ArticlePDF Available

A Finger Vein Recognition Method Based on Histogram of Oriented Lines and (2D)2FPCA

Authors:

Abstract

A finger vein recognition method based on two-dimensional histogram of oriented lines (2DHOL) and two-directional two-dimensional Fisher principal component analysis ((2D)²FPCA) is proposed in this paper. Firstly, according to the characteristics of different finger vein lines, the calculation method of gradient amplitude and direction in the histogram gradient histogram (HOG) is improved. The line responses and orientation of pixels in finger vein images are extracted by adopting the real part of 2-D Gabor filter. In this way, 2DHOL features are obtained. Secondly, considering the correlation between columns and rows of a finger vein image, and combining with the image's category information, the dimension of 2DHOL features is reduced by employing (2D)²FPCA. Finally, Euclidean distance is utilized in recognition. Experimental results applying different finger vein image databases show that, the proposed method can effectively improve the accuracy of finger vein recognition, and enhance the robustness to quantity changes of training samples. © 2018, Beijing China Science Journal Publishing Co. Ltd. All right reserved.
30 2 计算机辅助设计与图形学学报 Vol.30 No.2
2018 2 Journal of Computer-Aided Design & Computer Graphics Feb. 2018
收稿日期: 2017-03-21; 修回日期: 2017-06-29. 基金项目: 国家自然科学基金 (61572458). 张丽萍(1989), , 博士研究生,
要研究方向为图像处理与模式识别、手指静脉识别、智能信息处理; 李卫军(1975), , 博士, 研究员, 博士生导师, 论文通讯作者,
主要研究方向为智能计算、机器学习、图像处理与模式识别; (1989), , 博士研究生, 主要研究方向为图像处理与模式识
别、文字识别; 董肖莉(1985), , 博士研究生, 主要研究方向为图像处理与模式识别、机器视觉; 刘文杰(1989), , 博士研究
, 主要研究方向为图像处理与模式识别、手指静脉识别.
一种基于 2DHOL 特征与(2D)2FPCA 结合的手指静脉识别方法
张丽萍
1,2,3)
,
李卫军
1,2,3)*
,
1,2,3)
,
董肖莉
1,2,3)
,
刘文杰
1,2,3)
1) (中国科学院半导体研究所高速电路与神经网络实验室 北京 100083)
2) (中国科学院大学微电子学院 北京 100029)
3) (认知计算技术威富联合实验室 北京 100083)
(wjli@semi.ac.cn)
: 文中提出一种基于二维方向线直方图统计(2DHOL)特征与双向二维费希尔主成分分析((2D)2FPCA))相结合
的手指静脉识别方法. 首先针对手指静脉图像纹路走向的特点, 改进基于梯度直方图(HOG)特征中有关梯度幅值和
方向的计算方法, 采用二维 Gabor 滤波器获取静脉图像的线形响应和方向, 提取 2DHOL 特征; 然后综合考虑行列相
关性和类别信息, 采用(2D)2FPCA 2DHOL 特征进行降维处理,得到手指静脉特征向量; 最后计算特征向量的欧氏
距离. 应用不同手指静脉数据库进行实验的结果表明, 该方法能够有效地提高手指静脉识别率, 并对训练样本数变
化具有较强的鲁棒性.
关键词: 手指静脉识别; Gabor 滤波; 方向线直方图; 子空间学习; (2D)2FPCA
中图法分类号: TP391.41 DOI: 10.3724/SP.J.1089.2018.16302
A Finger Vein Recognition Method Based on Histogram of Oriented Lines and
(2D)2FPCA
Zhang Liping1,2,3), Li Weijun1,2,3)*, Ning Xin1,2,3), Dong Xiaoli1,2,3), and Liu Wenjie1,2,3)
1) (Laboratory of High Speed Circuit and Neural Network, Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083)
2) (School of Microelectronics, University of Chinese Academy of Sciences, Beijing 100029)
3) (Cognitive Computing Technology Wei Fu Joint Laboratory, Beijing 100083)
Abstract: A finger vein recognition method based on two-dimensional histogram of oriented lines (2DHOL)
and two-directional two-dimensional Fisher principal component analysis ((2D)2FPCA) is proposed in this pa-
per. Firstly, according to the characteristics of different finger vein lines, the calculation method of gradient
amplitude and direction in the histogram gradient histogram (HOG) is improved. The line responses and orien-
tation of pixels in finger vein images are extracted by adopting the real part of 2-D Gabor filter. In this way,
2DHOL features are obtained. Secondly, considering the correlation between columns and rows of a finger vein
image, and combining with the image’s category information, the dimension of 2DHOL features is reduced by
employing (2D)2FPCA. Finally, Euclidean distance is utilized in recognition. Experimental results applying
different finger vein image databases show that, the proposed method can effectively improve the accuracy of
finger vein recognition, and enhance the robustness to quantity changes of training samples.
Key words: finger vein recognition; Gabor filter; histogram of oriented lines; subspace learning; (2D)2FPCA
2 张丽萍, : 一种基于 2DHOL 特征与(2D)2FPCA 结合的手指静脉识别方法 255
近年来, 基于生物特征的身份识别技术越来
越受到研究人员的重视, 这项技术主要采用人的
生理特征(指纹、人脸、虹膜、掌纹、静脉等)或行
为特征(语音、步态等)进行身份识. 其中, 指纹、
人脸、虹膜等识别技术已经相对成熟, 能达到较高
的识别率, 并且市场上已有大量相关技术的产品.
随着技术的发展, 上述各项技术的应用局限性也
相应出现: 指纹识别准确性高, 但是接触式使用造
成用户体验不好, 同时造假情况时有发生; 人脸识
别被认为是最直观的生物识别技术, 但是近年来
越来越受到光照、姿态、表情等因素的约束; 虹膜
识别对采集设备及采集过程有过高的要求等.
脉识别是近年来出现的一种较为新颖的生物识别
技术, 虽然该技术起步较晚, 但是因为其具有非接
触性、活体检测、不易伪造、安全性更高等优点,
生物特征识别领域得到越来越多的关注[1]. 静脉识
别主要分为手掌静脉识别和手指静脉识别. 手指
静脉识别利用手指静脉血液中的血红蛋白对近红
外光(波长范围为 690~980
nm)的吸收特性来采集
静脉图像, 进而进行身份识别的技术. 由于静脉血
管分布的随机性, 使得个体的指静脉血管分布呈
现各异的结构, 即使是双胞胎的静脉分布特征也
不相同, 这也使得这项技术用于身份识别的唯一
性得到了保证.
