Font Size: a A A

Study On Quality Assessment And Feature Recognition Approach Of Finger-vein Image

Posted on:2013-07-27Degree:DoctorType:Dissertation
Country:ChinaCandidate:H F QinFull Text:PDF
GTID:1228330392953994Subject:Instrument Science and Technology
Abstract/Summary:PDF Full Text Request
Finger-vein recognition is a biometric using vein blood vessel under surface toidentify individuals. As finger-vein is the inner pattern, difficult to be copied and forged,it has received more and more attentions. This thesis has done a lot of researches onfinger-vein quality assessment, feature extraction, matching and classification, andproposes an effective finger-vein recognition approach which can achieve betterperformance. The main contribution of this thesis can be concluded as follows.(1) We proposed a finger-vein quality assessment approach by analyzing thesefactors which caused the degradation of finger-vein image. Firstly, the quality score ofeach block is computed by using Radon transformation to detect the vein pattern in eachblock, and then combining the scores of all blocks, the quality score of a grayscaleimage is obtained. In addition, to improve accuracy of identifying low quality image,we proposed three functions to compute the quality of binary image. We assess thequality of an inputted image by fusing the scores of binary image and grayscale image.The proposed approach is tested on our finger-vein database. Experimental results haveshown that the Radon transformation based approach and fusion approach can rejectlow quality image effectively. The fusion approach performs consistently better thanRadon transformation based approach at each level with around5%higher accuracy.However, the fusion approach will take more time for computation the quality of animage. To test the how the templates are chosen will have an impact on the performanceof a finger-vein recognition system, we select image corresponding to different qualityscores as template and compute the equal error rate (EER) of finger-vein recognitionsystem. From the experimental results, we can see that selecting image with middlescore can achieve a lower equal error rate (EER), which is very important to thetemplate selection in other biometric system.(2) We proposed a region growth based finger-vein pattern extraction method. Asthe vein blood vessel is continuous, we bring the region growth concept into extractingfinger-vein pattern. Based on the characteristics of finger-vein, the growth is moved onthe vein pixels. When the growth operator is performed at all points, the vein patternsare extracted from the vein image. We carry out the experiments on our finger-veindatabase and finger-vein database constructed by National Taiwan University of Scienceand Technology. When the false accept rate is0.1%, the genuine accept rates of our approach on the two databases are90.93%and98.33%, respectively.(3) We proposed a finger-vein verification approach based on multi-features fusion.As the single feature such as vein shape contains limited discriminating information,employing more features for finger-vein recognition may achieve higher accuracy. Inour thesis, the finger-vein recognition accuracy can be improved by combinedfinger-vein shape, orientation and Scale Invariant Feature Transform (SIFT) features. Inaddition, to match two images effectively, an image is divided into different blocks.Then we compute the distance based on matching each block from two images. As theSIFT-based partition approach can overcome the global changes and local changes, theproposed approach can achieve better performance. The experimental results on ourfinger-vein database and finger-vein database constructed by National TaiwanUniversity of Science and Technology have consistently demonstrated that the proposedapproach can improve the performance of finger-vein recognition system.(4) We proposed a feature extraction approach for finger-vein recognition.Different pixels in different image parts may have different discriminative powers. Toexplicitly exploit the different importance of each pixel of a sample, we proposed aweighted region covariance matrix (WRCM). Moreover, in order to capture morediscriminating information, we compute weighted Gabor region covariance matrix(GWRCM) by incorporating the Gabor features into weighted region covariance matrix(WRCM). Finally, the dissimilarity between two images is computed based on theGWRCM and the classifiers of nearest neighborhood are employed. The experimentalresults on two databases have demonstrated the efficiency of our approach.
Keywords/Search Tags:Finger-vein recognition, Quality assessment, Region growth, Multi-featuresfusion, Region covariance matrix
PDF Full Text Request
Related items