Research On Palmprint Liveness Detection Based On Single Image

For the past few years, with the rapidly development of biometric technology, a large amount of applications have been used for personal authentication in our daily life. Biometrics, such as face, palmprint, fingerprint and iris, have gotten its popularity in many security aspect. For instance, the fingerprint recognition in the function of unblocking in iphone5 s, the palmprint authentication used in the Time Card Machines, and the face identification in the online shopping which is still in testing phase, all of them receive the enjoyment and dependence. With the characters of unique for individuals, it is widely used in many areas instead of traditional identification authentication which has the disadvantages of easily being forged and loss. However, most of these applications only use the 2-D images information to analysis whether it is authenticated successfully. Unfortunately, once such personal biometric data is duplicated or stolen, imposters without access privileges can try to authenticate themselves as valid user. Images, videos or mask can be used by imposter to get through the identification system. Considering the above problems, it is necessary for us to detect whether the user is an imposter or an authorized person. Studies about liveness detection have been developed to determine whether a biometric image is from the real person or from a photograph, videos or other materials.Researches on liveness detection focus on face, fingerprint and iris while little researches on palmprint. In this thesis, we concentrated on the palmprint liveness detection which is the first attempt to distinguish the liveness in palmprint.The main content in this thesis is to discriminate whether the input image which is captured by iphone5 is a liveness palmprint or a recaptured palmprint. Palmprint liveness detection database was constructed which is divided into 3 sections considering the different light and background. Images are taken from 167 students that include their left hand and right hand that contain 2654 liveness palmprint and 2887 recaptured palmprint. After analyzing the trait between the liveness palmprint and recaptured palmprint, we find that recaptured images have the features of blur and less detail. From the discovery, we apply local descriptor of BSIF that combined by the image quality assessment to classify used by SVM. Experiments indicate that this method can obtain accuracy of 92%.In addition, considering the superiority of deep learning in many research fields,convolutional neural networks can learn deeper features from images for what are applied in image processing fields. This thesis use Alex Net network, caffe tools and superior property of computing in GPU to train the classify model. The neural network has five convolutional layers, some of which are followed by max-pooling layers and three fullyconnected layer. All we did in this neural network was fine-tuning. Database used in this experiment consists of 1062857 images which was divided by 544051 live palmprint images and 518806 recaptured palmprint images. Experiments on this database with the fine-tuning network show great efficiency and high accuracy which can obtain 100% after voting.