Professional Documents
Culture Documents
Face Detection Based On Viola-Jones Algorithm Applying Composite Features
Face Detection Based On Viola-Jones Algorithm Applying Composite Features
83
difference can be avoided effectively and the accuracy of
yH H T H
(WOpt ) x face recognition can be guaranteed. Figure 4 is a classifier
flow based on the Viola-Jones algorithm using composite
yL L T L
(WOpt ) x features.
(3)
Then the basic vector
YPool [Y1 ,!, Y2(c1) ]T
[ y1H , y2H ,! yCH1 , y1L , y2L ,!, yCL1 ]T is constructed
H L
by using the vector y , y obtained by NLDA. The
weight information of distinguishing features is used to
evaluate the validity of basic features in face recognition.
Then the feature with larger weight is selected to form the
composite feature vector. The discriminant weights of
each basic feature are measured by discriminant distance.
For Y j , it is the j-th feature of Y Pool . Its within-class
j j
distance ( DW ) and between-class distance ( DB ) can
be expressed by the following formulas (4), (5).
2 ( C 1) c
1
¦ ¦ N ¦ (Y
2
D
W
j
j
i
Y ji )
j 1 i 1 i Y Pool ci
(4)
2 ( C 1) c
Ni
¦ ¦ N (Y
Figure 4 A classifier using composite features based on
DBj j
i
Yj )2 Viola-Jones algorithm
j 1 i 1
(5) 3 Experimental results and analysis
i FDDB is the test data set used in this experiment.
Y j , Y j defines the mean value of the features
This data set contains 2,845 images with a total of 5,171
belonging to the class ci in the basic feature vector and faces. The face images in this data set have angle,
the mean value of the each feature in the basic feature illumination and scene changes.
vector. The discriminant distance of the j-th feature can be In the experiment, we randomly selected 100 images
expressed by equation (6): from the FDDB data set for face detection. And we count
the number of faces actually included in the selected 100
Fj DWj E DBj pictures, the number of false detection and missed
(6) detection. On this basis, a total of 10 experiments were
The discriminant distance of the j-th feature is used conducted and 1000 pictures were used as experimental
data. By comparing the composite feature method based
to measure the difference of each feature information. E
on Viola-Jones algorithm with the original Viola-Jones
is a parameter determined by the sample distribution. By algorithm in experiments, the accuracy of textual method
comparing the effects caused by different E , we finally used in the face recognition process is illustrated. And the
chose 2 as the value of E . The discriminant distance of phenomenon that the rigid object in the previous picture
affects the experimental effect is improved.
each feature in formula (6) is stored in the discriminant Table 1, table 2 and table 3 as shown below,are the
distance vector of the same size as Y Pool , namely specific experimental results on the FDDB data set of the
T original Viola-Jones algorithm and the application of
F [ F1 , F2 ,! , F2 ( c 1) ] . Then, according to the composite feature method based on Viola-Jones algorithm
discriminant distance Fj of each feature, a composite in this paper respectively .
feature vector (yCF) is constructed by selecting features
with larger Fj value. Finally, the constructed composite Table 1 Experimental results of original Viola-Jones
features are input into the classifier for face recognition. algorithm on FDDB data set
Experim Actual Missing False face Missin Error
2.2 Classifier Based on Viola-Jones Algorithms Applied ent face face number g detect
Compound Features number number number detecti ion
The original Viola-Jones algorithm does not handle on rate rate
the huge differences caused by the variation of different 01 126 16 3 0.127 0.024
rigid objects. By using the method of this paper, this 02 150 34 3 0.227 0.020
03 128 23 1 0.180 0.008
84
04 134 19 0 0.142 0 original face effectively, which makes a good foundation
05 152 22 1 0.145 0.007 for ensuring the accuracy of face recognition.
