Low-Quality And Small-Area Fingerprint Verification

Fingerprint verification is one of the most widely used authentication technologies.With the increasing demands of consumers for the portability and the increasing integration of mobile devices,the size of fingerprint sensors continues to decrease,resulting in the lack of effective information for fingerprint matching.In addition,the fingerprint capturing conditions are complex and diverse,some unfavorable conditions such as skin cracks and stains will decrease the quality of fingerprint image,which may affect the accuracy of fingerprint verification significantly.Therefore,the small-area and low-quality fingerprint recognition technology not only has broad prospects but also is a challenging problem which needs to be solved urgently.This thesis focuses on low-quality and small-area partial fingerprint verification.We propose a brand-new framework for small-area fingerprint verification,which achieves the best performance of small-area fingerprint verification in the public dataset and our self-built dataset.The main contributions and innovation are as followed:(1)Image enhancement algorithm.Considering the low image quality and the diversity of fingerprint capturing conditions,we firstly use cartoon-texture decomposition to eliminate the image appearance diversity introduced by different fingerprint capturing conditions and improve the generalization ability of the proposed algorithm.Secondly,we use neural convolution networks to estimate the orientation field of fingerprint image.Thirdly,we apply the Gabor filter on the image with the estimated orientation field to denoise and enhance the clearness of the fingerprint ridge structure.(2)Fingerprint alignment.Considering the large rotation and translation between two partial fingerprint images captured from the same finger,we design a Spatial Transformer Network(STN)to automatically align the input fingerprint image pairs.We use multi rotation and weight-sharing schemes,which can adapt to any rotation angle between two images,to improve the performance of affine transform parameters estimation.(3)Fingerprint comparison.Considering the lack of minutiae in small-area fingerprint image as well as the difficulty of minutiae and orientation field labeling in big data scenario,we proposed a binary-classification network Compare Net,which using sliding windows to crop multi-size fingerprint patches to learn multi-size features.In addition,we introduce the local self-attention mechanism to the network by a multiply-network to help Compare Net to focus multi-level features of fingerprints.Experiments show that our proposed Compare Net is able to learn level-1 and level-2features of fingerprint,such as the ridge flow and minutiae,without the prior information from fingerprint feature annotations and performs well in low-quality and small size fingerprint verification.This thesis focuses on the low-quality and small-area fingerprint verification,our proposed framework achieves the best performance of EER: 3.15% in FVC2006 smallarea verification competition,which outperforms the rank-1 entry of FVC2006 participants,whose EER is 5.56%,and the well-known commercial algorithm Veri Finger12.0,whose EER is 4.04%.In addition,we apply our algorithm on the consumer scenario of small-area under-screen fingerprint verification of smart phone and achieve the performance of the FNMR 3.103% with the FMR under 0.02%.