虽然手指静脉识别技术研究起步较晚, 但是在
20 年仍然有大量的研究方法被提出, 其中部分
研究致力于手指静脉识别的各类特征提取方法[2-14],
较有代表性的方法有Miura [2]提出的重复线性跟
踪法获取静脉纹路、各类基于 Gabor 滤波的特征提
取方法[3-6]、基于局部信息的局部二值编码(local
binary patterns, LBP)方法[7]基于频带受限的频域相
位匹配(band limited phase only correlation, BLPOC)
[8]、基于梯度方向直方图(histogram of oriented gra-
dient, HOG)特征提取[9]及各类子空间学习方法[10-11].
这些方法从不同的角度对手指静脉特征进行描述,
都取得了较为理想的结果.
HOG 特征已被成功应用于各类目标检测及识
别任务中, 并证明了其提取特征的有效性. Jia [15]
分析了将 HOG 特征用于掌纹识别的局限性: 由于
掌纹线具有不同的宽度并且线与线之间有复杂的
交叉点, 因此 HOG 特征中梯度的应用并不是一种
好的检测线形响应和方向的工具; 并首次提出基
于方向线直方图统计的掌纹特征描述方法. 该方
法在 HOG 特征的基础上, 利用线形滤波器来获得
掌纹图像的线形响应和方向. 子空间学习方法已
被应用于各类生物特征识别中, 包括手指静脉识
, 其中, 主成分分析(principal component analysis,
PCA)Fisher 线性判别(Fisher linear discriminant,
FLD)是最成熟的 2种算法. 文献[16]提出的
(2D)2FPCA 子空间学习算法是 PCA FLD 2
基本算法的改进, 被直接用于二维手掌静脉图像
的降维处理, 能同时克服一维降维方法协方差矩
阵维数高和单一二维降维方法只能进行一个方向
压缩的缺点.
与掌纹特征类似, 手指静脉也主要包含线形及
方向特征. 本文根据手指静脉识别的特点, 提出一
种基于 2DHOL 特征[15](2D)2FPCA 子空间学习[16]
结合的手指静脉识别方法, 既能利用 HOL 特征提取
获得静脉的线形响应和方向, 又能克服(2D)2FPCA
子空间学习直接应用于原始图像时特征描述不足
及冗余信息较多的缺点. 实验结果表明, 本文方法
无论是在自建数据库还是公开数据库上都能取得
很好的识别效果.
1 HOL 特征提取
HOL 特征提取以 HOG 为基础, 不同之处在于
可以采用不同的线形滤波器来提取指静脉图像的
线形和方向特征, 以取代 HOG 特征中梯度幅值和
方向的计算. Gabor 滤波已经多次被证明了其在计
算机视觉和模式识别中的有效性, 因此本文选择
Gabor 的实部来构建图像的幅值和方向响应.
通常, 在空间域中, 一个二维 Gabor 滤波器是
由一个高斯函数和一个复杂的平面波构成的复合
函数. Gabor 滤波器的实部被称为偶对称滤波器,
形式为
22
00
22
11
(,,, , ) exp cos(2 ).
2
2
xy
Gxy f fx

 









其中, 0
=cos sin, = sin cos,
x
xyyxyf



示滤波器的中心频率,
表示高斯包络线的标准
,
表示 Gabor 滤波器的方向. 通过以上函数,
可以构建一个单个尺度和k个方向的 Gabor 滤波器,
方向的取值为
π(1)
,1,2,,
mmmk
k
.
给定一幅输入手指静脉图像, 当取 k=8 ,
对应的角度为0,,,,
84

可以捕获不同方向的静
脉纹路信息.
256 计算机辅助设计与图形学学报 30
2 (2D)2FPCA 子空间学习算法
2.1 2DPCA 基本概念
普通的PCA 算法需要将数据表示为一维向量,
其原理是求取数据矩阵对应的协方差矩阵, 通过
求取此协方差矩阵最大的r个特征值对应的特征
向量组成变换矩阵, 该变换矩阵可使得数据矩阵
被最大程度分散化. 2DPCA 直接对二维图像进行
处理, 对于一系列输入图像矩阵 i
A, 相应的协方
差矩阵采用二维数据的形式表示为
TT
1
1
[( ) ( )]= ( ) ( );
M
ii
i
EE E
M
  
G AAA A AAAA
其中,
M
表示图像总数, A表示均值图像矩阵, i
A
表示第i个图像矩阵.
2.2 2DFLD 基本概念
FLD 变换的目标是通过寻找一组最佳鉴别矢
, 使得原始数据在变换时同类样本尽可能集中,
非同类样本尽可能分散, 即类间散布矩阵与类内
散布矩阵比值最大. 2DPCA 类似, 2DFLD 也是
直接对二维图像进行处理, 无需对图像进行一维
变换, 破坏图像的二维结构. 相应的求取方式也是
FLD 方法中类内和类间散布矩阵中的数据及均
值用二维图像的形式表示为
T
1()(),
L
Bii i
iN

S
AAAA
T
1()().
ki
L
Wkiki
iT



A
S
AA AA
其中, L表示类别数, Ni表示第i类图像数, i
A
示第i类均值图像矩阵, A表示所有均值图像矩阵,
k
A表示第k个图像矩阵.
2.3 (2D)2PCA 基本概念
对于一个图像矩阵, 采用 2DPCA 对其列向量
进行压缩变换相当于对图像进行水平压缩, 对其
行向量的变换相当于对图像进行垂直压缩,
(2D)2PCA 即为行列方向相结合的 2DPCA.
2.4 (2D)2FPCA 子空间学习算法
2DPCA 2DFLD 可以对图像去行相关, 同时
将水平方向上的分类信息压缩到相对较少的列向
量上, 但是没有考虑到图像的列向量存在的相关
, 导致这 2种方法的压缩率和压缩效果不佳.
(2D)2PCA 虽然考虑了行、列 2个方向的相关性,
是仅用 PCA 方法无法引入类别信息, 因此(2D)2PCA
是一种好的数据压缩方法, 却不是一种好的分类压
缩方法. (2D)2FPCA 方法通过将 2DPCA 2DFLD
相结合, 能很好地克服上述方法的局限性.
