Research On Multimodal Biometrics

Multimodal biometrics provides effective approach for person recognition in the modern world, so it has been paid more and more attentions.Feature level fusion and score level fusion system are researched in the thesis. Face, iris and palmprint are employed in our research. Experiments are done using the validation and open test set.The following contributions are made in the thesis.1. To overcome the shortcomings of CCA application in feature fusion, a novel supervised learning method, termed ad Enhanced Corelation Analysis (ECA) is proposed. With the help of kernel trick, the kernelized ECA (KECA) is further proposed to tackle the linenearly inseperable cases. The class information is employed in the ECA. Also, the ECA can overcome the difficulties due to the loss of samples in real applications. Experiments show that the ECA and KECA outperform other fearure fusion methods and KECA outperforms ECA.2. A novel high resolution NIR face and irises image device is designed. A noisy multimodal database is founded using the device. Two algorithms are proposed based on the database. 1) A novel joint iris code based both iris feaure fusion algorithm is proposed. Experiments are done using the database founded in this thesis and give promising results. 2) A composed fusion structure is proposed for fusion of face and irises. The feature fusion result of both irises is fused with face in score level.3. A least squares method based score fusion algorithm (LSMSF) is proposed. The parameters of the fusion function are estimated using the least squares method (LSM). The fusion function could be power series function, multivariate polynomial function and the reduced multivariate polynomial function. Experiments show LSMSF outperforms other score level fusion method under various conditions.