Two dimensional subspace analysis for pattern recognition

Abstract

This dissertation proposes four novel frameworks for image pattern recognition. The first framework is based on discriminant analysis of principal components framework or Fisherface. Since 2DPCA is more suitable for face representation than face recognition, thus 2DLDA is proposed for improving performance in recognition task. The second framework is based on Class-Specific Subspace (CSS) method. By applying CSS over 2DPCA, the class information is introduced to an unsupervised method. Each subspace of CSS learned from only the training samples within their own class. In this way, the CSS representation can provide a minimum reconstruction error, which it can be used to classify the input data. The third framework is based on our proposed method by generalizing form of image covariance matrix, called image cross-covariance matrix. Comparing to the covariance matrix of PCA, the image covariance matrix discards some information. In practice, this disregard information may possibly be useful for discrimination. The image cross-covariance matrices are formulated by two variables, the original image and its shifted version. By our shifting algorithm, many image cross-covariance matrices are formulated to use another information in which discarded by the image covariance matrix. The fourth framework is to apply the random subspace method to 2DPCA. Normally, the feature of 2DPCA is a matrix. In the row direction, the number of the columns of this matrix is affected by the number of selected eigenvalues of the image covariance matrix while the number of selected eigenvalues is not influenced in the column direction. Thus, the number of the rows is still equal to the height of original image and the random subspace method can be apply in the column direction. The random subspaces are constructed by randomly selecting a number of rows of the original feature matrix. The multiple classifiers are constructed in these random subspaces of the data feature space. These classifiers are usually combined by simple majority voting in the final decision rule. Moreover, this framework is also applied to DiaPCA that used to select the subspaces which constructed by the third framework. Experimental results on well-known image databases, both of face and non-face databases, show that all of our proposed techniques clearly gives a higher recognition accuracy than the conventional algorithm.