Hand Vein Recognition Based on the Fusion of Multi-classification
|School||North China University of|
|Course||Signal and Information Processing|
|Keywords||hand vein recognition texture feature SIFT feature multiple classifiertemplates fusing multiple kernel learning|
Biometric authentication systems based on hand vein patterns had gained considerable attention of researchers in recent years because hand vein patterns were unique, universal, reliable and immune to the imposter attacks. Anatomically, vein patterns were distinctive to individuals, and they did not change with time except by the surgical intervention. In addition, the vein vessels were underneath the skin and surrounded by tissues, which resulted in their partial visibility to the naked eyes. All these special properties of hand vein patterns made it a more stable biometric feature for authentication and a more suitable physiological feature for individual identification.This dissertation introduced the whole process of the hand vein recognition and discussed the every component namely:hand vein recognition theory, capture device, preprocessing, feature extraction and classification design. In the thesis, the main goal was discussing the fusion of multiple classifications such as the theory of fusion, fusion method and the result of every fused classification. The details were as follows:(1) The thesis summarized the capture device and preprocessing for hand vein images, analyzed the principle of the imaging and the structure of the capture device, delimited the parameters of the components based on the quality of the hand vein images, build the hand vein images database under the own copyright. The suitable methods were used which based on the property of hand vein images, such as ROI extraction based on the centroid of the images, the contrast enhancement based on CLAHE and the Winner method to reduce the amount of noise.(2) In the thesis, multiple features were extracted, which full describe the hand dorsal vein. The hand vein pattern was the texture information in the image and based on that the partition local binary patterns(PLBP) were used and some patterns were selected to be the final feature through the optimized. In the binarized hand vein images, the scale-invariant feature transform(SIFT) was used to form the feature keypoints. Two types of hand vein feature represented the identification of people.(3) This dissertation introduced the classifier design based on the PLBP and SIFT feature. Corresponding SIFT feature, the fused classifier based on the multiple classifier templates was proposed and the matching method was optimized, some wrong matches were removed. In multiple classifier templates, useful keypoints were maintained and the redundant information was eliminated which improved the recognition rate and calculation efficiency. About the PLBP feature, the multiple kernel learning based on the Radial basis function was proposed because of the isomerism of the feature space, in which the different train sets were in use of the different fusion method, as a result the improvement of recognition rate was implement. Finally, the two classification results were allocated the different weights to present the fusion of the decision level. The experiment results showed a better classification performance obviously.