3.2 Modified probabilistic neural network

advertisement
A Name Recommendation Photo Album Using Probability Neural Network
Ding-Hrong Chen and Shiang-En Tang
Department of Computer Science and Information, Southern Taiwan University, TAIWAN
chendh@mail.stut.edu.tw, m95g0223@webmail.stut.edu.tw
Abstract
In this paper, we propose a name recommendation
mechanism used in the photo albums. We apply a light
compensation method to improve the image quality,
and then a face detection algorithm is applied to detect
the face region in each image. We adopt the modified
probabilistic neural network (MPNN) to train the
sample faces for each individual’s photos. After
training, the new adding face photos can be
recognized and a name tag is recommended to this
face. The experiments show a promising result for an
accurate name tag annotation. The proposed method is
very appropriate to provide a name recommendation
function applied in photo album.
1. Introduction
2. The face recommendation architecture
The proposed system architecture is shown in Fig. 1.
The proposed system contains two steps: the first is the
pre-processing step and the second is the name
recommendation system. For the training purpose, the
face positions in each photo are obtained by user
selection. A lighting compensation is used as a preprocessing step to normalize the brightness distribution
to adjust each training photo to the same contrast range
[6]. We applied the OpenCV library to detect the
human faces. A standard face detection technologies
for facial image normalization [10][2] is also used for
further refinement. After pre-processing, we will use
the PNN to train each sample faces. By the PNN
learning and recalling process, the training samples are
used for the face detection phase.
Pre-Processing
Due to the mass amount population of digital photos,
the face detection and face recognition technology
[5][3] has been highly applied in many different fields.
If a face detection and name tag annotation mechanism
can be added into web photo albums, it will be more
convenient for users to search a specific photo via
different ways. The face recognition technology is also
used in public transportation safety monitoring in 2008
Beijing Olympic[9].
In this paper, we present a new method that locate
and annotate the human faces in each photos. The
modified probability neural network (MPNN) approach
is used to train the annotated example faces. After
training, if a new photo is added to the album, the
system will detect the new face, and gives this face a
name tag from the existed database.
This paper is organized as follows. In Section 2, we
describe the proposed system architecture. The face
recommendation method that obtains the new name tag
and their ranking is presented in Section 3. In Section
4, we present the training result of the proposed
method. The conclusion and future work are described
in Section 5.
Sample Face
Images
Light
Compensation
Normalized
Sample Faces
Face Detection
Name Tags
Learning
New Adding
Images
Recalling
Name
Recommendation
Modified PNN
Figure 1. System flowchart
3. Face recommendation algorithms
The above described pre-processing step can obtain
the sample faces, and this result will be treated as the
PNN input data. According the different user
categories, we give the name tags for these sample face
data. When a new photo is added to the album, a
proper name will be assigned to this photo.
3.1 Probabilistic neural network (PNN)
hi ci f i ( X )  h j c j f j ( X )
 j  i,
The probability neural network is proposed by D. F.
Specht for solving the problem of classification in 1988
[1]. The theoretical foundation is developed based on
Bayes decision theory, and implemented in a feedforward network architecture. The Bayesian decision
theory is the principle concept of classification that
uses statistical inference, mainly to find the maximum
expectation of the smallest risk.
The Bayes theory is used to solve the classification
problem in huge amount raw data, but a problem is
occurred that we need to know the probability density
function in advance. To solve this problem, Parzen
window method was proposed by Parzen in 1962[4].
Parzen uses a feature vector to estimate a class
distribution; the training data for each sample is
represented by an eigen-value as the center of the
Gaussian curve. The data point is assigned to a specific
class by comparing the probability density function
values to each class.
The probability neural network architecture is a
four-neuron layers network model as shown in Fig. 1.
These four layers include the input unit, the pattern
unit, the summation unit and the output unit. The input
unit contains a number of input image data and the
pattern unit node number equal the number of training
data, the output unit represents the classification
results.
X1
X2
X3
where hk is the prior probability of class k; ck is the
Input Unit
With the above equation, X and W ip have been
normalized with their unit length:
 ( X  W ip ) t ( X  W ip ) 
Pi ( X )  exp 

