Shilei Liu

DeepVein: Novel finger vein verification methods based on Deep Convolutional Neural Networks

Finger vein verification is using vein patterns to verify a persons identity, which is widely used in various fields. In practice, the method for verification is the most important part of a biometric system, which determines the reliability of the system. In this paper, we propose methods called DeepVein for finger vein verification based on deep convolutional neural networks and conduct experiments to evaluate our methods. The experimental results show that our proposed methods can achieve state-of-the-art performance in accuracy. In addition, we present how the amount of data for training affects the accuracy in the test datasets.

BMDT: An optimized method for Biometric Menagerie Detection

Biometric menagerie is an important phenomenon in biometric systems, which focuses on distinguishing the minority of people who perform poorly and cause the majority of the errors (FAR and FRR). It can help to evaluate biometric systems and improve their performance by analyzing the animal like users. The fundamental step of this theory is the detection of animals. If the detection is not accurate, it may lead to potential problems. However, the current theories carried out by Doddington et al. (1998) and Yager (2008) both neglected the threshold in biometric systems when detecting animals, which might reduce the accuracy of animal detection. To verify this conjecture, we apply the above two theories to detect the existence of animals on a special finger vein database PFVD - Perfect Finger Vein Database. The characteristic of PFVD is that its accuracy is 100%, indicating zero FAR and zero FRR. From the intuitive point of view, there should exist no goat, lamb or wolf in Doddingtons menagerie, and no worm, chameleon or phantom in Yagers menagerie. However, the experiments show the negative results, implying that the current theories are not perfect on animal detection. This paper analyzes the two theories and proposes BMDT - Biometric Menagerie Detection with Threshold, an optimized method based on Yager. By taking threshold into account, BMDT makes a significant improvement on the accuracy of animal detection, compared to the current theory. We apply BMDT on PFVD, and the results show that the falsely detected animals by Yagers method are removed. In addition, we evaluate BMDT in 3 more general cases, proving the advantage of the proposed method.

A study on the individuality of finger vein based on statistical analysis

Biometric recognition requires that the biometric characteristics used for the verification should be unique among individuals. The purpose of this study is to verify the individuality of finger vein, a new trait with superiority on accuracy and security. To carry out the research, we construct a large-scale finger vein database consisting of 710,399 images from 363,703 fingers. We adopt a score level fusion strategy to reduce the negative impact of algorithm deficiencies. We also design a distributed computing system for more than 83 billion impostor comparisons. The experimental results demonstrate that finger vein is sufficiently unique to distinguish one person from another in such scale.

A Study of Temporal Stability on Finger-Vein Recognition Accuracy Using a Steady-State Model

Stability has been one of the most fundamental premises in biometric recognition field. In the last few years, a few achievements have been made on proving this theoretical premises concerning fingerprints, palm prints, iris, face, etc. However, none of related academic results have been published on finger-vein recognition so far. In this paper, we try to study on the stability of finger-vein within a designed timespan (four years). In order to achieve this goal, a proper database for stability was collected with all external influences of finger-vein features (acquiring hardware, user behavior and circumstance situation) eliminated. Then, for the first time, we proposed a steady-state model of finger-vein features indicating that each specific finger owns a stable steady-state which all its finger-vein images would properly converging to, regardless of time. Experiments have been conducted on our 5-year/200,000-finger data set. And results from both genuine match and imposter match demonstrate that the model is well supported. This steady-state model is generic, hence providing a common method on how to evaluate the stability of other types of biometric features.

A Decision Reliability Ratio Based Fusion Scheme for Biometric Verification

Unimodal biometric verification has developed a lot and become more accurate, but there is still not a perfect algorithm. In the meantime, cases exist where unimodal verification system could not meet the requirements in practical use. It is proved that algorithms with the same overall accuracy may have different misclassified patterns. We could make use of this complementation to fuse individual algorithms together for more precise result. According to our observation, algorithm has different confidence on its decisions but this is seldom considered in fusion methods. Our work focuses on this confidence. We first define decision reliability ratio to quantify this confidence, and then propose the Maximum Decision Reliability Ratio (MDRR) fusion scheme incorporating Weighted Voting. Two experiments conducted on different datasets prove the effectiveness of the method. One is to fuse 4 finger vein verification algorithms on a set of 1000 fingers and 5 images per finger. The other experiment fuse the multimodal set in NIST-BSSR1. Experiment results show the fusion method could largely improve verification accuracy, from 91.29% to 99.81%. It also shows the MDRR outperforms the commonly used fusion methods as Voting, Weighted Voting, Weighted Sum or even the theoretically optimal method Likelihood Ratio-based fusion.

Parameter adjustment of finger vein recognition algorithms

Finger vein recognition is a biometric method utilizing the vein patterns inside ones fingers for personal identification. Recognition algorithm is the key part of a finger vein recognition system, dominating the system performance. There are usually a lot of parameters in algorithms, and different values of the parameters could lead to different system performance so that it is essential to set a proper value for each parameter in practice. In this paper, we conduct a set of experiments to study how the parameters influence the performance measured by equal error rate. We have made two observations from the results: 1.When an algorithm is applied on a dataset, the performance differs a lot as the parameter value changes even in a small range; 2.When an algorithm is applied on different datasets, the performance differs a lot, in other words, the optimized parameter value combination that maximizes the system performance differs significantly. These two observations reveal the importance of parameter adjustment in finger vein recognition. So this paper proposes two solutions: search algorithm and estimation by subset, which are fast, accurate and scalable methods to find the best parameters. The experiment results prove the effectiveness of our methods.

Which finger is the best for finger vein recognition

Finger vein recognition is a biometric method utilizing the vein patterns inside ones fingers for personal identification. Every user has 10 fingers in total, among which the index fingers and middle fingers from left and right hands are the most common used fingers in finger vein recognition systems, for their suitable length, width and flexibility. No evidence shows any significant relationship between ones different fingers, so they are usually treated as different classes in recognition. However, what is the difference between different fingers? How do they perform in finger vein recognition systems? Some researchers believe that index fingers perform better than middle fingers, while the others hold contrary opinions. There is not a consensus on this topic. In this paper, we conduct a set of experiments on different fingers, with different recognition algorithms and different databases. The result shows that considering the equal error rate and DET curve, the performance of different fingers is quite different in different algorithms and database. Therefore the performance of fingers is algorithm and database dependent. Based on the finding, we propose a method to improve the performance of finger vein recognition systems. The rationale of our method is that we assume every user has a “best” finger in certain scenarios, and the “best” finger for each user may be different. Then we design an algorithm to find their best fingers and suggest them to use it. Evaluation results show that if the users use their “best” fingers as our method suggests, the EER of the system decreases up to 60%. This means the proposed method can improve the performance of finger vein recognition system at a significant level without any change on the algorithm.

Performance of Finger-Vein Features as a Human Health Indicator

Biometric features of humans, which include behavioral features and biological features, have two main popular application areas: 1. Biomedical application as in human disease diagnosis and health monitoring; 2. Security application as in identity authentication.