3 本文方法
设输入图像 I, 大小为rc, 本文提出的基于
2DHOL+(2D)2FPCA 的手指静脉识别方法流程图
如图 1所示. 步骤如下:
输入. 手指静脉图像
I
.
输出. 图像特征数据
F
.
Step1. 对图像的每个像素 (, )Ixy, 获取 k个方向
上的 Gabor 滤波响应图谱
(,)=(,) (,, )
kk
FxyIxy Gxy
,
其中, (, , )
k
Gxy
是方向为kGabor 滤波器函数,
示卷积操作.
Step2. 构造对应的 Gabor 幅值 Gabor
(,)mxy 和方向
Gabor
(,)xy
响应
Gabor
(,) min( (,))
k
mxy F xy
,
Gabor
(,) argmin( (,))
kk
x
yFxy
.
Step3. 将图像分为互不重合的 mn个单元, 每个
单元包含 12
cc
个像素.
Step4. 12
bb
个单元构成一个块, 相邻的 2个块之
间可以有重合.
Step5. 对每个单元统计一个基于方向的子直方图
(histogram within cell, HC), 子直方图的位数为方向个数
k, 每个位置为对应方向的幅值累加求和
Gabor
() (,) ,
i
kmxy
HC
(,)
f
xy
单元 i并且 Gabor
(,)
x
yk
.
Step6. 串联每个块内单元对应的直方图, 构建基
于块的直方图(histogram of block, HB)
12
12
,,,
jbb
HB HC HC HC,
为了避免出现除数为零的情况, 采用 2
L范数对
HB归一化, 得到归一化的 HB向量(norm HB, NHB)
22
2
=+
j
j
j
HB
NHB HB ;
其中,
为很小的常数.
Step7. 将基于块的子直方图组合成 HOL 特征或
2DHOL 特征
12
,,,
N
NHB NHB NHBL,
2D 1 2
;;;
N
NHB NHB NHBL;
其中, N表示图像分块数.
通常, 采用 22
个单元构成一个块, 则图像块数为
(1)(1)Nm n
, HOL 原始特征维数为 22Nk
 ,
2DHOL 原始特征维数为 (22)Nk
 . 本文算法中, 采用
2DHOL 的形式, 11
,22Nmk n
.
Step8. 对得到的 2DHOL 特征的行方向上进行
2DPCA 变换, 得到变换矩阵
X
.
2 张丽萍, : 一种基于 2DHOL 特征与(2D)2FPCA 结合的手指静脉识别方法 257
Step9. 用变换矩阵
X
将输入特征转换到低维空间
: 11 11
2D
mn
md

X
LY, 并将降维后的数据进行转置
11
11 1dm
md

转置
YY
.
Step10. 对转置后的数据在行方向上进行 2DFLD
变换, 得到变换矩阵V.
Step11. 用变换矩阵V再次对数据进行降维, 得到
最终特征 C:11 12
1dm dd

V
YF
.
原始静脉图像存在数据维数巨大、信息冗余、
对有用信息表示不足等缺点, 因此本文提出的
2DHOL+(2D)2FPCA 手指静脉识别方法以 2DHOL
特征代替原始二维图像作为(2D)2FPCA 的输入,
可以很好地克服(2D)2FPCA 子空间学习方法直接
应用于原始图像时特征描述不足及冗余信息较多
的缺点.
1 本文方法流程图
4 实验与结果分析
4.1 实验数据
本文利用自行研制的手指静脉图像采集装置,
建立实验用手指静脉数据库. 自建手指静脉数据
集包含来自29 人的2
610幅图像, 每人左右手各采
集食指、中指和无名指 3个手指, 每次采集都是 6
个手指作为1次循环, 以增加每次手指按下的随机
, 每个手指采集15幅图像, 296=174个手指
对象. 原始采集图像如图 2所示, 图像大小为
640480. 在进行本文方法验证前, 需要先截取手
指静脉的感兴趣区域(region of interest, ROI), 具体
过程如图3所示. 由于采集设备中固定了手指有效
采集区域(如图 3a 中红色矩形框所示), 因此以红
色矩形框的上下中心为中心截取宽 130 像素作为
ROI 的垂直有效区域; 对于 ROI 截取的左右参考
线, 根据近红外光线可以穿透关节软骨的特性,
像在垂直方向上的投影, 关节处的灰度累加值比
其余地方高, 如图 3b 所示, 因此可以以指关节的
位置作为左右截取的参考线[17]; 本文以第一指关
节为参考线, 左右截取370 像素作为 ROI的水平有
效区域, 如图 3c 所示; 最终得到手指静脉 ROI
像大小为 130370(单位为像素) , 如图 3d 所示.
2 原始手指静脉图像
3 手指静脉 ROI 区域提取流程
4.2 参数选择及实验设置
HOL 特征提取中, 方向滤波器 Gabor 参数设置
如下: 0=0.0332, =10.99f
, 方向数 =8k. 本文采用 2
种统计度量方法进行分析: (1) 正确识别率(recog-
258 计算机辅助设计与图形学学报 30
nition rates, RR), 一对多识别, 即测试图像被正确
识别为本类的概率; (2) 等错误率(equal error rate,
EER), 一对一验证, 即错误接受率(false accept
rate, FAR)和错误拒绝率(false reject rate, FRR)相等
的点对应的值. 所有实验中, 每根手指被作为一个
独立的对象, 以每个对象的前5幅图像作为训练图
, 剩余 10 幅图像作为测试(除了后面关于训练样
本数的实验), 采用欧氏距离度量两两图像间的距离,
采用最近邻分类得到每幅测试图像对每类的度量
结果. 相应地, 一对多识别中, 识别次数为1
740
(17410), 一对一验证中, 共有正样本匹配次数 1
740
(17410), 负样本匹配次数为 301
020 (174
17310).
4.3 实验 1
实验 1. HOL 特征提取中最佳分块参数实验
除了上面提到的实验参数外, HOL 特征提取
, 图像的单元数mn会直接影响 HOL 提取特征
的效果. 本实验的目的是寻找最佳图像分块单元
, 由小到大改变mn, 选取33,34,12
分块方式. 由于分块策略的不同影响 HOL 原始特
征本身对图像的表达, 因此本文直接采用欧氏距
离度量HOL原始特征两两之间的距离,采用最近邻
分类来测试分块数对特征表示能力的影响. 4
示为不同单元数对应的 EER , 可以看出, 单元
数过多或过少对应的 EER 值都增大, 即输入指静脉
图像固定后, 对应的最佳单元数也相应确定, 针对
图像的大小, 当单元数为56
, 可以得到最小的
EER=0.683%. 因此, 在后续的实验中确定单元数
56. 此时, 对应的特征维数为 45822(
果表示为 2DHOL 的形式, 则原始维数为20 32
,
20 代表分块数, 32 代表每块的维数).
