Biometric Recognition Based On Samll Sample Size Problem

At present, biometric recognition technology is widely used authentication technology. In a monitoring system, gait recognition and face recognition are important biometric identification technology. Face feature is natural and direct and it’s technology is mature; gait feature has the advantage of be difficult to hide and be non-invasive, is promising in the video surveillance. Finger vein recognition is an emerging biometric identification technology; it has some unique advantage compared with other biometrics, such as high-security and anti-interference advantages.It is recognized as the most potential biometric identification technology.The small sample size problem is common in the practical application of the biometric identification technology. Currently, lots of biometric technology are ineffective or can do nothing facing this problem, which is a bottleneck to restrict the development of biometric identification technology. Most biometric identification methods are proposed based on the statistical analysis method, the accuracy of the biometric recognition system depends on the estimated model and the accuracy of the prediction function. In theory, the correct rate of the model and function can be guaranteed under the premise that the number of samples is infinite. As a result, the traditional methods are difficult to solve the small sample size problem in the biometric recognition system.In this paper, we proposed two methods to solve the small sample size problem in the biometric identification:face and gait recognition based on semi-supervised learning, finger vein recognition based on FSS method.The semi-supervised learning methods can effectively solve the small sample size problem in the video surveillance. In our work, we apply the semi-supervised learning method self-training in face recognition system and gait recognition system respectively. Besides, we apply the semi-supervised learning method co-training in the multi-modal face and gait recognition system. To improve the efficiency of the semi-supervised learning method, we use(2D)2PCA method to reduce the dimension of the face image, and use PCA method to reduce the dimension of the gait feature. On one hand, semi-supervised learning methods can make full use of the existing labeled samples; on the other hand, it can take full advantage of the unlabeled samples, so it can improve the accuracy of the training model.The traditional method only extracts little information for the finger vein, and it is poor to solve the small sample size problem in the finger vein recognition system. The FSS (full matching score sequence) method is capable of digging much more information, extracting rich distinctive features, so, it can fully enhance the performance of the finger vein recognition system in the small sample size problem. Our method overcomes the shortcomings of the traditional matching method which only focus on the highest score, considering the value and the sequence of all the matching score, digging the dependencies between different samples and expressing the feature information well for the samples. In addition, the fusing of the LBP and FSS feature method on the score level fusion for finger vein recognition can gain the advantages of each feature extraction method, and comprehensively improve the performance of finger vein recognition system.