Dissertation > Excellent graduate degree dissertation topics show

A Research on Feature Selection and Fusion in Palmprint Recognition

Author: YuanCaiMao
Tutor: SunDongMei
School: Beijing Jiaotong University
Course: Signal and Information Processing
Keywords: Feature selection Feature Fusion Composition theory Eigenvectors Correlation
CLC: TP391.41
Type: Master's thesis
Year: 2010
Downloads: 134
Quote: 1
Read: Download Dissertation

Abstract

With the social progress and the development of information technology, information security is increasingly conventional password, the password has been unable to meet the people's demand for security authentication. In this case, the biometric technology for its easy-to-use, highly reliable characteristics of importance for the people and extensive research. The palmprint identification technology has many advantages, biometric identification technology to become a high-profile widespread concern and discussion, and has broad market prospects. Many scholars have done a lot of research on palmprint identification proposed palmprint identification algorithms. However, many problems still exist palmprint recognition. Very important issue is the use of a single feature identifying the presence of palmprint recognition rate is limited, and poor robustness shortcomings. In order to overcome the defects of single feature recognition, feature fusion method is introduced into the the palmprint identification field. Fusion methods at all levels, feature-level fusion of both overcome the large quantities of raw data, unstable characteristics, but also to maximize retains the original information, the most likely to bring the best recognition effect, hence the research focus in recent years. The feature fusion mainly refers to the fusion of feature vectors, palmprint feature level fusion eigenvectors paper mainly discusses two issues: (1) With the increase in the number of categories to be identified, the dimension of the feature vector constant case, to identify the effect will gradually decline, but also increase the feature vector dimension has not been brought recognition rate rise. This is because the feature extraction process may be introduced to the classification unfavorable \To solve this problem, we propose a KL distance measure feature selection framework for the evaluation function to improve recognition performance by removing the weak component. Experiments show that after feature selection, the recognition performance of the feature vector has been improved, and the dimensions also been reduced. Other key issues that need to be addressed in feature selection is the estimated probability density of the random variable. Because every time you need to estimate the random variables many another value is quite different and can not be described by the statistics of the distribution model, this paper proposes a method based on composition theory of probability density, according to the characteristics of the data itself the probability density of the model and parameters, it is possible to better adapt the requirements. The experiment proved the feasibility of such a probability density estimation method. Another problem is that in (2) feature fusion eigenvectors related issues. Converged eigenvectors of the correlation between each other the smaller, the more it can bring to enhance recognition performance, or complementary, amounting to less than feature fusion purpose. This article using the Spearman rank correlation coefficient of correlation between the eigenvectors metrics, and experimental study of the relationship between the strength of correlation with the final recognition effect. Experiments, respectively, using the different characteristics of integration programs, each program in the correlation between the eigenvectors are quite different. The experiments showed that the correlation characteristics of fusion can bring a higher recognition rate, and the feature selection frame in the previously discussed to be fusion feature vector after the application, the enhancement of the recognition rate will also have a positive effect.

Related Dissertations

CLC: > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device
© 2012 www.DissertationTopic.Net Mobile