4 HOL 中不同分块单元数对应的 EER
4.4 实验 2
实验 2. HOL 特征与不同子空间降维方法相结
合的识别效果实验
本实验对比不同子空间学习方法对原始 HOL
特征降维后的识别效果, 主要对比方法有 PCA,
2DPCA, 2DFLD, (2D)2PCA 以及(2D)2FPCA, 1~
5所示为 HOL 特征+不同子空间学习方法随着
降维主成分数的变化所对应的 EER , 其中,
粗的为最佳 EER 及对应的主成分数. PCA 需要
HOL 特征串联成一维特征外, 其余 4种二维子
空间学习方法的输入都是 2DHOL 特征.
1 HOL+PCA 方法不同主成分数对应的 EER
特征维数 EER/%
640 0.683
400 0.632
200 0.575
150 0.575
120 0.575
100 0.586
80 0.632
2 2DHOL+2DPCA 方法不同主成分数对应的 EER
特征维数 EER/%
32×20 0.683
32×15 0.675
32×12 0.624
32×10 0.690
32×8 0.684
32×6 0.736
3 2DHOL+2DFLD 方法不同主成分数对应的 EER
特征维数 EER/%
32×20 0.683
32×15 0.517
32×12 0.460
32×10 0.403
32×8 0.413
32×6 0.460
4 2DHOL+(2D)2PCA 方法不同主成分数
对应的 EER
特征维数 EER/%
32×20 0.683
25×12 0.524
20×12 0.621
20×8 0.632
17×8 0.690
17×6 0.752
15×6 0.751
2 张丽萍, : 一种基于 2DHOL 特征与(2D)2FPCA 结合的手指静脉识别方法 259
5 2DHOL+(2D)2FPCA 方法不同主成分数
对应的 EER
特征维数 EER/%
20×32 0.683
12×25 0.460
12×20 0.402
8×20 0.287
8×17 0.287
6×17 0.290
6×15 0.336
6所示为上述5种方法在最佳主成分数时对
应的 EER、数据维数和压缩率((原始维数降
降维
后维数)/原始维数)统计. 可以看出, 2DHOL+
(2D)2FPCA 方法在这几个方面都明显优于对比方
. 综合考虑识别效果和数据量, 后续实验选取
2DHOL+(2D)2FPCA 的最优维数为6×17=102, EER=
0.290%.
6 不同子空间学习方法在识别及压缩效果上的比较
算法 EER/% 特征维数 压缩率/%
HOL+PCA 0.575 120 81.25
2DHOL+2DPCA 0.624 384 40.00
2DHOL+2DFLD 0.403 320 50.00
2DHOL+(2D)2PCA 0.524 300 53.12
2DHOL+(2D)2FPCA 0.290 102 84.06
0.287 136 78.75
4.5 实验 3
实验 3. 本文方法与 5种经典方法的比较实验
为了验证 HOL 特征及本文方法在手指静脉识
别中的识别效果, 按照与上述实验相同的设置与
一些经典方法进行对比, 对比方法有 LBP[7], Gabor
滤波[3], HOG[9], BLPOC[8], (2D)2FPCA[16]. 为公平
起见, 各种方法中的参数均采用多次实验选取最
优设置. 5所示为不同方法对应的 ROC 曲线,
对应的 EER 值如表 7所示.
从图 5可以看出, 7 种对比方法中, 单独的 LBP
编码特征表示方法效果最差, 原因是该方法主要
是对原始图像采用 LBP 方法编码然后分块统计直
方图作为特征, 处理方法比较简单; (2D)2FPCA
一种子空间学习方法, 其直接对原始二维图像进
行处理, 是结合了2DPCA2DFLD的一种改进方
, 但是由于原始图像中信息冗余较严重, 而且未
进行任何预处理, 因此学习到的特征对手指静脉
表示有限; BLPOC Gabor 滤波都是频域处理方法,
得到的识别效果明显优于前 2, 特别是 Gabor
5 不同方法得到的 ROC 曲线
7 不同方法得到的 EER
算法 EER/%
LBP 4.200
(2D)2FPCA 3.735
BLPOC 2.348
Gabor 1.667
HOG 0.804
HOL 0.683
2DHOL+(2D)2FPCA 0.290
波近些年在很多模式识别领域都得到了不错的效
, 是一种优质的图像滤波器; 从图 5和表 7的结
果可以看出, HOG 特征用于指静脉识别能得到不
错的识别效果, 其识别结果远好于前面 4种方法;
本文提出的 HOL 特征是针对 HOG 特征的一种改
, 并且用到了 Gabor 滤波作为线形提取滤波器,
因此同时具有 Gabor HOG 2种方法的优点,
EER=0.683%; 而本文提出方法既能利用 HOL
征描述获得静脉的线形响应和方向, 又能克服
(2D)2FPCA 子空间学习直接应用于原始图像时特
征描述不足及冗余信息较多的缺点, 与其他方法
相比, 该方法能得到最小的 EER=0.290%.
4.6 实验 4
为了验证算法的普适性, 选用公开的 FV-USM
手指静脉库进行实验. FV-USM 数据库[18]包含 123
人共 492 个手指对象, 分为训练集和测试集, 每个
数据集、每个手指有 6幅图像, 直接采用提供的
ROI 图像, 部分图像如图 6所示(6a, 6b 中同
一位置表示同一个手指对象的训练图像和测试图
). 由于前面已经在自建数据集上详细比较了本
文方法与几种经典方法的识别效果, 本实验只比
较提出的方法与HOL, (2D)2FPCA 方法单独使用时
的识别效果, 结果如表 8所示.
260 计算机辅助设计与图形学学报 30
a. 训练图像 b. 测试图像
6 FV-USM 数据库示例图像
8 THU-FV 数据库上的实验结果
算法 EER/%
HOL 4.471
(2D)2FPCA 11.410
2DHOL+(2D)2FPCA 2.235
从图6和表 8可以看出, 由于FV-USM 数据库
中的图像质量较差, 因此得到的识别率均低于自
建库; 但本文的方法能在很大程度上提高 HOL
(2D)2FPCA 2种方法的识别率, 得到更优的结果,
EER=2.252%.
