【作者】 李海啸；

【导师】 舒华忠；

【作者基本信息】 东南大学 ， 生物医学工程， 2005， 硕士

【摘要】 纹理分割是纹理分析中的难点,其中的关键技术是特征提取,特征提取的好坏直接影响着分割结果的好坏,矩作为一种特征描述子可以有效地提取纹理特征。综合考虑矩在图像描述能力、信息冗余度和噪声敏感性等方面的性能,Zernike矩具有比几何矩、Legendre矩等更优良的性能。但是几何矩和正交矩只能提取图像的全局特征,新出现的小波矩很好的克服了这一缺陷,它可以借助小波变换的多分辨率特性提取图像的局部特征。我们分别将这两种矩应用于提取纹理特征,得到了良好的分割结果。本课题首先提出基于Zernike矩和BP神经网络的纹理分割方法。它的特征提取分两步:首先在图像每个像素周围的小窗口内计算Zernike矩,得到纹理特征的几何模型,然后通过一个非线性变换器将第一步得到的矩值转换成特征向量。特征提取完成,接下来的分类工作交给BP神经网络。比较基于Zernike矩和基于Legendre矩的纹理分割结果,可以看出前者是一种更有效的分割方法。接下来讨论了Zernike矩的阶数及窗口的大小对分割结果的影响,得出结论:1、提高Zernike矩的阶数可以减少分割错误率,但是更高阶矩对噪声敏感,反而使分割错误率上升;2、矩值计算时窗口大小的选取与纹理本身的特点相关,较致密的纹理需要选取较小的窗口,较疏松的纹理需要选取较大的窗口,而非线性变换器的窗口应尽可能的大,以提高分割结果的区域连续性,但同时要考虑到计算效率问题。我们还提出了基于小波矩和BP神经网络的纹理分割方法。首先根据小波变换和旋转不变矩的概念引入了小波矩,主要讨论了两种小波矩——三次B样条矩和Haar矩,并且提出了特征选取的方法,提高了效率和特征性能。通过对差别细微的纹理对的分割结果进行比较,发现三次B样条矩的分割错误率低于Haar矩的,而Haar矩的分割错误率又低于Zernike矩的。更多还原

【Abstract】 The most important technology in texture segmentation is texture features extraction. Moment, as a feature descriptor is employed to extract the texture features efficiently. Considered the aspects of image description, information redundancy and noise sensitivity, Zernike moment is more better than Geometrical moment and Legendre moment. But these kinds of moments can only extract global features from images, Wavelet moment solves this problem. In this paper, Zernike moment and Wavelet moment is used to extract texture feature to accomplish segmentation.First, a texture segmentation algorithm based on Zernike moment and BP neural network is presented, in which the feature extraction is divided into two steps: first, the Zernike moments in small local windows of the image are computed; second, a nonlinear transducer is used to map the moments to texture features and these features are used to construct feature vectors served as input data of the clustering algorithm. Then a BP neural network is employed to perform segmentation. Compared the segmentation result based on Legendre moment, the result based on Zernike moment is better.Then the problem of selecting Zernike moment order and windows size is discussed, and such conclusions are gained: 1. If high Zernike moment order is used, good segmentation result is gotten, but the more higher order is sensitive to noise, thus the segmentation result becomes bad. 2. The first window size is determined by the texture itself character, textures with large texture tokens require large window sizes whereas textures with finer texture tokens require smaller window sizes; The second window is should set as large as possible to enhance the continuity of the segmentation result.We also presented the segmentation algorithm based on Wavelet moment and BP neural network. According to the conception of wavelet transition and moment, we introduced Wavelet moment. In our experiment, we mainly used two kind of Wavelet moment—B Spline moment and Haar moment. Besides these we presented an algorithm of feature selection. Compared the segmentation results of texture pairs which have little difference between them, the results based on B-Spline moment are better than that of Haar moment, than that of Zernike moment.更多还原