A Research On Feature Selection And Fusion In Palmprint Recognition

Along with social progress and development of information technology, people are increasingly demanding information security. The traditional password and smart cards have been unable to meet the people's safety certification requirements. In this case, biometric technology, with its easy to use and highly reliable, has been the focus of extensive research. For having many advantages, the palmprint identification technology has become a high-profile biometrics, and been widespread concerned and discussed. The market prospect for palmprint identification is broad.Many scholars have been doing a lot of research on palmprint recognition, and lots of algorithms have been proposed. However, many problems still exist. One very important issue is when using single feature vector for identification, the recognition rate is limited by characteristics of the feature vector, and less robust. To overcome the shortcomings of single feature vector recognition, feature fusion is introduced into the field of palmprint recognition. In all fusion ways, feature level fusion (FLF) has been emerging an effective way to improve system performance. Feature vectors extracted from palmprint and fusion strategy are key factors in this procedure. This paper discusses issues about the two aspects respectively.(1) The performance will decline considerably as the number of categories increase. This is determined by the intrinsic characteristics of the selected feature vectors. To overcome this problem, we propose a feature selection method based on K-L distance between feature vectors of each person and the population in each dimension. Experimental results show that the performance of the system keeps a stable level when the categories increase. A key issue in the feature selection is how to estimate the probability density of a random variable. As there are many random variables to estimate each time and the variables are quite different to each other, the traditional methods are no longer available. This paper presents a approach based on Theory of Composition, which estimates the probability model and parameters according to the characteristics of sample data, to meet the need better. Experimental results show the feasibility of this method.(2) Whether the fusion of different feature vectors can bring improved performance, it depends on which feature vectors should be chosen. For FLF, feature vectors should have weak correlation between each other. In this paper, we studied the correlation based on Spearman correlations coefficient. If the matrix reflects a strong correlation, this fusion strategy should be a failure. We test some fusion strategies in experiment, and the correlation between feature vectors in each fusion strategy is different from others. Experimental results show that strategy in which the feature vectors with weak correlation brings an improved performance. Besides, the feature selection framework discussed above plays a positive role in improving the recognition rate.