2008 Congress on Image and Signal Processing

A New Image Enhancement Method for Face Recognition

with Single Training Sample

Jiazhong He1, Di Zhang1,2

1
School of Information Engineering, Shaoguan University
2
School of Computer Science＆ Engineering, South China University of Technology

Abstract training. In [1] and [2], face recognition algorithms are

proposed to solve the problem of single training
At present there are many methods that could deal images. Instead of focusing on the feature extraction or
well with frontal view face recognition when there is the learning stage of face recognition systems, these
sufficient number of representative training samples methods focus on the preprocessing stage. (PC)2A [1]
per person. However, few of them can work well when combines the original training image with its vertical
only single training sample per person is available. In and horizontal projections and then performs principal
this paper, we present a training sample enhancement component analysis (PCA) on the combined training
method based on Wavelet Transform (WT) low- images set. SPCA [2] combines the original training
frequency band to improve the performance of face image with its derived image gotten by perturbing the
recognition with a single training sample. In order to image matrix's singular values and then performs PCA
enhance the classification information of the single on the joined training images set. These algorithms
training sample, each training sample is combined achieve higher recognition accuracy with far fewer
with its WT reconstructed image based on low- eigenfaces than the standard eigenface algorithm on the
frequency band into a new training sample. By using FERET database when only one training image per
Fourier transform, the resulting spectrum person is available.
representation of face image is used as the feature of Principal component analysis (PCA) is a well-
the face image for recognition. The experimental known data representation technique widely used in
results on ORL face database indicate the effectiveness pattern recognition and signal processing. However
of the proposed method. image registration is required in PCA-based approach.
This approach gives a very good result if the face to be
1. Introduction recognized is aligned properly. Lai et al. [3] have
demonstrated that the recognition accuracy using PCA
With the growing interest in the development of falls from 85.5, 81.7, 76.7, and 70.0-67.5% if the
human and computer interface and biometric alignment is deviated from 0 to 4 pixels error.
identification, automatic recognition of faces is Fourier transform is a traditional and effective tool
considered as one of the fundamental problems in of signal analysis. It has been widely used in face
computer vision and pattern analysis, and many recognition. Fourier transform is invariant against
scientists from different areas have addressed it. Great spatial translation. In the process of face recognition,
progress has been made in the area of machine applying the 2D discrete Fourier transform to the face
recognition and many novel methods have been image and using the resulting Fourier spectrum instead
introduced in past 20 years. of the spatial data from the face image, the translation
Face recognition is a hard problem because there error which caused by the spatial position offset can be
are many image variations in real-life such as lighting, eliminated entirely and as a result, the recognition
illumination, facial expression, partial occlusion and accuracy will be improved.
imprecise localization of face area. The typical To overcome the alignment problem in template-
approach in handling these variations is to use large based approach and enhance the classification
and representative training samples set. However if the information of the single training sample, along the
training set is small, the performance of these line of [1] and [2], instead of focusing on the feature
algorithms decreases. Typically, for law enforcement extraction or the learning stage of face recognition
applications only a single face sample is available for system, this paper focuses on the preprocessing stage.

978-0-7695-3119-9/08 $25.00 © 2008 IEEE 216

DOI 10.1109/CISP.2008.715
We present a training sample enhancement method catches less attention. This paper focuses on the
based on Wavelet Transform (WT) to improve the preprocessing stage. Each original training image is
performance of face recognition with a single training combined with its WT reconstructed image based on
sample. Each original training image is combined with low-frequency band into a new training image.
its WT reconstructed image based on low-frequency For face image f ( x, y ) , let f r ( x, y ) is its WT
band into a new training image. Fourier transform is reconstructed image based on low-frequency band.
used for feature extraction, and then L1 nearest
Then we combine f ( x, y ) with f r ( x, y ) to generate a
neighbor classifier is employed for classification.
newly combined training image according to the
2. Enhancement of training image following equation:
f ( x, y ) + α ⋅ f r ( x, y )
fα ( x, y ) = (1)
Wavelet Transform (WT) has been proven effective 1+α
for image analysis and feature extraction. It represents where α is called combination parameter，in our
a signal by localizing it in both time and frequency experiments α is set to 0.25.
domains. WT divides an image into four different
frequency bands as shown in Fig.1. By further 3. Fourier transform
decomposing the low-frequency band, we can perform
the multi-resolution analysis, which has been widely Fourier transform is an effective tool for signal
adopted in image processing. analysis. It has been widely used in face recognition.
For 2D image f(x, y) with the size of N×M, 2D discrete
Fourier transform is defined as:
LL LH 1 M −1 N −1   ux vy  
F ( u, v ) = ∑ ∑ f ( x, y ) exp  −2π i  M + N   (2)
MN x = 0 y =0  
For u = 0,1,2," , M − 1 , v = 0,1,2,", N − 1 . Fourier
HL HH transform has the following property:
  ux vy  
f ( x − x0 , y − y0 ) ⇔ F (u , v) exp  − j 2π  0 + 0   (3)
 M N 
Figure 1. 1-level WT decomposition
Formula (3) indicates that Fourier transform is
invariant against spatial translation. In the process of
In Figure 1, the band LL is the low-frequency band
face recognition, applying the 2D discrete Fourier
of the original image. The bands LH and HL record the
transform to the face image and using the resulting
changes of the image along horizontal and vertical
Fourier spectrum instead of the spatial data from the
directions. While the HH band shows the high
face image, the translation error which cause by the
frequency component of the image. Nastar et al. [4, 5]
spatial position offset can be eliminated entirely and as
have investigated the relationship between variations in
a result, the recognition accuracy will be improved.
facial appearance and their deformation spectrum.
They found that facial expressions and small occlusion
affect the intensity manifold locally. Under frequency- 4. Outline of the proposed approach
based representation, only high-frequency spectrum is
affected, called high-frequency phenomenon. The main processing steps of the proposed
Moreover, changes in pose or scale of a face affect the algorithm may be summarized as follows:
intensity manifold globally, in which only their low- Step 1: normalize the intensity of the face image.
frequency spectrum is affected, called low-frequency after normalized, each original training image is
phenomenon. Only a change in face will affect all combined with its WT reconstructed image based on
frequency components. The above statements show low-frequency band into a new version of the original
that the band LL is insensitive to the facial expressions training image.
and small occlusion. Therefore, we focus on the WT Step 2: In the training stage, 2D discrete Fourier
reconstructed image of the 3-level low-frequency band transform is applied to the newly combined training
in this paper. image and the resulting Fourier spectrum is used as the
According to Jain [6], a pattern recognition system feature of training image for recognition.
includes three modules, i.e. the preprocessing stage, Step 3: In the recognition stage, 2D discrete Fourier
the feature extraction, and the learning stage. The first transform is applied on the normalized test image and
stage is important to recognition system but often the resulting Fourier spectrum is used as the feature of
the test image for recognition.

