Research On Downsampling For Coding-based Palmprint Recognition

Biometrics(face,fingerprint,iris,etc.)has been a promising authentication way.Compared with other modalities,palmprint,as a novel modality,contains lots of advantages,such as rich discriminative,user-friendliness and high privacy.Due to these advantages,palmprint has attracted much attention in recent years.Among different palmprint recognition method types,coding-based methods are considered as one of the most promising practical mainstream types for they are free from training,not data-driven,and small storage cost and low computational complexity.In coding-based methods,palmprint images are usually evenly divided into 4×4pixels block,each block represents a pixel in the template,so the template size is only1/16 of the original image size.In the traditional downsampling method(TDM),the upper left pixel of each block is directly selected as the feature and completely ignores the other pixels,which leads to the serious information loss and it is hard to reflect the data in the block comprehensively,resulting in many problems,such as high randomness,weak robustness and low discrimination.Hence,three novel downsampling methods and a novel joint feature are proposed and designed in this dissertation,they can all directly embedded into the existing coding-based methods,and the accuracy and robustness of these methods are improved significantly.In this dissertation,the related works and the classic coding-based methods are reviewed at first,the importance and significance of downsampling are analyzed.Sufficient results have shown that the interference of dislocation and rotation can be effectively resisted by only changing the downsampling method.The main contributions of this dissertation are shown as follows:(1)Democratic Voting Downsampling Method(DVDM)is proposed in this dissertation,which gives each pixel equal power to decide the final feature.Due to the number of juries is increased,the feature extracted can reflect the main information in the block better than TDM.Meanwhile,an extended information entropy formulation and some statistical models are proposed to quantify the stability and robustness of DVDM.(2)Extreme Downsampling Method(EDM)is proposed in this dissertation,which selects the best response pixel as the representative of the block to effectively avoid the side effects caused by the poor response pixels.When two images cannot be aligned perfectly,the former winner pixel has a greater chance to be the winner again in winning the representative election of the dislocation block.Therefore,the downsampled feature stable,robust and have a high discriminative ability.(3)In this dissertation,a Joint Feature Method(JFM)is proposed,the feature of JFM is the combination of the features of EDM and TDM.TDM contains the spatial feature and neglects the discriminative and robust features,and EDM is opposite.Therefore,the two methods are complementarity.The score-level fusion strategy is used to reduce the variances of imposter and genuine distributions,and JFM can achieve an impressive performance in palmprint recognition.(4)A Uniformly Window Downsampling Strategy(UWDS)is proposed in this dissertation,which selects the upper left 4 pixels(2×2)of the block as the jury rather than all pixels,and the jury is the input of EDM and DVDM.Thus,it not only retains the spatial information in TDM,the advantages of EDM and DVDM are also inherited,such as high discrimination and strong robustness.As mentioned above,the importance of downsampling is demonstrated and analyzed in this dissertation,and the existing problems of coding-based methods are also discussed.In order to relieve these problems,three novel down sampling method and a novel joint feature are proposed in this dissertation.Sufficient experiments have proofed that these methods could be easily embedded into the existing coding-based methods and they can achieve remarkable performance improvement.