Research Of Vein Recognition Based On Layered Structure Of SVM And Fusion Identification

Biometric identification is a kind of authentication technology. It is based on the inherent physical characteristics or the behavioral characteristics of the human body. Owing to its better security, reliability and validity, many scholars are paying more and more attention to it, and began to affect all areas of our social life. After decades of research, people have reached remarkable achievements in the fields of face, fingerprints, iris, hand shape, palm, voice, signature gait, and venous etc.Vein recognition is a kind of biological recognition. During to the non-contact, simple acquisition and uniqueness characteristics, it has become the focus in the field of high technology and application of the market. It must have a good prospect for development.This article aims to realize the image classification and matching of the palm-dorsal vein and finish the authentication. It studies the information contained in veins, stratified feature extraction, cascade SVM classification, information fusion decision-making and hierarchical feature veins matching. The research is under the assumption of image preprocessing (sampling, de-noising and smoothing, etc).In the first step, this article finished the analysis of global image characters, image peak signal-to-noise ratio and image information entropy. After that, it studies the image features descriptor which is based on the shape. The descriptor has the properties of translation, rotation and extends invariants. Then, this paper gets the results that Hu invariant moments and affine invariant moments are based on the shape of the border region while border invariant moments are based on shape boundary. Finally, it refines the whole image geometric structure and finishes the extraction and analysis of texture structure.In the second step, this paper uses cascade SVM classifier to classify the characters from each layer, and make use of information fusion method to get the final decision results. Firstly, classify the characters in each layer by SVM method separately and form three parallel structures. Then, compute the reliability of sample classification, namely the right values reliability. Finally, get the results by weighted function fusing.In the final step, according to the characteristics of the three layers extraction, compare the classified samples with signal-to-noise threshold then finish intersection, endpoint and nearest matching on the condition of threshold requirements to achieve the purpose of layered matching after eliminating the unexpected resultThe experiment results show that the methods can achieve the purpose of recognition and can give some references for future researches.