基于特征的纹理特征提取、分类与检索方法研究

【作者】 许漫坤；

【导师】 平西建；

【作者基本信息】 中国人民解放军信息工程大学 ， 信号与信息处理， 2003， 硕士

【摘要】 纹理是图象中一个基本且重要的特性。纹理分析是图象理解、分析与识别中的重要研究内容之一，不仅对于视觉智能活动的模拟具有重要的理论意义，而且具有非常广阔的应用前景。纹理分析通常可分为基于特征和基于模型两类方法。纹理分析的研究主要为纹理描述、纹理分割、纹理分类、纹理检索等方面。本论文概述了早期和近期的多种纹理分析方法，主要研究和讨论了基于特征的纹理特征提取、分类和检索方法。论文的工作主要包括： 一、提出一种改进的离散极坐标付立叶变换(DPFT)的快速算法，并在此基础上，提出一种基于DPFT的纹理分类算法。提取纹理的归一化DPFT幅度谱在不同方向和尺度上的信息作为纹理特征，应用于纹理分类，取得了良好的分类效果。本文提出的改进的DPFT算法能够快速有效地完成二维离散极坐标付立叶变换。由付立叶变换在工程应用中的广泛性和极坐标系在处理二维数据旋转和尺度变换上的优越性，该算法也可用于图象配准、图象检索和雷达信号处理等应用领域。 二、提出一种基于提升小波变换的纹理分类算法。利用小波分析的时频局域化特性以及提升小波的本位运算等优点，对纹理图象进行提升小波分解，再对得到的高频子图继续做第二层小波分解，然后以各频率子图的图象熵作为图象的纹理特征进行纹理分类。该算法提取的特征维数较低、计算量较小，具有较强的纹理分类能力。 三、提出一种基于纹理谱直方图和自组织特征映射网络(SOFM)的纹理分类算法。引入象素的八近邻付立叶级数，计算随机选取的纹理区域的纹理谱并量化得到谱直方图，将其作为SOFM网络的特征模式输入并训练网络。网络的输出层对应的不同纹理的类别，从而将不同模式的纹理归为各自的类别。算法对均匀纹理的分类快速有效。 四、提出基于高阶累积量、多通道Gabor滤波以及提升小波变换的纹理检索算法。在基于高阶累积量的方法中，利用图象中象素的窗口邻域具有相关性的特点计算图象的三阶累积量作为纹理的特征。基于多通道Gabor滤波和提升小波变换的方法是利用它们的时频局域化特性，提取纹理的不同尺度和方向上的信息作为纹理的特征。分别用上述算法对纹理图象进行处理，提取相应的纹理特征，对实验纹理库进行检索，并给出详实的实验数据、图示与分析。基于多通道Gabor滤波和基于提升小波的检索算法可以提取纹理图象不同尺度和方向上的纹理信息，因此整体检索效果优于基于高阶累积量的方法。 本文算法的实验纹理库由108类Brodatz纹理图象组成。更多还原

【Abstract】 Texture is one of the essential and important characteristics of imagery. As an important component of image understanding, analysis and recognition, texture analysis is valuable not only in theory, but also in wide-range applications. There are two basic types of methods in texture analysis which are the method based-on feature and the method based-on model. Texture analysis composes of texture description, texture segment, texture classification, texture retrieval, etc. In this paper, some approaches on texture feature extraction for purpose of classification and retrieval are discussed and studied.The major work implemented in this paper is presented as follows:(1) An improved discrete fast polar Fourier transform algorithm (called DPFT for short) is proposed which fast and effectively computes the 2D discrete Fourier transform in polar coordinate systems. And based on it, a new texture classification method in presented that utilizes the information of the DPFT serialized amplitude spectrum in different scales and orientations as texture features, and has good performance in texture classification. Because the Fourier transform is widely applied in engineering and the polar coordinate systems has superiority in data’ rotation and scaling, the improved DPFT algorithm can also be implemented in many fields such as image registration, image retrieval and radar signal processing.(2) A new texture classification algorithm based-on the lifting wavelet transform in suggested utilizing wavelets’ perfect performance in spatial-frequency localization and the lifting wavelets’ in-place operations. It computes the entropy values of the sub-frequency images disposed by the two-layers’ lifting wavelet transform, and has merits of low feature dimensions, low computational cost and good capability.(3) An algorithm for classification of texture image based-on spectrum histogram and self-organized feature mapping network in proposed. It induces the idea of pixel’s 8-neighbor Fourier series, randomly chooses the local region of texture and computes the spectrum of it. After quantizing the spectrum, the spectrum histogram of the local area is extracted and then provided to the SOFM network as a feature vector to train the net. The neurons in the topological output layer correspond to different textures when the training process is finished. Experiments on six samples of Brodatz textures demonstrate the simplicity and efficiency of this algorithm.(4) Approaches for texture retrieval based on high-order cumulants, multi-channel Gabor filters and lifting wavelet transform are suggested separately which extract the texture features to implement the texture queries. In these methods, the approach based-on high-order cumulants utilize the local window’s correlation of image to compute the three-order cumulants as features. And the other two methods are perfect in spatial-frequency localization. In a whole these twospatial-frequency localized methods are more effective than the method based-on high-order cumulants for the superiority in embodying images’ information at different scales and orientations.The texture’ test database in this paper is composed of 108 different kinds of Brodatz textures.更多还原