Study On The Crucial Vein Recognition Technology

Hand vein Information-based identification technology has become one of the most promising biometrical feature recognition, and the key is that the vein feature of one person is unique and it has good performance of anti-interference. It performs better than the fingerprint feature and face feature in personal identification. The paper focuses on the main procedures of vein recognition including vein image capture, vein image enhancement, vein feature extraction and representation, vein recognition and so on.(1)To improve the situation that lacking of open vein image databases for research, the paper publishes a set of capture device with circle structured near-infrared light design, and it also imports the design of the double-layer vein image quality estimation system to realize the PID control on the light system. With the improving design on the device, the paper publishes a set of high-quality vein image database with independent intellectual property rights.(2)To solve the problem of capturing vein image in poor quality affect the later process, the paper conducts deep research on image enhancement technology and design three state-of-the-art enhancement methods for the problem. The self-designed enhancement methods are respectively Multi-Scale Top-hat Transformation(MSTHT), Local Gray-level Information-based methods( LGLIT), and super-resolution-based reconstruction and enhancement(SRRT). And it also imports quality evaluation system to compare the quality of before-enhancement and after-enhancement to demonstrate the advantage of the proposed enhancement methods.(3)For the purpose of decrease the negative effect bringing by the non-vein information of the image on the later processing procedure, the paper brings about a new kind of ROI extraction method based on effective area location distribution theory, and then imports a series of pre-process methods including vein image segmentation, noise removal and normalization on both grayscale and size to get the final high-quality vein image skeleton. And in the feature extraction and representation section, the paper firstly introduces the data-dimension reduction and feature mapping theory, and respectively conduct feature description experiment with PCA, 2DPCA, and self-designed(2D)2FPCA methods on the self-published vein databases and get relatively great recognition rate with the highest rate of 98% by the self-designed(2D)2FPCA.(4)To solve the problem of negative influence by the uncontrollable factors such as rotation change, size change, light condition change during the capturing section on the later recognition, the paper introduces the moment theory in mathematics to get the geometrical-invariant feature representation. The paper design three important experiments in this section: In the first experiment, the paper makes some manned changes on the image including rotation, size change and so on, and then designs methods to get the Zernike moment-based feature descriptors to get the final feature matrix to prove the anti- geometrical effect; In the second experiment, the paper design the comparison one to decide which moment are the best one to be chosen for feature representation; In the last one, the paper respectively adopts Hu and PCET moment for final feature representation and classification, and get good result, in which the calculating speed are up to 0.04 s and the recognition rate up to 98.3%.(5)To solve the problem that the traditional recognition methods are manned-design feature and classifiers, and it needs repeated attempts to adjust the parameters for higher recognition rate, the paper introduces the feature learning theory which refers to combining low-hierarchy feature to get the high-hierarchy feature so that the distributed feature-based deeps structured network. Firstly, the paper designs the k-means clustering and SIFT feature combination method to demonstrate the necessary of importing deep learning theory and then the CNN network based on RNN for classifying part is introduced(RCNN), which is a high-level deep network structure monitoring the working theory of visual perception system, and it gets incredibly good effect in recognition.