Research On The Finger Vein Image Segmentation Method

Biometrics recognition can employ some biometrics of bodies for identity verification,which is more convenient and secure.In recent years,biometrics in finger(i.e.,finger vein,finger dorsal texture and finger knuckleprint)have attracted increasing attention due to the convenience and security of these biometrics.In order to employ the effective information in fingers and remove the useless background information,finger segmentation should firstly be performed.Therefore,robust finger segmentation method plays an important role for improvement of recognition performance.Some methods can be used for finger segmentation.However,several problems still exist.Intensity between background and finger regions,and the noise existed in finger edge can degrade the performance.In addition,the pixel-level operation and the ignorance of relationships among pixels reduce the efficiency and effectiveness of segmentation methods.This thesis focuses on two problems of finger segmentation in finger vein images:intensity between background and finger regions,and the noise existed may degrade performance.In addition,high computational complexity and the ignorance of relationships among pixels may reduce the efficiency and effectiveness of the methods.The contributions of this thesis are summarized as follows:(1)Finger segmentation method based on sparse representation and level set is proposed to deal with problems of intensity overlap and noise.The spare representation is used to extract the sparse feature of pixels.After that,the coarse segmentation result is obtained based on the sparse features.And then the coarse finger segmentation region is used as the initial region of level set,and level set is performed to obtain finer finger segmentation result.The spare feature can represent the pixels by using intensity,local and spatial characteristics,which is robust to intensity overlap.In addition,level set is robust to noise and can obtain smooth edge,which can further improve the segmentation performance.(2)Finger segmentation method based on multi-scale superpixels and conditional random fields is proposed to improve the efficiency and effectiveness of the methods.The image is divided into multi-scale superpixels by using SLIC.Superpixels-level is used as the elementary processing unit.After that,intensity,texture and local features are extracted for each superpixel and the conditional random field is used for superpixels classification.The number of generated superpixels is always far less than pixels,leading to computation complexity reduction.Multi-scale analysis can combine the effective information at different scales.In addition,conditional random fields can employ the spatial relationships between superpixels,which can improve the effectiveness of the method.This thesis analyzes problem of intensity overlap and noise,and problem of high computational complexity and ignorance of useful information.To solve these two problems,finger segmentation method based on sparse representation and level set,finger segmentation method based on multi-scale superpixels and conditional random fields are developed.The experimental results on two public finger vein databases demonstrate the effectiveness of the proposed methods.The development of this study establishes solid foundation for performance improvement of finger biometrics recognition systems.