从数学的角度推导了误差逆传播算法(BP算法)，同时指出了BP算法存在的问题。构建了一种用于多目标识别的改进的BP算法。 阐述了Gabor滤波器原理。给出了Gabor小波滤波器的表达式。根据理论分析和实际需要，设计了多通道Gabor滤波器，提取了灰度图像目标纹理特征。相应地给出了一个基于多通道Gabor滤波器特征神经网络识别算法。 给出了相位叠加不变量pc的定义。从理论上详细探讨了用Gabor小波求解相位叠加不变量PC的数学过程，给出了相应的PC不变量的公式。 详细分析了log gabor函数性能，给出了用log gabor求相位叠加(PC)不变量的修正公式。最后，设计了基于相位叠加(PC)不变量的神经网络目标识别新算法。 阐述了三种不同的目标表示法，即基于特征的目标描述法、基于模板的目标描述法和基于Gabor小波神经网络目标描述法，同时分析了三种表示法的优缺点。详细阐述了Gabor小波神经网络构建方法，重点构建了Gabor小波神经网络的训练算法，给出了理论分析和算法的具体步骤。 本文的主要研究思想就是：运用并行的Gabor小波神经网络算法对灰度图像目标进行实时识别。研究的思路：根据前馈神经网络模型(BP网络)和Gabor小波理论在目标特征提取和识别中的处理方法，有机地把二者结合起来。同时构建了用于灰度图像目标识别的Oabor小波神经网络模型，应用优化理论和自适应技术，使目标识别系统达到实时处理的目的。 本文研究的主要成果和创新点有： (1)提出了一种适于多CPU目标识别系统改进的BP神经网络训练算法。主要是用变步长方法及输入向量归一化方法，对所选激励函数在理论上做了调整，最后，在多CPU的目标识别的神经网络系统上实现了该算法。 (2)基于多分辨分析原理，设计一种新的基于Gabor小波的多通道滤波器。该滤波器能对低质量的灰度图象目标进行特征提取，并有很好的鲁棒性。滤波器的中心频率是一个从低到高的范围，滤波器采用6方向，4尺度，对灰度图象直接进行小波变换，用Gabor小波变换系数的模的平均值和其标准方差来表示抽取的灰度图象目标的纹理特征，最后，把获得的小波特征输入到改进的BP神经网络分类器进行分类识别。 (3)提出了一种基于Log Gabor小波的低层次的图象分割、边缘特征中国科学院博士学位论文:Gabor小波神经网络及其在灰度图象目标识别中的应用研究提取的方法。主要是运用小波变换频域中的相位信息提取目标特征不变量，这种不变量对目标的亮度、光线变化具有不变性。运用PC特征不变量特征结合神经网络算法对灰度图象目标进行了识别，识别率明显的提高。 (4)提出了Gabor小波神经网络模型及其训练算法。主要从理论方面对Gabor小波神经网络进行了分析，同时运用BP算法原理构建了一种适合目标识别的Gabor小波神经网络训练算法。理论和仿真实验表明，Gabor小波神经网络的收敛性和鲁棒性均明显优于BP网络。该算法应用于目标识别时，不仅提高了识别的精度，而且克服了BP算法易陷入局部极小的缺陷。 (5)对多神经网络集成算法进行了初步的探讨研究，给出了一种用于自动目标识别(ArR)的神经网络集成算法，并结合算法设计了一个基于多CPU并行结构的多目标识别神经网络系统。该算法是运用不同规模和初始条件下形成的同种类型的神经网络分类器。主要运用的是改进的BP神经网络，并把每一个神经网络输出映射为后验概率，然后进行加权平均判决。最后利用设计的神经网络系统进行了实时识别。更多还原

【Abstract】 The dissertation is the study of Gabor wavelet neural network algorithm and its application in gray image target recognition. It is involved with the neural network theory, the wavelet analysis theory ,the wavelet neural network and their application technology of the target recognition. The key algorithms and technologies in the system realization are studied and discussed from the theory and practice , the aim of the research is the designing and realizing the neural networks target system based on many CPU .This thesis mainly consists of the following parts.the actuality of the BP neural network algorithm, the wavelet transforms theory and Gabor wavelet neural network algorithm and their application of target recognition in the world are introduced. The problems are analyzed. The basic theory and the key technologies are expounded, the main toll route are gray image target recognition.The multiplayer forward neural network and its training algorithm are thorough analyzed, Error back propagation algorithm is derived from the mathematic , the problem of BP algorithm is indicated .The improved BP algorithm with many target recognition is constructed.The principle of Gabor filter is expounded. The expression of Gabor wavelet filter is presented. The multi channel Gabor filter is designed based on theory and practicality, the texture features of gray image target are extracted. The neural network recognizing algorithm based on multi channel Gabor filter feature is presented .The define of phase congruence (PC) invariant is introduced. The mathematic process of deriving phase congruence (PC) invariant is discussed from theory ,the formula of PC invariant is presented.The performance of log gabor wavelet is detailedly analyzed, the modified formula of PC invariant is presented by the log gabor wavelet. Finally.the new algorithm of neural network recognition based on the PC invariant is designed.The three target representations are expatiated, that are the target representation based on the feature , based on the template and based Gabor wavelet neural network, the advantage and disadvantage are discussed. The construct method of Gabor wavelet neural network is expounded. The training algorithm of Gabor wavelet neural network is constructed,the theory analysis and the concrete step of algorithm are presented.The mostly thought t are real time recognizing gray image target with Gabor wavelet neural networks algorithm. The train of thoughts are the forward neural networks (BP net) and Gabor wavelet are organically combined based on they were applied in target feature extraction and recognition. A model of Gabor wavelet neural network is constructed with automatic target recognition, the real time process aim is realized with the automatic target recognizing system applying optimizing theory and self-adapt technique. The neural networks ensemble algorithm is realized with neural network processing system designed based many CPU ,and the good impact is gained when it is applied target recognition .Mostly research findings and innovations are focused on :(1) The improved BP neural networks training algorithm was given out suit to many CPU target recognition system Principally with alter step and input vector , excitation function was adjusted from theory, finally this algorithm was realized in target recognizing system based on many CPU.(2) According to multi resolve principle,a new multi channel filter based Gabor wavelet was designed. It can extract the feature of low quality gray image target, and had good robust. Its center frequency was the range from low frequency to high frequency, its orientation is 6 and scale is 4. Gary image was directly transformed by these wavelet filters, the feature of extracting gray image target was denoted by the coefficients of Gabor wavelet transform and its standard variance, the wavelet feature was input to the improved BP neural networks to classify.(3) The method of image segment and edge feature extraction of low level was put f更多还原