多径中心化二进制模式直方图投影提升人脸识别性能
75 浏览量
更新于2024-08-27
收藏 321KB PDF 举报
本文档探讨了一种改进的面部识别方法，即"多径中心化二进制模式直方图投影"（Multi-radius Centralized Binary Pattern Histogram Projection）。传统的局部二进制模式（Local Binary Pattern, LBP）在人脸识别中存在一些问题，如特征向量维度较高、鉴别力较弱以及对噪声敏感。为了克服这些局限性，作者提出了一种新型的特征提取技术——中心化二进制模式（Centralized Binary Pattern, CBP）。
CBP操作通过集中处理像素邻域内的二进制信息，显著降低了直方图的维度，提高了特征表达的效率。这种操作方式减少了冗余信息，增强了特征的区分度，使得算法在处理噪声环境下的性能得到了提升。此外，为了进一步提高识别的准确性和速度，作者采用了多径CBP直方图作为人脸的表征，并将其投影到保持局部结构的主成分分析（Locality Preserving Projections, LPP）空间中。这样做可以进一步降低特征维度，有利于后续的高效处理和分类。
在实验部分，研究者将提出的多径CBP直方图投影方法应用到了FERET和CAS-PEAL两个知名的人脸识别数据库上。对比实验结果显示，新方法在识别精度和速度上都优于现有的LBP和其他传统方法，证明了其在实际场景中的有效性。这篇论文提供了一种有效的解决方案，通过结合CBP特性和LPP，实现了更加鲁棒且高效的面部识别系统。这对于计算机视觉和生物特征识别领域具有重要意义，特别是在实时和大规模数据处理的应用场景中。
June 10, 2009 / Vol. 7, No. 6 / CHINESE OPTICS LETTERS 475
Multi-radius centralized binary pattern histogram
projection for face recognition
Xiaofeng Fu (付付付晓晓晓峰峰峰)
1∗
and Wei Wei (韦韦韦 巍巍巍)
2
1
Institute of Computer, Hangzhou Dianzi University, Hangzhou 310018, China
2
Institute of Electrical Engineering, Zhejiang University, Hangzhou 310027, China
∗
E-mail: saqieer98@yahoo.com.cn
Received July 16, 2008
The existing local binary pattern (LBP) operators have several disadvantages such as rather long his-
tograms, lower discrimination, and sensitivity to noise. Aiming at these problems, we propose the central-
ized binary pattern (CBP) operator. CBP operator can significantly reduce the histograms’ dimensionality,
offer stronger discrimination, and decrease the white noise’s influence on face images. Moreover, for in-
creasing the recognition accuracy and speed, we use multi-radius CBP histogram as face representation
and project it onto locality preserving projection (LPP) space to obtain lower dimensional features. Ex-
p eriments on FERET and CAS-PEAL databases demonstrate that the proposed method is superior to
other modern approaches not only in recognition accuracy but also in recognition speed.
OCIS codes: 100.5010, 100.2000, 100.2960.
doi: 10.3788/COL20090706.0475.
In recent years, many approaches have been developed
for face recognition, including Gabor wavelet
[1,2]
, prin-
cipal component analysis (PCA)
[3]
, linear discriminant
analysis (LDA)
[4]
, manifold
[5]
, and local binary pat-
tern (LBP)
[6]
. Gabor wavelet-based approaches have
the drawback of expensive computation and thus are
not appropriate to construct the fast and efficient face
recognition system. The approaches based on PCA and
LDA preserve the global structure of image space, while
manifold-based methods preserve the local structure of
image space. However, local structure is more important
than global structure in real-world applications. LBP-
based approaches are arousing researchers’ high attention
due to the advantages of simple computation, robustness
to illumination variation, and discriminative ability.
Nevertheless, the existing LBP operators have several
unsatisfactory aspects. Firstly, they produce rather long
histograms, which slow down the recognition speed es-
pecially on large-scale face database. Secondly, under
some certain circumstances, they miss the local struc-
ture as they do not consider the effect of the center
pixel. Thirdly, the binary data produced by them are
sensitive to noise. Therefore, we propose the central-
ized binary pattern (CBP) operator which overcomes the
above shortcomings of existing LBP operators. In detail,
CBP operator decreases the histograms’ length largely
by comparing pairs of neighbors in the operator. Be-
cause of taking the center pixel into consideration and
giving it the largest weight, CBP operator’s discrimina-
tion is improved. CBP operator is insensitive to noise
owing to its modified sign function. Furthermore, in or-
der to improve the recognition accuracy, we use multi-
radius CBP histogram (MCBPH) as face representation.
Out of consideration for recognition speed, we do not
choose Chi square statistic
[6]
, as used by existing LBP-
based method, but project MCBPH onto locality preserv-
ing projection (LPP)
[5]
space and finish the classification
in this low-dimensional space. In this way, the proposed
method’s recognition rate is improved further due to
LPP’s powerful discrimination. The approach of multi-
radius CBP histogram projection (MCBPHP) has many
advantages such as significant dimensionality reduction,
more powerful discrimination, insensitivity to noise, and
high recognition speed. Experiments on two well-known
large-scale face databases show that the proposed method
outperforms the existing LBP-based method not only in
recognition rate but also in recognition speed.
The conventional LBP operator
[7]
is shown as
LBP(M, R) =
M−1
X
m=0
s(g
m
− g
c
)2
m
, (1)
in which R is the radius, s(x) =
½
1, x ≥ 0
0, x < 0
, g
c
repre-
sents the center pixel, and g
m
(m=0,
...
, M −1) are the
neighbors of g
c
. The image pixels are first labeled by
thresholding the difference between g
c
and g
m
using the
sign function s(x). The concatenation of the neighboring
labels is then used as a unique descriptor for each pat-
tern.
The patterns are uniform if the transitions between “0”
and “1” are less than or equal to two. The histogram
of the uniform patterns in the whole image is used as
the feature vector. For efficient face representation, the
extracted feature should also retain spatial information.
Fig. 1. (a) A face image divided into 8×8 small regions, (b)
weight set for different regions. Black squares indicate weight
0.0, dark gray 1.0, light gray 2.0, and white 4.0.
1671-7694/2009/060475-04
c
° 2009 Chinese Optics Letters
下载后可阅读完整内容，剩余3页未读，立即下载
2021-02-12 上传
2009-07-18 上传
2021-02-09 上传
2021-05-16 上传
2021-07-07 上传
2021-03-18 上传
2021-03-16 上传
2010-05-26 上传
2021-03-11 上传