4.7 实验 5
实验 5. 训练样本数对识别准确率的影响实验
对子空间学习方法, 训练样本的数量会直接
影响算法的性能, 最终影响识别效果. 选择 HOL
特征与 5种子空间学习结合的方法各自对应的最佳
主成分数, 改变训练样本数量, 统计各种方法对应
的识别率 RR, 结果如图 7所示.
7 5 种方法随着训练样本数变化对应的识别率
从图 7可以看出, 随着训练样本数的增加, 5
方法对应的识别率均有不同程度的提高, 当训练
样本数为 12 (训练样本测试样本=12:3), 识别率均
达到 100%; HOL+PCA; 2DHOL+2DPCA; 2DHOL+
(2D)2PCA 3种方法的识别率较 2DHOL+ 2DFLD,
2DHOL+(2D)2FPCA , 当训练样本数减少时,
文方法识别率下降缓慢, 其余4种方法识别率都有
较大程度的下降. 因此, 本文提出的方法当训练样
本数从 3变化为 12 , 都能得到令人满意的识别
效果, 在训练样本数增加为 10 之后, 对测试集的
识别率已经达到 100%.
4.8 算法复杂度及运行时间
HOL 特征中采用 Gabor 滤波器来获取图像的
幅值和方向响应, 涉及Gabor 模版与图像的卷积
, 因此获取 HOL 特征的主要耗时取决于图像大
小和 Gabor 滤波器模版大小(单位为像素). 本文采
Gabor模版大小为81×81, 图像大小为 370
130,
可采用快速傅里叶变换来加速卷积操作.
(2D)2FPCA 方法的主要耗时在于训练阶段求取协
方差矩阵, 可采用离线训练的方式来节省这部分
时间. 本文实验在 Matlab2014 下运行, 主频为 3.2
GHz CPU, 内存为 8
GB, 本文方法处理一幅测试
图像的总时长为 1
s左右.
5
本文采用一种新的手指静脉特征描述方法
HOL, 该方法针对手指静脉纹路走向, 能有效的
描述手指静脉特征, 并在此基础上研究了不同子
空间学习方法与该特征描述方法的融合, 以提高
手指静脉识别的识别效果, 并最终提出一种基于
2DHOL+(2D)2FPCA 的手指静脉识别方法. 该方法
既能利用 HOL 特征描述获得静脉的线形响应和方
, 有效地对手指静脉纹路进行描述, 又能利用
(2D)2FPCA HOL 特征进行有效的行列压缩,
降低数据量和提高数据表示有效性的同时, 还能
克服(2D)2FPCA 子空间学习方法直接应用于原始
图像时特征描述不足及冗余信息较多的缺点.
自建手指静脉数据集上的实验结果表明, 本文方
法无论对于训练样本数多少都能得到令人满意的
识别效果, 不仅能提高手指静脉识别的识别率,
且能提高手指静脉识别的鲁棒性, 减小方法对于
训练样本集的依赖性. 但是, 耗时较长使得本文方
法离实际应用还有一定的距离, 另外(2D)2FPCA
也存在小样本问题, 在单样本训练图像时会无法
使用. 这些问题也是下一步的研究重点.
参考文献(References):
[1] Hashimoto J. Finger vein authentication technology and its fu-
ture[C] //Proceedings of the Symposium on VLSI Circuits. Los
2 张丽萍, : 一种基于 2DHOL 特征与(2D)2FPCA 结合的手指静脉识别方法 261
Alamitos: IEEE Computer Society Press, 2006: 5-8
[2] Miura N, Nagasaka A, Miyatake T. Feature extraction of finger-
vein patterns based on repeated line tracking and its application
to personal identification[J]. Machine Vision and Applications,
2004, 15(4): 194-203
[3] Xie S J, Yang J C, Yoon S, et al. Guided Gabor filter for finger
vein pattern extraction[C] //Proceedings of the 8th International
Conference on Signal Image Technology and Internet Based
Systems. Los Alamitos: IEEE Computer Society Press, 2012:
118-123
[4] Yang J F, Yang Y L. Multi-channel Gabor filter design for finger-
vein image enhancement[C] //Proceedings of the 5th Interna-
tional Conference on Image and Graphics. Los Alamitos: IEEE
Computer Society Press, 2009: 87-91
[5] Lu Y, Yoon S, Xie S J, et al. Finger vein recognition using his-
togram of competitive Gabor responses[C] //Proceedings of the
22nd International Conference on Pattern Recognition. Los
Alamitos: IEEE Computer Society Press, 2014: 1758-1763
[6] Khellat-Kihel S, Abrishambaf R , Cardoso N, et al. Finger vein
recognition using Gabor filter and support vector machine[C]
//Proceedings of the International Image Processing, Applica-
tions and Systems Conference. Los Alamitos: IEEE Computer
Society Press, 2014: 1-6
[7] Lee E C, Lee H C, Park K R. Finger vein recognition using
minutia-based alignment and local binary pattern-based feature
extraction[J]. International Journal of Imaging Systems and
Technology, 2009, 19(3): 179-186
[8] Mahri N, Suandi S A S, Rosdi B A. Finger vein recognition al-
gorithm using phase only correlation[C] //Proceedings of the
International Workshop on Emerging Techniques and Chal-
lenges for Hand-Based Biometrics. Los Alamitos: IEEE Com-
puter Society Press, 2010: 1-6
[9] Dalal N, Triggs B. Histograms of oriented gradients for human
detection[C] //Proceedings of IEEE Computer Society Confer-
ence on Computer Vision & Pattern Recognition. Los Alamitos:
IEEE Computer Society Press, 2005, 1: 886-893
[10] Guan Fengxu, Wang Kejun, Liu Jingyu, et al. Bi-direction
weighted (2D)2PCA with eigen value normalization one for
finger vein recognition[J]. Pattern Recognition and Artificial
Intelligence, 2011, 24(3): 417-424(in Chinese)
(管凤旭, 王科俊, 刘靖宇, . 归一双向加权(2D)2PCA
手指静脉识别方法[J]. 模式识别与人工智能, 2011, 24(3):
417-424)
[11] Wu J D, Liu C T. Finger-vein pattern identification using prin-
cipal component analysis and the neural network technique[J].
