Research On Palmprint Feature Extraction And Matching Methods For Personal Authentication

Biometrics is the technology that uses human behavior or physical characteristics for personal identification. It is based on the biologic data and information technology, and has the advantages of safety, validity and ease of use. Till now, a variety of biologic features have been used for personal identification. For example, fingerprint, face, iris and signature recognition have been investigated for many years and have been applied for a wide range of applications.As an emerging biometric technology, palmprint recognition has received a large amount of attentions since it was proposed. Compared with other biometric technologies, palmprint recognition has many distinctive features. For example, palmprint has much larger area and more information than fingerprint. Only a low-resolution device is required for a high-precision palmprint recognition system. Twins can be easily distinguished by palmprint recognition, which is very hard for face recognition. Palmprint is also easier to collect than iris pattern, and its feature is more stable and reliable than signature. Consequently, palmprint recognition is a very promising biometric technology and receives increasing research interest.Palmprint recognition system is composed of four modules: image acquisition, pre-processing, feature extraction and matching. In these four modules, feature extraction and matching are the most crucial ones. Based on the analysis of current progress of palmprint recognition technology and the evaluation of state-of-the-art recognition methods, we have investigated several key problems in the two modules to improve the precision and speed of palmprint recognition system. We propose a serious of solutions to these problems, which are described as follows:(1) Orientation selection method for competitive code: The competitive code method uses six oriented Gabor filters for feature extraction. This method, however, doesn't take the characteristics of palmprint orientation into account. Based on the orientation distribution of a set of real palmprint images, we improve the competitive code by selecting the filter orientation using modified fuzzy C-means (FCM) algorithm. In order to meet the requirement of competitive code that the neighboring orientations should have the same angular distance, we first formulate the requirement as a regulation term and then add it to the objective function of FCM algorithm, and finally give the new update rules. Experimental results show that the modified FCM algorithm can find more representative orientations, and recognition accuracy will be improved if we set them as the orientations of Gabor filters.(2) Feature extraction by steerable filters: We propose a novel feature extraction method, which can extract palm line feature and orientation feature simultaneously. Compared with filter-bank based orientation feature extraction method, this method can extract more accurate features and is more computationally efficient. Compared with modified finite Radon transform (MFRAT) based palm line feature extraction method, this method is more effective and easier to implement.(3) Generalized angular distance: Based on the continuous orientation representation, we investigate the appropriate distance measure for palmprint recognition. We propose a generalized angular distance, and also give its discrete form for discrete orientation representation. We show that other two distance measures in the literature can be viewed as special cases of generalized angular distance. To make it more valuable in practice, we propose a novel coding scheme and a look up table based fast matching method.(4) Fast palmprint identification methods without losing accuracy: Palmprint identification system often adopts brute force searching. In order to improve the identification speed without accuracy losing, we propose a fast identification method based on competitive code and cover tree method. We give an equivalent form of distance measure of competitive code which makes it a metric, and then combine it with cover tree method. We further propose a palm tree based fast palmprint identification method and its improved version. Both of them utilize the relationship between templates of the same person to build a tree structure, and reduce unnecessary matchings to speed up the identification process.(5) Palmprint registration and matching methods: There are always some rotation and non-linear deformation in palmprint image after pre-processing. This will affect the genuine matching, and decrease the recognition accuracy. In order to solve this problem, we first propose a palmprint registration method based on principal line and ICP algorithm, which can correct the holistic translation and rotation. We further propose a palmprint matching method based on MRF model. It first divides the template into small patches and then models the non-linear deformation by different translation and rotation of the patches. Finally it estimates the deformation parameters and the matching score by belief propagation algorithm.