Feature Level Information Fusion Based On Subspace-and Its Application On Human Identification

With the rapid development in the information, digital technologies, and network of our society, using the traditional methods like password can't meet the needs of security authentication. The corresponding biometrics identification technologies have been the trend on human identification. The unimodality can't meet the practical need for either the identification rate is low, or it can't overcome the defect of themselves. Multibiometrics feature fusion technologies become the hot focus in biometrics identification currently, including the fusion at multi-sensorial, multi-algorithmic, and multimodal level. According to the levels established by international literature, information fusion can occur at the data, feature, score or decision level. And the paper is just completed under the project "Research of Feature level information fusion theory and its application" , funded by the National Natural Science Fund Committee.With the idea of feature level information fusion, multi-algorithmic and multimodal information fusion technologies were developed under the subspace analysis in the paper in order to make identification rate higher and authentication rate lower, including the False Accept Rate (FAR), and the False Reject Rate(FRR). The main works of the paper are as fellows:1. With the fusion feature of the linear Principal Component Analysis (PCA) and the Principal Component Analysis based kernel, its application on the palmprint identification can get higher identification rate. PCA and KPCA can get the linear feature and nonlinear feature respectively, and the best fusion coefficient can be calculated by making the total distance of between-classes largest.2. The dual subspace, i.e. the range space and the null space of within-classes scatter, were analyzed. And a algorithm fusing the discriminant information existing in the dual subspace was given. Linear Discriminat Analysis (LDA) based Fisher principle is classic algorithm under the subspace analysis. However, the small sample size problem results in singularity of the within-classes scatter matrix. Some traditional methods, such as Fisherpalm and Discriminant Common Vector (DCV), are adopted to avoid the singularity of with-classes scatter matrix. But Fisherpalm discards the discriminate information existing in the null space of with-classes scatter matrix, while the DCV only uses the discriminant information existing in the null space of with-classes scatter matrix. Because all of the discriminant information is not utilized, we apply the dual space analysis with palmprint recognition in the paper, and it can outperform the DCV and Fisherpalm methods.3. An algorithm, Fisher discriminant analysis based kernel (KFDA), was proposed under the relation measurement framework, which was used to fuse the face and palmprint feature. The experiment results show that the method can achieve a higher identification rate and a lower authentication rate than the uniomodality.