06 153 16 4 0.105 0.026
07 131 18 4 0.137 0.031
REFERENCES
08 126 14 1 0.111 0.008
09 147 18 2 0.122 0.014 [1] Liu Yihui, Bai Li. Salient Information of Mass Spectra of
10 125 9 2 0.072 0.016 Prostate Cancer Dataset[C]. Proc. of IEEE GRC’08. Hangzhou,
China:[s. n.], 2008
[2] C.Sagonas,G.Tzimiropoulos,S.Zafeiriou,and M.Pantic, “A
Table 2 Experiments of using composite feature semiautomatic methodology for facial landmark annotation,” in
based Viola-Jones method on FDDB data set Proc. CVPR,Jun. 2013, pp. 896–903.
Experim Actual Missing False face Missi Error [3] Z. Zhang, P. Luo, C. C. Loy, and X. Tang, “Learning deep
ent face face number ng detect
number number number detecti ion
representation for face alignment with auxiliary attributes,”
on rate IEEE Trans. Pattern Anal. Mach. Intell., vol. 38, no. 5, pp.
rate 918–930, May 2016
01 126 5 1 0.040 0.008 [4] Wu Kai, Zhu Hengliang, Hao Yangyang, Malizhuang.
02 150 8 0 0.053 0 Cascade regression multi-pose face registration [J]. Chinese
03 128 8 0 0.063 0 Journal of Image and Graphics. 2017, 22 (2): 0257-0264
04 134 6 0 0.044 0 [5] Bai Xiaoming, Wang Chengzhang, Shiqin. Color face
05 152 10 0 0.065 0 recognition based on two-dimensional linear discriminant
06 153 10 0 0.065 0 analysis [J]. Journal of Beijing University of Technology. 2010,
07 131 7 0 0.053 0 12 (36): 1717-1721.
08 126 8 0 0.063 0 [6] Wang Hongyong, Liao Haibin, Duan Xinhua, Ding Miao.
09 147 9 0 0.061 0 Adaptive Face Recognition Based on Singular Value and
10 125 4 0 0.032 0
Feature Fusion Matrix [J]. Computer Engineering and
Application, 2010, 46 (7): 162-174.
Table 3 Comprehensive comparison of the two [7] Bai Xiaoming, Wang Chengzhang, Shiqin. Color face
algorithms recognition based on two-dimensional linear discriminant
Total Total Total Total Gross analysis [J]. Journal of Beijing University of Technology, 2010,
Algorithm
name
face missing error missing error 36 (12): 1717-1721.
number count count rate rate [8] Wei Chongchong, Wang Xiaopeng, Yan Jianwei. A Face
Viola-Jone Recognition Method for Eliminating Illumination Effects [J].
189 21 0.138 0.015
s Electronic Test, 2012, (7): 19-23.
1372
Article [9] Wang Shoujia. Research on Image-based human detection,
75 1 0.055 0.0007
method tracking and face recognition [D]. Jilin University, 2013.
[10] Ren S, Cao X, Wei Y, et al. Face alignment at 3000 fps via
4 Conclusion regressing local binary features[C]//Proceedings of the IEEE
By analyzing the experimental results in the previous Conference on Computer Vision and Pattern Recognition. 2014:
section, we can know that the original Viola-Jones 1685-1692.
algorithm has a lot of missed and false face detections on [11] Yi D, Lei Z, Li S Z. Shared representation learning for
the FDDB data set. However, the method used in this heterogenous face recognition[C]//Automatic Face and Gesture
paper is obviously lower than Viola-Jones algorithm in Recognition (FG), 2015 11th IEEE International Conference and
Workshops on. IEEE, 2015, 1: 1-7.
terms of both missed detection rate and false detection [12] Zhu J Y, Zheng W S, Lai J H, et al. Matching nir face to vis
rate. From this, we can see that the combination of face using transduction[J]. Information Forensics and Security,
Viola-Jones algorithm and the constructed composite IEEE Transactions on, 2014, 9(3): 501-514.
features avoids the missed and false detection of the
85