Research On Fingerprint Identification Method Based On Semi-Supervised Learning

Fingerprint identification technology, as a means of determining the identity by computers, is one of the most widely biometric recognition technologies so far. In the process of sampling the fingerprints, fingerprint image captured always include some variations, called intra-class variations, due to various uncontrolled like scratching, drying, etc. However, those fingerprint intra-class variations may affect the performance of the fingerprint recognition system. In order to adapt to fingerprint variations, this paper makes some related research.Firstly, we describe the fingerprint template update methods and replacement methods applying to the field of fingerprint recognition. Among them, template update methods are mainly to choose the representative fingerprint template to better adapt to the fingerprint intra-class variations. After that, the fingerprint template replacement methods are proposed in the condition of the limited database which is used to store fingerprint templates. In the process of templates updating constantly, when the template number has reached the maximum storage capacity, it need to employ the corresponding replacement method, and replace the unrepresentative fingerprint template with a new fingerprint template, which is also in order to make the fingerprint template database more representative. However, when a new sample matches with fingerprint templates of all users, it will get a set of matching scores. While updating or replacing the template set by using the above methods, it only uses the maximum matching score to determine the corresponding user’s identity, and does not use the rest matching scores. So it does not make full use of the whole matching scores, and may affect the recognition rate of fingerprint identification system.Secondly, we improve the FSS based fingerprint identification method, and apply a semi-supervised learning paradigm by using a lot of unlabeled samples to improve the performance of fingerprint recognition system. In model application, it need to a lot of labeled samples to design the classifier with high accuracy, but it is too difficult to collect those samples. However, collecting unlabeled samples is very easy and low-cost. Therefore, it is worthy trying to use a small amount of labeled samples and a lot of unlabeled samples at the same time. So we can use a large number of unlabeled samples to adapt well to fingerprint intra-class variations, thus improve the performance of fingerprint recognition system. Finally, we propose a fingerprint identification method based on semi-supervised FSS by constantly optimizing the FSS Center of each user. This method is divided into two stages:the registration stage and the identification stage. During the registration stage, we use the FSS based fingerprint identification method to train the FSS Center of each finger; During the recognition stage, we collect a lot of unlabeled fingerprint samples to constantly train the FSS Center under the semi-supervised learning framework, in order to make the updated FSS Center to better represent the identity of the corresponding user. We evaluate our method on the DIEE Fingerprint database. The experimental results show favorable performance of our method as compared to state-of-the-art, and can improve the performance of fingerprint recognition system.