Expert Systems with Applications, 2011, 38(5): 5423-5427
[12] Jin Huaiguo, Sun Dongmei, Wang Shenghui. Fusion of multi-
finger vein based on Fisher criterion[J]. Journal of Computer-
Aided Design & Computer Graphics, 2013, 25(2): 183-188(in
Chinese)
(金怀国, 孙冬梅, 王升辉. 基于 Fisher 准则的多指静脉融合
[J]. 计算机辅助设计与图形学学报, 2013, 25(2) : 183-188)
[13] Mohan M, Prem Kumar R, Agrawal R, et al. Finger vein rec-
ognition using integrated responses of texture features[C]
//Proceedings of the 4th International Work Conference on Bio-
inspired Intelligence. Los Alamitos: IEEE Computer Society
Press ,2015: 209-214
[14] Chen X, Zhang J. Optimized discriminant locality preserving
projection of gabor feature for biometric recognition[J]. Inter-
national Journal of Security and its Applications, 2012, 6(2):
321-328
[15] Jia W, Hu R X, Lei Y K, et al. Histogram of oriented lines for
palmprint recognition[J]. IEEE Transactions on Systems Man,
& Cybernetics: Systems, 2014, 44(3): 385-395
[16] Wang J, Li H J, Wang G Q, et al. Vein recognition based on
(2D)2FPCA[J]. International Journal of Signal Processing, Im-
age Processing and Pattern Recognition, 2013, 6(4): 323-332
[17] Lu Y, Xie S J, Yoon S, et al. Robust finger vein ROI localiza-
tion based on flexible segmentation[J]. Sensors, 2013, 13(11):
14339-14366
[18] Asaari M S M, Suandi S A, Rosdi B A. Fusion of band limited
phase only correlation and width centroid contour distance for
finger based biometrics[J]. Expert Systems with Applications,
2014, 41(7): 3367-3382
... EER/% LBP [24] 4.200 (2D) 2 FPCA [57] 3.735 Gabor [53] 1.667 DBC [33] 1.190 HOG [58] 0.864 HOL [59] 0.683 HCGR [60] 0.806 Adaptive Learning Gabor [44] 0.731 HOPGR [61] 0.678 Ours 0.331 ...
... EER/% MeanC [19] 10.64 RLT [17] 8.25 MaxC [20] 2.65 LBP [24] 6.9 LLBP [26] 5.7 LDC [25] 3.31 HOG [59] 2.54 HOL [59] 2.35 SVDMM [36] 2.45 MBTM [62] 7.88 BSPHVT [63] 5.08 DBC [33] 1.44 CNN [64] 1.21 HOPGR [61] 2.04 SCNN-LSTM [65] 1.12 Modified Densenet-161 (Softmax/LMCP/AAMP) [66] 1.21/ 0.86 /1.01 Ours 1.17 ...
... EER/% MeanC [19] 10.64 RLT [17] 8.25 MaxC [20] 2.65 LBP [24] 6.9 LLBP [26] 5.7 LDC [25] 3.31 HOG [59] 2.54 HOL [59] 2.35 SVDMM [36] 2.45 MBTM [62] 7.88 BSPHVT [63] 5.08 DBC [33] 1.44 CNN [64] 1.21 HOPGR [61] 2.04 SCNN-LSTM [65] 1.12 Modified Densenet-161 (Softmax/LMCP/AAMP) [66] 1.21/ 0.86 /1.01 Ours 1.17 ...
Article
Full-text available
Finger vein recognition has attracted considerable attention from the biometric identification technology community owing to its convenience and security. Unlike most previous works only pay attention to one part of finger vein recognition, we propose a joint Bayesian framework in this paper, which is based on partial least squares discriminant analysis (PLS-DA). It involves three major stages: 1) robust feature description, 2) discriminative feature mapping, and 3) separative verification. In stage 1), we extract line responses and orientation of finger veins using a bank of Gabor filter, and histograms are constructed in local patches as primitive features. Subsequently, in stage 2), a discriminant feature mapping based the PLS-DA (PLS-DA-FM) method is proposed to project these primitive features into low-dimensional forms in a supervised manner. Thus, highly compact and discriminative features are obtained in this stage. Finally, in stage 3), we directly build a Bayesian model based on the joint distribution of finger vein feature pairs to measure the similarity between the features. Extensive experiments on five finger vein datasets demonstrate the superior performance of the proposed method to most state-of-the-art finger vein recognition methods.
... However, gradient information used in HOG is not well suited to accurately represent the orientation and patterns in a finger vein image. As variants of HOG, histogram of competitive Gabor responses (HCGR) [21] and histogram of oriented lines (HOL) [22] use Gabor wavelets to detect the line responses and orientations of pixels. However, these methods use generalpurpose filters that do not consider finger vein-specific features, which limits the adaptability and efficiency of the filters for finger vein images. ...
... It includes 6000 images of 100 volunteers; 10 images of 6 fingers of each volunteer were collected. The SEMI-FV database [22] is a self-built finger vein database that consists of 2610 finger vein images captured from 29 volunteers, with images of six fingers per subject. Some image samples are depicted in Fig. 4. ...
... However, gradient information used in HOG is not a good representation of finger vein patterns, because the patterns have different widths and crisscrossing vein patterns. As variants of HOG, histogram of competitive Gabor responses (HCGR) [21] and histogram of oriented lines (HOL) [22] use Gabor wavelets to detect the line responses and orientations of pixels. However, these methods use general-purpose filters that do not consider finger vein-specific features, which limits the adaptability and efficiency of the filters for finger vein images. ...
... It includes 6000 images of 100 volunteers; 10 images of 6 fingers of each volunteer were collected. The SEMI-FV database [22] is a self-built finger vein database that consists of 2610 finger vein images captured from 29 volunteers, with images of six fingers per subject. Some image samples are depicted in Fig. 4. ...
Preprint
Local feature descriptors exhibit great superiority in finger vein recognition due to their stability and robustness against local changes in images. However, most of these are methods use general-purpose descriptors that do not consider finger vein-specific features. In this work, we propose a finger vein-specific local feature descriptors based physiological characteristic of finger vein patterns, i.e., histogram of oriented physiological Gabor responses (HOPGR), for finger vein recognition. First, prior of directional characteristic of finger vein patterns is obtained in an unsupervised manner. Then the physiological Gabor filter banks are set up based on the prior information to extract the physiological responses and orientation. Finally, to make feature has robustness against local changes in images, histogram is generated as output by dividing the image into non-overlapping cells and overlapping blocks. Extensive experimental results on several databases clearly demonstrate that the proposed method outperforms most current state-of-the-art finger vein recognition methods.