217
 Step 4: The L1 distance is used to measure the
similarity between the two feature vectors. We can see from Figure 3 that as α gradually
increases, the recognition accuracy increases at the
5. Experiment beginning and decreases at the end. This may suggest
that choosing an appropriate value for α is important
We use the ORL database (http://www.cam- to the recognition performance. According to Figure 3,
orl.co.uk) that contains a set of faces taken between the proper value of α lies in the range of
April 1992 and April 1994 at the Olivetti Research 0.15 < α < 0.35 .
Laboratory in Cambridge, UK. There are 10 different
images of 40 distinct subjects. For some of the
subjects, the images were taken at different times, 100
varying lighting slightly, facial expressions 90
(open/closed eyes, smiling/non-smiling) and facial 80
details (glasses/no-glasses). All the images are taken
70
against a dark homogeneous background and the

rank 1 accuracy (%)

subjects are in up-right, frontal position (with tolerance 60
for some side movement). The size of each image is 92 50
×112, 8-bit grey levels. One of the persons is shown 40
in Figure 2. In this paper experiments were performed
30
with one normal training image and 9 test images for
each person. 20

10

0
-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
combination parameter α

Figure 3. Accuracy versus combination parameter

6. Conclusion
In this paper, we present a training image
Figure 2. One person in the ORL database
enhancement method based on WT to solve the single
training sample problem. To enhance the classification
Comparative experimental results are shown in
information of the single training sample, each original
Table1. Here, α is set to 0.25. We can see from Table
training image is combined with its WT reconstructed
1 that the proposed training image enhancement
image based on 3-level low-frequency band into a new
method gives 14% improvement in recognition
training image. The experimental results on ORL face
accuracy over the result based on the original training
database indicate the effectiveness of the proposed
image.
method.
The proposed method introduces the combination
parameter α . In order to know the influence of
combination parameter α on the recognition
7. References
performance, we perform some experiments based on
[1] J.Wu, Z.-H, Zhou, “Face recognition with one training
the combined training image with different values of image per person”, Pattern Recognition Letters, 23(14), 2002,
α . The rank1 accuracy is shown in Figure 3. Here, pp.1711-1719.
WT reconstructed image is based on 3-level low-
frequency band. [2] D. Zhang, S. Chen, and Z.-H. Zhou, “A new face
recognition method based on SVD perturbation for single
Table 1. Recognition accuracy comparison example image per person”, Applied Mathematics and
Computation, 163(2), 2005, pp.895-907.
Training image Rank 1(%) Rank 3(%)
[3] J.H. Lai, P.C. Yuen, G.C. Feng, “Spectroface: a Fourier-
Combined 85.83 89.72 based approach for human face recognition”, in: Proceeding
of the Second International Conference on Multimodal
Original 71.67 83.61 Interface, Vol. 2, Hong Kong, 1999, pp. VI 115-120.

218
[4] C.Nastar, B.Moghaddam, A.Pentland, “Flexible images: [6] A.K. Jain, R.P.W. Duin, J.C. Mao, “Statistical pattern
matching and recognition using learned deformations”, recognition: a review”, IEEE Transactions on Pattern
Comput. Vision Image Understanding, 65(2), 1997, pp.179- Analysis and Machine Intelligence, 22(1), 2002, pp.4-37.
191.
8. Acknowledgement
[5] C.Nastar, N.Ayache, “Frequency-based non-rigid motion
analysis”, IEEE Transactions on Pattern Analysis and
Machine Intelligence, 18(11),1996, pp.1067-1079. This work was supported by the National Natural
Science Foundation of China (Grant No: 60772117).

219