2 2


The summation unit will be calculated from the
pattern unit by summarizing and averaging the output
of N i neurons to generate an output probability vector,
I1
P1
I3
I2
P2
P3
S1
P4
S2
O1
Pattern Unit
Summation
Unit
Output Unit
Fig. 1 The probabilistic neural network architecture
The raw data are divided into k classes, which
represent the k situations to be explained. Each class
has m-dimensional observations, i.e. X  [ X 1 , X 2 ,   X m ] .
The following Bayes decision rule is used to classify
the raw data into k classes:
representative center that belongs to the class k, but the
loss function has been a miscarriage of justice;
and f k is the probability density function of class k
The probability density function for each class can
be expressed as follow:
  ( x  xij ) t ( x  xij ) 
1
1 Ni
f i ( x) 

 exp 
d
N i j 1 
2 2

d
2
(2 ) 
where fi (x) denotes the dimension of training
vector,  is the smoothing parameter and d is the
dimension of training vector. N ij denotes the total
number of training vector in category i , and xij is the
neruon vector and x is test vector.
When a new input image X is added, the pattern
unit will be calculated with the individual weight vector
W ip of the product to make a nonlinear transformation
to Z i :
Z i  X W ip .
The probability of neural network with back
propagation networks can approximate any continuous
nonlinear function. In this paper, we use the Gaussian
function as the activation function:
 Z 1 
P( Z i )  exp ( i 2 )
 

