Search

Two Dimensional Linear Discrimination Analysis (2DLDA) is an effective feature extraction approach for face recognition, which manipulates on the two dimensional image matrices directly. However, some between-class distances in the projected space are too small and this may bring large error classification rates. In this paper we propose a new 2DLDA-based approach that can overcome such drawback in the 2DLDA. The proposed approach redefines the between-class scatter matrix by putting a weighting function based on the between-class distances, and this will balance the between-class distances in the projected space iteratively. In order to design an effective weighting function, the between-class distances are calculated and then used to iteratively change the between-class scatter matrix. Experimental results showed that the proposed approach can improve the recognition rates on the ORL database, the Yale database and the YaleB database in comparison with other 2DLDA variants. [ABSTRACT FROM AUTHOR]

This paper investigates face image enhancement based on the principal component analysis (PCA). We first construct two types of training samples: one consists of some high-resolution face images, and the other includes the low resolution images obtained via smoothed and down-sampling process from the first set. These two corresponding sets form two different image spaces with different resolutions. Second, utilizing the PCA, we obtain two eigenvector sets which form the vector basis for the high resolution space and the low resolution space, and a unique relationship between them is revealed. We propose the algorithm as follows: first project the low resolution inquiry image onto the low resolution image space and produce a coefficient vector, then a super-resolution image is reconstructed via utilizing the basis vector of high-resolution image space with the obtained coefficients. This method improves the visual effect significantly; the corresponding PSNR is much larger than other existing methods. [ABSTRACT FROM AUTHOR]

In this paper, we aim to develop a new classifier for increasing the worst case performance for individual person. Technically, we adopt the idea from LDA and improve the worst recognition performance for individuals. This is achieved via introducing different weighting coefficients in LDA optimization process for obtaining the projection matrix. By increasing the weighting coefficients associated with the smallest between-class distance, the pair of classes with the nearest distance can exert more powerful influence on the optimization process in derivation of the projection matrix. The algorithm is tested on the Extended YaleB dataset and the ORL dataset. [ABSTRACT FROM AUTHOR]

In this paper, we will revisit the problem of generalized low rank approximation of matrices (GLRM) recently proposed in [1]. The GLRM problem has been investigated by many researchers in the last two years due to its broad engineering applications. However, it is known that there may be no closed-form solution for this problem and the convergence for the proposed iterative algorithm Algorithm GLRAM is actually still open. Motivated by these observations, we will derive a necessary condition for the GLRAM problem via manifold theory in this paper. Further, the relationship between this derived necessary condition and the Algorithm GLRAM is investigated, which will provide some insights for the possible complete analysis of the convergence problem in the future. Finally, some illustrative examples are presented to show different cases for convergence and divergence with different initial conditions. [ABSTRACT FROM AUTHOR]

Feature selection is an important data preprocessing step in data mining and pattern recognition. Many algorithms have been proposed in the past for simple patterns that can be characterised by a single feature vector. Unfortunately, these algorithms are hardly applicable to what are referred as complex patterns that have to be described by a finite set of feature vectors. This paper addresses the problem of feature selection for the complex patterns. First, we formulated the calculation of mutual information for complex patterns based on Gaussian mixture model. A hybrid feature selection algorithm is then proposed based on the formulated mutual information calculation (filter) and Baysian classification (wrapper). Experimental results on XM2VTS speaker recognition database have not only verified the performance of the proposed algorithm, but also demonstrated that traditional feature selection algorithms designed for simple patterns would perform poorly for complex patterns. [ABSTRACT FROM AUTHOR]

In this paper, we propose a new machine learning approach based on AFS (Axiomatic Fuzzy Sets) fuzzy logic, in attempt to provide a better model with interpretability. First, we will concisely present the AFS theory. Second, we will propose new membership functions for fuzzy sets and their logic operations. Third, we will design a new machine learning algorithm based on the new membership functions and their logic operations. This algorithm has two advantages. One is that it can mimic the human reasoning comprehensively and offers a far more flexible and effective means for the study of large-scale intelligent systems. Another is its simplicity in implementation and mathematical beauty in fuzzy theory. Finally, a credit data example is used to illustrate its effectiveness. [ABSTRACT FROM AUTHOR]

Inspired by the human immune system, and in particular the negative selection algorithm, we propose a learning mechanism that enables the detection of abnormal activities. Three types of detectors for detecting abnormal activity are developed using negative selection. Tracks gathered by people΄s movements in a room are used for experimentation and results have shown that the classifier is able to discriminate abnormal from normal activities in terms of both trajectory and time spent at a location. [ABSTRACT FROM AUTHOR]