... Researchers have explored this topic extensively over the last two decades, investigating techniques for extracting features from ear images and their subsequent comparison [10,28]. Successful feature extraction techniques in ear recognition and other biometrics include Principal Component Analysis-(PCA) [29][30][31][32]37], wavelet-based [5,13,18,25], Support Vector Machine (SVM) [4,26,27] and neural network-based and other [1,2,7,9,11,15,22,24,27,33,39,40] methods. Amongst these techniques, PCA has been used for both feature extraction in the form of eigenvectors and dimensionality reduction. ...
Article
Full-text available
Principal Component Analysis (PCA) has been successfully applied to many applications, including ear recognition. This paper presents a Two Dimensional Multi-Band PCA (2D-MBPCA) method, inspired by PCA based techniques for multispectral and hyperspectral images, which have demonstrated signi�cantly higher performance to that of standard PCA. The proposed method divides the input image into a number of images based on the intensity of the pixels. Three di�erent methods are used to calculate the pixel intensity boundaries, called: equal size, histogram, and greedy hill climbing based techniques. Conventional PCA is then applied on the resulting images to extract their eigenvectors, which are used as features. The optimal number of bands was determined using the intersection of number of features and total eigenvector energy. Experimental results on two benchmark ear image datasets demonstrate that the proposed 2D-MBPCA technique signi�cantly outperforms single image PCA by up to 56.41% and the eigenfaces technique by up to 29.62% with respect to matching accuracy on images from two benchmark datasets. Furthermore, it gives very competitive results to those of learning based techniques at a fraction of their computational cost and without a need for training.
... e top-down image detection algorithm based on deep learning [14][15][16][17][18] extracts global image information and high-level semantic information by self-learning image features. In contrast, the bottom-up image detection algorithm based on low-level visual features extracts low-level image color and texture. ...
Article
Full-text available
To enhance the effect of motion detection, a Gaussian modeling algorithm is proposed to fix holes and breaks caused by the conventional frame difference method. The proposed algorithm uses an improved three-frame difference method. A three-frame image sequence with one frame interval is selected for pairwise difference calculation. The logical “OR” operation is used to achieve fast motion detection and to reduce voids and fractures. The Gaussian algorithm establishes an adaptive learning model to make the size and contour of the motion detection more accurate. The motion extracted by the improved three-frame difference method and Gaussian model is logically summed to obtain the final motion foreground picture. Moreover, a moving target detection method, based on the U-Net deep learning network, is proposed to reduce the dependency of deep learning on the number of training datasets. It helps the algorithm to train models on small datasets. Next, it calculates the ratio of the number of positive and negative samples in the dataset and uses the reciprocal of the ratio as the sample weight to deal with the imbalance of positive and negative samples. Finally, a threshold is set to predict the results for obtaining the moving object detection accuracy. Experimental results show that the algorithm can suppress the generation and rupture of holes and reduce the noise. Also, it can quickly and accurately detect movement to meet the design requirements.
... Key words: optical coherence tomography; internal and external fingerprints; structure point clustering; pseudo-structure point removal , . [1] [2-3] [4] . ...
Article
Vein pattern-based methods have powerfully promoted the performance of finger vein recognition. However, it is not easy to precisely extract vein patterns from images, especially from low-quality images, and the non-vein area have been proved to be helpful for recognition. This paper proposes to use low-rank representation to extract as much noiseless discriminative information as possible from finger vein images. However, image deformation and image quality variations weaken the correlation of genuine images, and therefore damage the low-rank linear representation. To further deal with this problem, the class labels of training images and the local geometric structure between testing images and training images, reflected by sparse reconstruction errors of testing images, are used as constraints of low-rank coefficients. In particular, vein backbone decomposition based sparse representation is proposed to fast compute the deformation-robust reconstruction errors of each testing image. The reconstruction error on sub-backbones of one training image are used and modified as the constraint of the low-rank coefficient on this training image. We evaluate the proposed method on three widely used finger vein databases, and experimental results show that the proposed method performs well in finger vein recognition.
Chapter
In this work, we propose an efficient joint Bayesian model that expanded with soft biometric traits for finger vein recognition, which is different from the existing mainstream finger vein recognition methods that extract features from the region of interest (ROI) of finger vein images. First, a two-branch convolutional neural networks (CNNs) is designed to extract feature simultaneously from finger shape images and ROI vein images, respectively, the aim of which is to extract more information from different parts of fingers. Then, two different features are fused by concatenating. Finally, a joint Bayesian recognition loss is used to train the proposed two-branch CNN which is to promote intra-class variance minimization and inter-class variation maximization. Further, in our framework, a end-to-end learning method of CNNs and joint Bayesian recognition loss is achieved to optimize the parameters. Experiments performed on two public finger vein databases demonstrate that the proposed model is efficient and achieves better performance than state-of-the-art methods.
Article
In the process of image acquisition, the contrast between veins and non-veins in finger vein images is not high due to the influence of the fuzzy light source, skin scattering and finger movement. To solve this problem, a finger vein image enhancement method is proposed (GTGFs), which enhances finger vein patterns by setting guided image as input image firstly. On this basis, the tri-Gaussian model is based on disinhibitory properties of the concentric receptive field used to locally enhancing the image. The parameters of the tri-Gaussian model are determined based on the finger vein width information. The experiment results show that the proposed enhancement method can significantly enhance the finger vein patterns and improve the recognition effect of the methods based on vein pattern segmentation.
Article
Full-text available
Nowadays biometric identification systems are widely spread since the safety of those systems has been proven. They exhibit a large number of advantages when compared to other identification systems such as key and password that are subject to falsification and loss. Among biometric systems, finger vein recognition based on venous network has been considered recently in the literatures. This paper aims to present a finger vein recognition system using Support Vector Machine (SVM) based on a supervised training algorithm. The proposed system is divided in several phases, each performing a specific task. Two pre-processing schemes are employed in order to assess the efficiency in terms of recognition rate. Simulation results show that using Gabor filters in preprocessing for codifying the venous network and SVM for the classification can improve the recognition rate when compared to the existing methods.
