【作者】 王雨春；

【导师】 王展青；

【作者基本信息】 武汉理工大学 ， 应用数学， 2006， 硕士

【摘要】 典型相关分析是多元统计分析的一个重要研究课题。它借助主成分的思想，用少数几对综合变量来反映两组变量间的线性相关性质。目前它已经在众多领域的相关分析和预测分析中得到广泛应用。但将其应用于数字图像处理，目前相关的研究还很少。 本文首先深入研究典型相关分析的各种线性算法和非线性算法，并通过实验数据的比较说明了它们各自的特点。然后主要从理论和实验两个层面论证了典型相关应有于数字图像处理方面的可行性和优越性。同时通过实验数据说明了典型相关的线性判别和非线性判别的分类能力和特点。 本文提出一种基于图像邻域信息的分割方法。首先，根据象素点邻域信息得到高维特征向量；然后采用典型相关分析(CCA)改进线性判别分析(LDA)中的变换矩阵，使得特征向量的降维具有自适应性；最后用最近邻法对降维后的特征向量进行分类，从而实现了图像的分割。试验中，选取人脸图像分割来验证该方法，结果显示出其具有良好的分类效果。在图像匹配方面，基于典型相关分析的思想，对相邻图像进行匹配拼接，确定出合适的拼接点，并选用不同场景的图像进行试验，所得结果令人满意。更多还原

【Abstract】 Canonical Correlation Analysis (CCA) is a important research subject of multivariate statistical analysis. With the idea of principal component analysis, CCA reflects the linear correlation between two sets of variables with a few basis vectors. Now, the method has been applied to many fields for correlation analysis and forecast analysis, But there is few relevant reference about applying it to the digital image processing.Firstly the linear algorithms and the nonlinear algorithms of CCA are presented in this paper. Secondly, some experiments have been designed to summarize and compare the characteristics of various kinds of algorithms. Their feasibility and validity in digital image processing is confirmed by the experiments and theory. Finally, to apply the method in multiple classifiers, the combination of multiple classifiers based on belief value is performed.An image segmentation algorithm based on contextual information is proposed in this paper. Firstly, the high-dimensional feature vectors of pixels are extracted through the contextual information. Then, the LDA transformation matrix is improved with CCA, which makes the reduction of dimensions adaptive. Finally, the labels of low-dimensional vectors are obtained using 1 nearest neighbor classifier. The experiments were conducted on face pictures; the result shows that the pixel classification is excellent.更多还原