Minutiae-Guided Fingerprint Embeddings via Vision Transformers Steven A. Grosz Michigan State University

2025-05-02 0 0 3.39MB 10 页 10玖币

侵权投诉

Minutiae-Guided Fingerprint Embeddings via Vision Transformers

Steven A. Grosz

Michigan State University

groszste@msu.edu

Joshua J. Engelsma*

Rank One Computing

josh.engelsma@rankone.io

Rajeev Ranjan

Amazon

rvranjan@amazon.com

Naveen Ramakrishnan

Amazon

navramk@amazon.com

Manoj Aggarwal

Amazon

manojagg@amazon.com

Gerard G. Medioni

Amazon

medioni@amazon.com

Anil K. Jain

Michigan State University

jain@msu.edu

Abstract

Minutiae matching has long dominated the ﬁeld of ﬁn-

gerprint recognition. However, deep networks can be used

to extract ﬁxed-length embeddings from ﬁngerprints. To

date, the few studies that have explored the use of CNN ar-

chitectures to extract such embeddings have shown extreme

promise. Inspired by these early works, we propose the ﬁrst

use of a Vision Transformer (ViT) to learn a discrimina-

tive ﬁxed-length ﬁngerprint embedding. We further demon-

strate that by guiding the ViT to focus in on local, minu-

tiae related features, we can boost the recognition perfor-

mance. Finally, we show that by fusing embeddings learned

by CNNs and ViTs we can reach near parity with a com-

mercial state-of-the-art (SOTA) matcher. In particular, we

obtain a TAR=94.23% @ FAR=0.1% on the NIST SD 302

public-domain dataset, compared to a SOTA commercial

matcher which obtains TAR=96.71% @ FAR=0.1%. Ad-

ditionally, our ﬁxed-length embeddings can be matched or-

ders of magnitude faster than the commercial system (2.5

million matches/second compared to 50K matches/second).

We make our code and models publicly available to encour-

age further research on this topic: github.com/tba.

1. Introduction

Over the past several decades, ﬁngerprint recognition

systems have become pervasive across the globe in a num-

ber of different applications, from mobile phone unlock to

national ID programs [2]. The widespread adoption of Au-

tomated Fingerprint Identiﬁcation Systems (AFIS) can be

*This author’s afﬁliation was with Amazon at the time of writing this

paper, but is now with Rank One Computing.

Figure 1: The most prevalent ﬁngerprint representation is

comprised of a variable length, unordered minutiae (key-

point) set. (a) A full minutiae set from a computer gener-

ated (synthetic) ﬁngerprint [1]. Each minutia point has a

location (x, y)and an orientation θindicating the position

and direction, respectively. (b) Examples of the two types

of ﬁngerprint minutiae (Termination and Bifurcation).

primarily attributed to two major tenants:

Accuracy: According to the ongoing Proprietary Finger-

print Template III (PFT III 1) evaluations conducted by the

National Institute of Standards and Technology (NIST), ﬁn-

gerprint recognition systems are now able to obtain recog-

nition accuracies across multiple operational datasets (col-

lected for various use-cases) in excess of 99%.

Scientiﬁc Understanding: Fingerprints were long believed

to be both permanent (retaining the same high accuracy over

time) and unique (different for every person, even differ-

1https://www.nist.gov/itl/iad/image-group/

proprietary-fingerprint-template-pft-iii

arXiv:2210.13994v2 [cs.CV] 26 Oct 2022

Figure 2: An example illustrating minutiae correspondences

between a pair of synthetic ﬁngerprints [1] of the same

ﬁnger. A total of 23 minutiae points are in correspon-

dence shown with green lines. Correspondences were au-

tomatically established using the graph matching algorithm

from [6]

ent ﬁngers of the same person). Rigorous statistical analy-

ses have demonstrated that these central tenants are indeed

backed by strong evidence [3, 4].

Both of these tenants have been established primarily

via the use of long-standing discriminative features widely

known as ﬁngerprint minutiae [5]. Minutiae points are

anomalous key-points located throughout the ﬁngerprint’s

friction ridge pattern. These anomalies occur as either i)

terminations or ii) bifurcations (see (b) of Figure 1 for ex-

amples). Furthermore, each minutiae point is a 3-tuple of

(x, y, θ)where x, y indicate the location of the minutiae

point and θis the direction of the ridge ﬂow at the minu-

tiae point’s location.