i.e.
Si ( X ) 
1
Ni
 ( X  W ip ) t ( X  W ip ) 
exp



2 2
p 1


Ni
Finally, one or many larger values are chosen as the
output unit that indicates these data points are in the
same class via a competition transfer function from the
output of summation unit [7], i.e.
O( X )  arg max( Si ( X )), i  1,2,  , m
3.2 Modified probabilistic neural network
The PNN can process the classification problem
effectively, but it’s a hard clustering method. Zaknick
proposed a modified probabilistic neural network
(MPNN) architecture in 1995[8]. The major difference
between MPNN and PNN is that MPNN calculates the
output unit with a “centroid method” instead of the
“winner-take-all” method used in PNN. If f i (x) is the
Gaussian activation function, and
There are 5 users and their corresponding name tags
are used in this experiment. We divide the image
database into two sets. One is used for training, and the
other is used for recalling. The image sets are as shown
in Fig. 5.
yi is the
representative of pattern unit, then the output unit is
computed by:
M
Oi 
 Z y f ( x)
i 1
M
i
Z
i 1
i
i
i
f i ( x)
4. Experimental Results
We adopt the experimental samples of this study
from the HiAnts[11]. There are total 5110 photos in the
album; and 2697 photos have been annotated as the
training samples. The experimental steps are performed
with image pre-processing and face detection. Then the
modified probability neural network is used as the
training kernel. Fig. 2 shows a 16x16 standard face and
Fig. 3 is the result of the face detection using OpenCV
library. We also adopt the light compensation to make
the image quality homogenous as shown in Fig. 4.
Fig. 2 Standard image face
Fig. 3 Sample faces detection
(a)
(b)
Fig. 4. (a) The images before light compensation (b)
The images before light compensation
(a)
(b)
Fig. 5. (a) The training image set (b) The images
used in the recalling experiments
We randomly pick 10 training samples of 5
categories in the training image set to train the MPNN,
and then chose other un-recommended 50 images in the
recalling image set to test the name recommendation
performance. We repeat this process 5 times and
compute its average recalling rate as shown in Table. 1.
Table 1 shows the recalling rate for a new testing face
image is more than 70%.
Table 1. The recalling success rate
user
no.
A
B
C
D
E
1
11
8
9
12
10
2
8
11
11
10
10
3
8
7
10
11
14
4
9
11
6
10
14
5
8
12
9
10
11
error
15
14
12
11
7
Recalling
rate
70%
72%
76%
78%
86%
We perform another experiment to validate the
recognition accuracy. For each specific user, we
selected arbitrary 20 images for testing. We want to
know the accuracy rate of a testing image is annotated
correctly to its category. The result is shown in Table 2.
Table 2. The recognition accuracy for 5 categories
user
A
B
C
D
E
Accuracy
A
0.78
0.70
0.73
0.66
0.75
75%
B
0.68
0.86
0.72
0.75
0.76
100%
C
0.72
0.75
0.76
0.73
0.75
30%
D
0.63
0.74
0.70
0.78
0.70
85%
E
0.73
0.74
0.73
0.69
0.79
75%
The result in Table 2 shows that the user’s photo has
better recognition accuracy in his own category, but the
difference is not significant comparing with other
categories. For example, the recognition accuracy of
user C is almost equal in every category, which will
make a miss classification in the recalling. One of the
possible reason is the facial expression of each users
makes the training performance is not so good.
We also test the different face sizes used in training.
Table 3 shows the size effects for the recalling
experiment.
Table 3. Recommend result of different sizes
Training
Test
Correct
Size
samples
samples
Rate
16×16
50
50
41
[2] G.T. Suen,” A Study on Personal Identification by Face
Recognition Based on Neural Networks” Master Thesis,
National Kaohsiung First University of Science and
Technology,Taiwan,2003.
[3] Mpiperis, I.,Malassiotis, S.,Strintzis, M.G., “3-D Face
Recognition With the Geodesic Polar Representation”, IEEE
Trans., Information Forensics and Security, Vol. 2, pp 537 –
547, Sept. 2007.
[4] Parzen, E, “On estimation of a probability density
function and mode.” Annals of Mathematical Statistics, Vol.
33, pp. 1065-1076, 1962.
[5] Pentland, T. Choudhury, “Face Recognition for Smart
Environments,” Computer, Vol. 33, no. 2, pp. 50-, Feb.
2000.
[6] T.S. Hung, “A Smart Digital Surveillance System with
Face Tracking and Recognition Capability” Master Thesis,
University of chung yuan christian,Taiwan,2004.
Accuracy
82%
32×32
50
50
38
76%
64×64
50
50
37
74%
5. Concluding remarks
In this paper, we proposed a MPNN method for
human face name recommendation used in album. A
light compensation method is performed to adjust the
image contrast distribution to provide a much more
homogenous training patterns. We use a number of
standard faces with name tags to train the MPNN. After
training is complete, the MPNN can detect other
images with name tags and suggest a proper name to
this face.
The MPNN owns several advantages as the training
kernel. The training speed is fast and that’s make it
suitable for a large amount of image dataset. The
choice of the sample faces will also affect the training
result. Although this study is still need to improve its
consistence and integrity, the proposed method is very
appropriate to provide a name recommendation
function applied in photo album.
6. References
[1]D. F. Specht, “Probabilistic Neural Network for
Classification, Mapping, or Associative Memory,”
Proceeding of the IEEE International Conference on Neural
Networks, Vol. 1,pp.525-532,1988.
[7] Whei-Min Lin, Chia-Hung Lin, and Zheng-Chi Sun,
“Adaptive MultipleFault Detection and Alarm Processing for
Loop System with Probabilistic Network,” Power delivery,
IEEE Transactions on, Vol.19, Issue:1, Jan. 2000 Pages:6469.
[8] Zakinch, A. “Introduction to the Modified Probabilistic
Neural Network for general signal processing applications”
IEEE Transactions on Signal Processing, 1998, Vol. 46, pp.
1980-1990.
[9] http://www.eepw.com.cn/article/62775.htm
[10] http://pages.cs.wisc.edu/~beechung/facedet/report.html
123
[11] http://hiants.tw
Download