Research On Key Issues Of Contactless Palmprint Recognition

As an emerging biometric technology, palmprint recognition has its own advantagescompared with other biometric technologies, such as high use-friendliness and high dis-criminability. At the same time, the acquisition devices for capturing palmprint imagesare low in price. Therefore, palmprint recognition has its high application potential capac-ity. Most of the existing palmprint recognition methods usually employ a contact device,which guarantees that the captured images have pure background and balance illumina-tion, and at the same time, avoids the image blurring caused by hand movement. There-fore, the traditional contact palmprint recognition achieves a high accuracy. However, thecontact palmprint recognition has to be improved on the user-friendliness.Contactless palmprint recognition is a prominent technology for improving the user-friendliness and broaden the applications of palmprint recognition. In contactless occa-sions, due to the removement of contact guiding pegs, the accuracy of the system willbe influenced by palmprint deformations. Aiming at solving the problems in contactlesspalmprint recognition, this dissertation focuses on two of the key issues, i.e., contactlesspalmprint image acquisition, and deformed palmprint recognition.This work designs and implements a contactless multiple hand feature acquisitionsystem, which can capture palmprint, palm vein, and dorsal hand vein images simulta-neously, and during the capturing, users need not to touch any part of the device, andtherefore, the system achieves a high user-friendliness. Besides, we establish a contact-less palmprint image database containing about 4,000 images, which provides us with adata supplement.For deformed palmprint recognition, we firstly establish a universal model for palm-print recognition, in which the matching of palmprints is taken as a question of opti-mization. Based on the proposed model, deformed palmprint recognition is then studiedfrom two aspects, i.e.,linearly-deformed palmprint recognition and non-linearly-deformedpalmprint recognition. As for the former one, a method based on palmprint image reg-istration is proposed for solving the linear deformation problem. In the method,a lineartransformation model is computed using the matched scale invariant feature transform(SIFT) point between palmprint images. The computed transformation model is thenused for palmprint registration, and palmprint feature extraction and matching are per-formed on the registration palmprint image. The method is proved to be e?ective forsolving linear deformation problem.As for the non-linearly-deformed palmprint recognition, we firstly propose a methodfor modeling a non-linear deformation, in which a non-linearly-deformed palmprint isapproximated as piece-wise linear transformations based on matched SIFT points. Thepiece-wise linear transformations are obtained through the proposed iterative randomsample consensus(I-RANSAC) algorithm. Based on the non-linear transformation model,a palmprint recognition method using I-RANSAC and local palmprint descriptors is pro-posed. This method firstly computes the non-linear transformations using I-RANSAC,and based on the computed model, outliers are removed. Local palmprint descriptorsare then employed to further remove mis-matched points. The method can maintain asmuch as possible truly matched SIFT points between palmprint images with non-lineardeformations, and therefore, the recognition accuracy can be improved.For employing the discriminative areas on palmprints while considering palmprintnon-linear deformations, a palmprint recognition method based block growing is pro-posed. After approximating a non-linear transformation using piece-wise linear transfor-mation models, this method partitions a palmprint into several regions according to thelinear transformation models, and performed image registration on each of the regionsrespectively. Through this way, most of discriminative areas on palmprints are employedfor recognition, and subsequently the recognition accuracy is improved.