Research Of Multi-level Feature Extraction And Recognition For Palmprint Images

Palmprint recognition has received increasing research interesting in recent years.It is believed that the palmprint recognition methods are essentially associated with the types of palmprint images.In general,plamprint images can be classified in terms of different criteria.For example,according to the resolution,palmprint images can be divided into low-resolution and high-resolution palmprint images.In addition,the palmprint images can also be categorized into contacted-based and contactless palmprint images according to the way of palmprint image acquisition.Furthermore,based on the dimension,they can be grouped into 2D and 3D palmprint images.In general,a low-resolution palmprint image mainly depicts the line and texture features,and a high-resolution palmprint image contains the minutiae point features a palmprint.By contrast,a 3D palmprint image mainly preserves the structure information of a palmprint surface.Therefore,it is necessary to develop different kinds of palmprint recognition methods for different types of palmprint images.In this dissertation,palmprint images are divided into four categories: low-resolution contacted-based palmprint image,contactless palmprint image,high-resolution palmprint image and 3D palmprint image.We carefully analyze different types of palmprint images and correspondingly propose different kinds of palmprint recognition methods.Specifically,the major works of this dissertation are as follows.(1)For low-resolution palmprint image recognition,we proposed a neighboring direction indicator(NDI)based method and a half-orientation based method,respectively.The orientation features have successfully used in low-resolution palmprint image recognition.However,the conventional palmprint recognition methods usually extract one dominant orientation features of palmprint images,which is sensitive to the rotation and noise,and further many points in a palmprint have multiple orientation features.To this issue,we propose a neighboring direction indicator(NDI)based method,which shows good robustness and can better represent the multiple orientation features.In addition,the conventional palmprint recognition methods are all based on a assumption that the points of palmprint images are in some straight lines.However,many lines of palmprint images are curves and even fold lines.To this end,we propose a half-orientation code(HOC)method to describe the orientation features of curves in palmprint images,and further the discrepancy of the double HOCs can effectively reflect radian of the curves.(2)This work proposes a model integrating the low-rank representation(LRR)with principal line distance for contactless palmprint recognition.Contactless palmprint images are captured under free conditions.So the contactless palmprint images are always influenced by the positions of palm,illumination changes and even suffer from noise.Actually,it is possible to assume that contactless palmprint images are drawn from different subspaces,which motivates us to use LRR for contactless palmprint recognition.LRR is able to produce good clustering results,and the principal lines are the most stable features that can be correctly extracted from contactless palmprint images in most scenarios.Therefore,the proposed method combining the LRR with principal line distance can effectively improve the accuracy of contactless palmprint recognition.(3)High-resolution palmprint images possess rich minutiae-based features with high discriminative ability.However,wrinkles are widely presented in high-resolution palmprint images,which interpret the distribution of ridge flows,and then produce a lot of spurious minutiae points.In this dissertation,the neighbor-area consistency is used to update the extracted ridge orientation,and then a improved Gabor filter is designed to enhance the palmprint images.Thus,the ridge flows can be recovered and spurious minutiae points are effectively reduced.In the matching stage,we rotate the the principal ridge direction of palmprint images according to a certain template for the rotation alignment,which effectively improve the performance of feature matching.(4)This thesis proposes to fuse the 2D orientation features with the 3D surface type features for 3D palmprint image recognition.3D palmprint images save the depth information of the palm surface,and meanwhile contain the texture features of the palmprint.In this thesis,we propose an enhanced competitive code method to extract both the orientation feature and its stability information,both of which can effectively represent the 2D orientation feature of 3D palmprint images.Further,we use the blocked-based surface type(ST)histogram to represent the 3D features of the 3D palmprint surfaces.Experimental results show that the proposed fusion method performs better than the conventional3 D palmprint image recognition methods in most cases.In summary,we carefully analyze the characteristics of different types of palmprint images and design effective feature extraction and matching methods for them.