Research On Algorithm For Palmprint Recognition

Palmprint recognition is a new developing personal identification method. Palmprint refers to principal lines, wrinkles and ridges on a palm. Palmprint pattern is determined by gene and very stable. Everyone has different palmprin patternt. Even for twins, their palmprints are only similar to each other.The Hong Kong Polytechnic University and Tsinghua University proposed the idea of palmprint recognition during 1997 to1998. After that many domestic and international famous research institutions began to make study into this field. Palmprint recognition, as a well- accepted biometric technology, draws great attention from the researchers in many fields recently. In 2003, the Hong Kong Polytechnic University set up the biggest palmprint images database in the world. Also they establish the first civil online palmprint recognition system. On the other hand, NEC and PRINTRAK both have developed palmprint recognition systems for criminal research. David. Z et al have done much creative work. Although there are several published papers so far, the algorithm for palmprint recognition is still a rather challenging field.The palmprint images can be obtained online or offline. They both have disadvantages. Offline image capture can't realize real time palmprint recognition. Online image capture usual use contact way, which has the problem of sanitation, low quality image and low resolution. In this paper, a contactless palmprint image acquisition device is designed with high resolution, which is composed of LED light source, triangular prism, ultraviolet radiation filter and infrared filter, global len and high resolution CMOS camera.When capturing the palmprint image, the hand is put under the oblique LED light source. The hand is above the prism contactlessly. The output data of the CMOS image sensor is in Bayer format, which keeps the original information of every bit in the pamlprint image with resolution of 5 million pixels.The captured palmprint contains noise. The sizes of the palms differ between different people. Even for the same person, the way the palm is put is different at different occasion, which leads to the shift and rotation in the image. Furthermore, the device applies the same focal distance. It is possible to get defocus images. For the above reasons, we have to use the appropriate method to preprocess the image before feature extraction.First we locate the palmprint image by use of the points at the root of the fingers, and then select the center region of the palmprint as the region of interest (ROI). Then we restore the image, including separating the image into several overlap blocks, reducing noise by ridgelet transform and restoring the defocus palmprint using the regularization method. After that the image is normalized for size, grey mean and variance. Finally we use unsharp mask method to compensate the illumination.Palmprint imge consists of linear edges with different scales. This requires an analysis tool with muti-scale analysis property. At the same time, the textures of the palmprint are anisotropic, which require a tool with anisotropy too. Furthermore the palmprint is in high resolution, it needs the analysis tool must be effective and efficient.In this paper, we propose the hierarchical multi-scale feature extraction method based on ridgelet.Divide the image into overlap blocks of size 64×64 to reduce the influence of palmprint location error on the recognition accuracy.We use the wavelet to decompose the image blocks based on the property of multi-scale analysis to get the approximation, horizontal details, vertical details and diagnose details W 1, W2,W3,W4.Then ridgelet is applied to W 1, W2,W3 and then R1,R2,R3 are obtained.Because palmprint image has many different scale linear edges, the resulting matrices are sparse. We combine the first four biggest coefficients in each block in R1 to form the vector R1 . We get R2, R3 in the same manner.Then the palmprint image feature is expressed as three different level vectors R1 , R2, R3, where Rv1 represents large and medium-scale features while R2, R3represent small-scale features.Here R = ( R1,R2,R3) is defined as the hierarchical multi-resolution ridgelet feature vector, where R_j~j is the first four biggest coefficients in each block after ridgelet decomposition, i =1,2,3, j = 1, 2,L,49×49×4.The defined hierarchical multi-resolution ridgelet features can capture the information of palmprint image from all kinds of scales, locations and directions. It also represents the features from coarse level to fine level. The proposed method retains the complete palmprint information and enhances the operation speed and reduces the complexity of computation.Palmprint feature matching is a process to classify the palmprint images. That is to match the test palmrpint against the enrolled palmprints in the database and use the proper matching strategy to determine whether the person is the authorized user and recognize his identity.After performing the ridgelet transform, the definition of hierarchical multi-scale ridgelet feature is given. To test the distinguishing ability of the feature vector, we use the following distance formula to evaluate the similarities of multi-scale ridgelet at different levels:Here R ijand S ijare the combination of the first four coefficients of Rr ifor two image blocks from two palmprint images, i =1,2,3,j = 1, 2,L,49×49×4.We provide the hierachical pamprint matching by use of the three different levels features. First we use the first level feature Rv1 , which represents the large and medium-scale feature, to perfom coarse-level classification. It can find the similar image subset in the database quickly. If there is big variation at this level, the palmprint must be different from the test one. Otherwise move to the next level matching.R2and R3, the second and third level features, represent the small-scale features. If the two palmprint images are still alike at the second leve, then match at third level. The final matching result is the enrolled image, which has the shortest distance in the database.We match each test sample against the enrolled images and obtain the genuine and imposter distribution of the first to third level and ROC curves. The EER for the first level is 0.53, 0.65for the second level and 0.4 for the third level. The experimental results show the effectiveness and accuracy of the method.