In almost all state-of-the-art (SOTA) ﬁngerprint recogni-

tion systems, a full minutiae set (shown in (a) of Figure 1)

is ﬁrst extracted from a given ﬁngerprint. Subsequently,

this variable length, unordered set of minutiae key-points

is compared to another set of minutiae key-points extracted

from an enrolled ﬁngerprint image using graph matching

techniques (Figure 2). At the most basic level, if a simi-

larity score aggregated from corresponding points is more

than a speciﬁed threshold, the ﬁngerprint pair is determined

to be a genuine match, otherwise a non-matching imposter

pair.

Although the success of the minutiae based features have

led to minutiae and automated ﬁngerprint recognition be-

ing nearly synonymous terms, minutiae based ﬁngerprint

matching systems do have several signiﬁcant limitations:

Inefﬁciency: Minutiae matching is computationally expen-

sive. First minutiae must be detected. Oftentimes, descrip-

tors are also extracted for each minutiae point. Then, these

variable length, unordered sets need to be compared via ex-

pensive graph matching techniques. In contrast, most SOTA

face recognition systems which rely on deep feature repre-

sentations require only a dot product between ﬁxed-length

embeddings of query and enrollment images to compute a

match score (dmultiplications and d−1additions for a d

dimensional embedding). Vulnerability: Matching ﬁnger-

print templates2in a secure encrypted manner is extremely

challenging. In contrast, fully homomorphic encryption

(FHE) schemes on deep embeddings have now shown the

ability to match biometric templates in the encrypted do-

main in real-time [7, 8].

These limitations of the minutiae template and also the

demonstrated success of SOTA deep networks to extract

highly discriminative biometric embeddings from faces [9,

10], has initiated exploration in alternative representations

to ﬁngerprint minutiae. In particular, works from [11, 12,

13, 14, 15, 16] all explore the use of deep networks to

embed ﬁngerprint images into a compact, discriminatory,

ﬁxed-length ﬁngerprint representation. These works show

tremendous promise in terms of both the accuracy and speed

needed to either supplant or at least complement the es-

tablished minutiae template. For example, Engelsma et al.

showed in [17] that a 192-dim deep ﬁngerprint embedding

could reach near parity with COTS matchers in terms of

authentication and search accuracy on NIST SD4 [18] and

NIST SD14 [19] datasets while matching at 3 or 4 orders

of magnitude faster speed (300 milliseconds search time vs.

27,000 milliseconds search time on a gallery of 1.1 million

ﬁngerprints). On resource constrained devices and for civil

ID and law enforcement databases with hundreds of million

images in the datasets, these improvements in search time

are invaluable.

Given the complementary strengths of the minutiae tem-

plate and the deep ﬁngerprint templates (human inter-

pretability, statistical understanding, and interoperability of

the minutiae template) vs. (accuracy and speed of the deep

templates), a natural idea is to learn a ﬁngerprint embedding

which somehow distills knowledge of ﬁngerprint minutiae

into the parameters of the deep network. Rather than com-

pletely discarding the minutiae template, we lean on this

domain knowledge to learn more discriminative and gener-

alizable deep ﬁngerprint embeddings. Indeed, the models

in [17, 13, 15] all aim to do just that using Convolutional

Neural Networks (CNN) in combination with distillation of

minutiae domain knowledge.

Inspired by the merging of minutiae domain knowledge

into deep CNNs, in this work, we explore the ﬁrst ever use

of a Vision Transformer (ViT [20]) to learn a ﬁxed length

embedding from a ﬁngerprint image. Similar to prior work

(which utilized CNN building blocks), we build on top of

the vanilla ViT with a strategy for incorporating minutiae

domain knowledge into the network’s parameters. The use

of the multi-headed self attention blocks (MHSA [21]) built

into ViT for the extraction of ﬁngerprint embeddings is

2We use the term templates, representations, and embeddings through-

out to denote a set of features extracted from a ﬁngerprint image.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Minutiae-GuidedFingerprintEmbeddingsviaVisionTransformersStevenA.GroszMichiganStateUniversitygroszste@msu.eduJoshuaJ.Engelsma*RankOneComputingjosh.engelsma@rankone.ioRajeevRanjanAmazonrvranjan@amazon.comNaveenRamakrishnanAmazonnavramk@amazon.comManojAggarwalAmazonmanojagg@amazon.comGerardG.MedioniAm...

展开>> 收起<<

Minutiae-Guided Fingerprint Embeddings via Vision Transformers Steven A. Grosz Michigan State University.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Minutiae-Guided Fingerprint Embeddings via Vision Transformers Steven A. Grosz Michigan State University

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: