Research On The Key Problems Of Finger Vein Recognition

Finger vein recognition is the technique to identify a person by her or his finger vein image, which is captured by a finger vein image capturing device. The imaging principle is that, near-infrared light can be absorbed intensively by the hemoglobin in the blood of vein, but transmits other tissues of finger easily,therefore vein pattern in finger will be captured as shadows. Finger vein has some advantages over other biometrics:(1) non-contact: captured images are not affected by the surface of the finger, which is convenient and acceptable for the users;(2) live body recognition: the image is captured only in the condition of the live body;(3) high security: finger vein is internal feature which is difficult to forge;(4) small device size: small size is more convenient for application. Finger vein recognition has been widely applied in some fields such as bank, ATM, car and so on. Due to its advantages, finger vein has been a promising biometrics for identity verification.Finger vein recognition includes four steps: image capturing, image preprocessing, feature extracting and matching. Feature extraction is the key step, which may seriously affect the performance of the recognition systems. Although most of vein feature can achieve satisfactory performance to a certain degree, there are still some problems to be solved. For example, the feature can be further improved (feature contains redundant information or noise and ignors the difference of individuals).Most of features are sensitive to the variation of the capture conditions. Besides, multiple biometrics fusion can improve the performance of recognition system significantly, and how to employ the information of multiple biometrics effectively is also a key problem to be solved.This thesis focuses on three problems:inefficiency of vein features, sensitivity of features to the variation of the capturing condition and inefficiency of multiple biometrics fusion.The main contributions can be summarized as follows:(1)To solve the problem of inefficiency of the features, finger vein recognition based on personalized feature selection is proposed. Fusion feature, called Pyramid Histograms of Gray, Texture and Orientation Gradients (PHGTOG) is firstly proposed. And then, a personalized feature selection method based on LASSO is developed. PHGTOG is extracted by combining the histogram of gray, histogram of LBP, histogram of Orientation Gradients with spatial pyramid representation. Therefore, PHGTOG can reflect the information such as intensity, texture and the shape of the finger vein. In addition, spatial pyramid representation makes PHGTOG contain global and local information of the finger vein. The feature obtained by the proposed feature selection method can reduce the redundant information or noise of PHGTOG, and what is more, it considers the difference of the individuals, and retains the personalized discriminative information of PHGTOG for each individual, which is more effective.(2)To solve the problem of the sensitivity of features to the variation of capturing conditions, finger vein recognition based on the hyperinformation feature (HIF) is proposed. The concept of base attribute is firstly defined, which is inspired by the semanitic attribute in computer vision, and then, HIF extraction framework is proposed, as a case study, a simple but effective feature extraction method based on K-means and sparse learning is designed in HIF extraction framework. HIF is constructed by some base attributes, and these base attributes can represent certain characteristics of the finger vein image captured in certain conditions.Compared with the traditional feature which only represents the characteristics (such as texture or shape) of finger vein in a single view, HIF can describe more information which can reflect the characteristics of the finger vein images captured in certain condition from multiple views. The abundant discriminative information makes HIF more robust to the variation of capturing conditions.(3)To solve the problem of the the inefficiency of multiple biometrics fusion, personalized fusion of the finer vein and finger contour based on Classification Confidence Score is proposed. Considering the convenience of the image capturing and the complementarity of the biometrics, finger vein and finger contour are used for fusion on the score level. For most multiple biometrics fusion on score level,two problems are ignored:(1) discriminative information contained in score is insufficient (2) the fusion does not consider the difference of the individuals. The two problems make multimodal biometrics system less effective.To improve the fusion efficiency, the classification confidence score is firstly proposed. The classification confidence score can be obtained by calculating the distance between the instance and the classification hyperplane .And then, SVM is employed for learning the personalized fusion weight for each individual. Compared with the traditional socre, classification confidence score contains the class information, which may provide more effective information for the fusion. Personalized fusion weight can represent the difference of individuals, and further improves the performance of multimodal biometrics system.This thesis analyzes there problems of finger vein recognition: inefficiency of features, the sensitivity of features to the variation of capturing conditions and the inefficiency of multiple biometrics fusion, the solutions to these problems are proposed, i.e, finger vein recognition based on personalized feature selection, finger vein recognition based on the hyperinformation feature and personalized fusion of the finer vein and finger contour based on Classification Confidence Score respectively. The experimental results on our finger vein databases demonstrate the efficiency of the proposed methods. The development of this study not only extends the research field of the personalized biometric recognition, but also enriches the techniques of feature extraction in finger vein recognition, and futher improves the performance and the robustness of the finger vein recognition systems.