Conference Paper
Full-text available
In this paper, a novel explicit image filter, called Guided Gabor filter, is proposed to extract the finger vein pattern without any segmentation processing, and lower performance reduction for poor quality images which result from low contrast, illumination, or noise effects, etc. The proposed filter is contributed for finger vein enhancement, noise reduction, and haze removal without being affected by the brightness of the vein. It performs well not only on ridge detection like the Gabor filter, but on image enhancement as an edge-preserving smoothing operator without the gradient reversal artifacts. The experimental results show that the proposed method is able to get vein pattern more clearly and faster than the existing methods, and improve the matching performance with higher recognition rate.
Article
Full-text available
Subspace learning methods are very sensitive to the illumination, translation, and rotation variances in image recognition. Thus, they have not obtained promising performance for palmprint recognition so far. In this paper, we propose a new descriptor of palmprint named histogram of oriented lines (HOL), which is a variant of histogram of oriented gradients (HOG). HOL is not very sensitive to changes of illumination, and has the robustness against small transformations because slight translations and rotations make small histogram value changes. Based on HOL, even some simple subspace learning methods can achieve high recognition rates.
Article
Full-text available
Finger veins have been proved to be an effective biometric for personal identification in the recent years. However, finger vein images are easily affected by influences such as image translation, orientation, scale, scattering, finger structure, complicated background, uneven illumination, and collection posture. All these factors may contribute to inaccurate region of interest (ROI) definition, and so degrade the performance of finger vein identification system. To improve this problem, in this paper, we propose a finger vein ROI localization method that has high effectiveness and robustness against the above factors. The proposed method consists of a set of steps to localize ROIs accurately, namely segmentation, orientation correction, and ROI detection. Accurate finger region segmentation and correct calculated orientation can support each other to produce higher accuracy in localizing ROIs. Extensive experiments have been performed on the finger vein image database, MMCBNU_6000, to verify the robustness of the proposed method. The proposed method shows the segmentation accuracy of 100%. Furthermore, the average processing time of the proposed method is 22 ms for an acquired image, which satisfies the criterion of a real-time finger vein identification system.
Conference Paper
We study the question of feature sets for robust visual object recognition, adopting linear SVM based human detection as a test case. After reviewing existing edge and gradient based descriptors, we show experimentally that grids of Histograms of Oriented Gradient (HOG) descriptors significantly outperform existing feature sets for human detection. We study the influence of each stage of the computation on performance, concluding that fine-scale gradients, fine orientation binning, relatively coarse spatial binning, and high-quality local contrast normalization in overlapping descriptor blocks are all important for good results. The new approach gives near-perfect separation on the original MIT pedestrian database, so we introduce a more challenging dataset containing over 1800 annotated human images with a large range of pose variations and backgrounds.
Article
To carry out the finger vein recognition quickly and effectively, an algorithm of finger vein recognition is proposed according to the characteristics of bi-direction two-dimensional principal component analysis ((2D)2PCA) reducing the dimensions. The algorithm is bi-direction weighted (2D)2PCA with eigenvalue normalization one ((OW2D)2PCA) based on preprocessing image of the figure vein image. The effect of the rate of cumulate eigenvalue on (2D)2PCA is analyzed, and the effect of the weighted value, the weighted value with eigenvalue normalization one and the rate of cumulate eigenvalue on W(2D)2PCA, OW(2D)2PCA, (W2D)2PCA and (OW2D)2PCA are analyzed as well. Experimental results on our database of finger vein images show that the presented method achieves high recognition accuracy. The redundant information of eigenvectors extracted by (2D)2PCA is restrained strongly, and the bi-direction weighted effect is better than the one direction weighted effect. The average recognition rate of (OW2D)2PCA is higher than those of 2DPCA, (2D)2PCA, W(2D)2PCA, (W2D)2PCA and OW(2D)2PCA.
Article
To improve the single-feature finger vein recognition, multi-feature fusion technology can be utilized not only to improve recognition rate, but also reduce false accept rate. In this paper a new fusion algorithm based on Fisher discriminant criterion is proposed for finger vein recognition. In the proposed algorithm, the forward mean Hausdorff distance (FMHD) and reverse mean Hausdorff distance (RMHD) between the feature points of the enrolled finger vein and those being matched are first computed. Then Fisher criterion is employed to determine the fusion parameters of FMHD and RMHD, with which FMHD and RMHD are fused to generate a new matching score for recognition finger vein. To improve the recognition rate, we further integrate three finger veins (i.e. index finger, middle finger and ring finger) using the proposed new matching score. Experiments demonstrate that the fusion of FMHD and RMHD has a better performance than traditional FMHD matching score, and the integration of three finger veins leads to a significant decrease of false accept when compared to the single finger vein recognition.
Article
Discriminant locality preserving projection(DLPP) can not obtain optimal discriminant vectors which utmostly optimize the objective of DLPP. This paper proposed a Gabor based optimized discriminant locality preserving projections (ODLPP) algorithm which can directly optimize discriminant locality preserving criterion on high-dimensional Gabor feature space via simultaneous diagonalization, without any dimensionality reduction preprocessing. The proposed method is applied to face and finger vein recognition problems and is compared with some other related Gabor based dimensionality reduction techniques. Experimental results conducted on the VALID face database and a subset of PKU finger vein database indicates the effectiveness of the proposed algorithm.
Article
With recent increases in security requirements, biometrics such as fingerprints, faces, and irises have been widely used in many recognition applications including door access control, personal authentication for computers, Internet banking, automatic teller machines, and border-crossing controls. Finger vein recognition uses the unique patterns of finger veins to identify individuals at a high level of accuracy. This article proposes a new finger vein recognition method using minutia-based alignment and local binary pattern (LBP)-based feature extraction. Our study makes three novelties compared to previous works. First, we use minutia points such as bifurcation and ending points of the finger vein region for image alignment. Second, instead of using the whole finger vein region, we use several extracted minutia points and a simple affine transform for alignment, which can be performed at fast computational speed. Third, after aligning the finger vein image based on minutia points, we extract a unique finger vein code using a LBP, which reduces false rejection error and thus the equal error rate (EER) significantly. Our resulting EER was 0.081% with a total processing time of 118.6 ms. © 2009 Wiley Periodicals, Inc. Int J Imaging Syst Technol, 19, 179–186, 2009