Finger Vein Recognition Based On Two-Dimensional Non-negative Matrix Factorization

The rapid development of technology has brought great convenience to the social life,and the universal application of Internet makes information exchange quickly and widely. At the same time, information security and privacy become particularly important, and people have higher requirements for identity authentication. Biometrics, the new generation of identification technology which uses the inherent characteristics (physical or behavioral), is now widely used due to its high security. Currently, finger vein recognition technology comes up as the second generation of biometric identification technology. Because it can overcome the shortcomings of other hand features such as fingerprint, hand geometry and so on, so finger vein recognition has already attracted attention of many researchers.This article focuses on the region of interest acquisition and feature extraction algorithms, and puts forward a complete program of finger vein recognition.First, change the finger-vein image from RGB to gray and do normalization operations to reduce storage space. The full finger regions are extracted by using Otsu threshold method based on the column direction and getting rid of the small block noise through connected domain area. In order to reduce the influence of nonlinear translation and rotation when using finger vein images obtained with non-contact devices, we propose a region of interest extraction method. First through fitting finger midline to identify the deflection angle and rotation correction; then, using the diameter of the fingertip contour to locate, find the finger specified width maximum inscribed rectangle and finally size normalization. Comparative experiments are done between complete finger-vein region and finger-vein region of interest,the results show this algorithm can accurately extract the region of interest, and effectively improve the system recognition performance.Subspace method is commonly used in feature extraction, data dimensionality reduction via matrix factorization to find the algebraic properties of the data. Non-negative matrix factorization (NMF) is a new decomposition, adding non-negative constraint in the decomposition process and making the decomposition results also satisfy non-negative.Therefore, this paper studies using NMF to extract local characteristics of the finger vein. The paper first studies the traditional non-negative matrix factorization theory, and then it researches two improved sparse algorithms in analyzing finger vein characteristics extraction.The article contrasts these three algorithms in feature-based images, convergence rate,verification accuracy and identification accuracy, and then draws conclusion that nonnegative matrix factorization with sparse constraints (NMFSC) algorithm works best.To overcome the problem of long training time, the paper continues to research on the two-dimensional nonnegative matrix factorization (2DNMF) which contains twice nonnegative matrix factorization. The decomposition takes the original two-dimensional images as objects to retain the integrity of structural information, so the identification rate is proven to be increased. Furthermore, without quantization process, the data dimension can be much lower, so the complexity of the calculation and the training time are greatly reduced.Then, the paper improves the 2DNMF and put forward parallel-2DNMF algorithm in which the two nonnegative matrix factorization processes are parallel and (2D)2 NMF algorithm which is based on the diagonalization of original images and the orthogonalization of basis matrix. Experiments show that the speed of convergence,the accuracy of verification and the precision of identification are superior to the NMF and its improved methods.Finally, this paper introduces the entire finger vein recognition system to verify the performance of the proposed algorithms. Experimental results show that the proposed algorithms not only ensure fast convergence, but also significantly shorten the training time,besides, the recognition rate is also increased.