An MRF-based Kernel Methods for Nonlinear Feature Extraction

• Main Authors: Po-Wen Chou, 周博文
• Other Authors: Pi-Fuei Hsieh
• Format: Others
• Language: en_US
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/00007884978772069596

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 95 === 　　Feature extraction has been intensively used in pattern recognition for removing redundant features and accelerating data processing. When classes are separated by a nonlinear boundary, the linear feature reduction methods may not as efficiently as nonlinear methods separate classes in a low dimensional feature space. By replacing the inner product with an appropriate positive definite function, the kernel-based nonlinear feature extraction methods implicitly perform a nonlinear mapping of the input data into a high-dimensional feature space and then compute a linear feature extraction method. The radial basis kernel function (RBF) is a commonly seen kernel trick. However, the RBF kernel method may not fulfill feature extraction perfectly in classification problem since its similarity measure is based merely on the Euclidean distance. 　　This study incorporates contextualand class information into the RBF kernel function to improve the performance of kernel-based feature extraction. Samples are considered closer in case of the same class labels or in case that their neighbors are alike in the MRF sense. This leads to an increased discrimination between samples when using kernel function implicitly for a nonlinear mapping of data into a higher-dimensional space. 　　 　　Experiments have been tested on three synthesized dataset and real images. In our experiments, the proposed MRF-based kernel method yielded comparatively higher classification accuracy than the traditional RBF-based kernel method and other linear feature extraction methods. Also, the MRF-based kernel feature extraction, followed by a pixel-wise classifier, out performed the MRF-based contextual classifier.