Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
© 2016, IJARCSSE All Rights Reserved Page | 106
Volume 6, Issue 12, December 2016 ISSN: 2277 128X
International Journal of Advanced Research in Computer Science and Software Engineering Research Paper Available online at: www.ijarcsse.com
Comparative Analysis of Illumination Compensation with Different
Histogram Equilization Algorithms and Its Face Application Saba Tahseen (M.tech Student)
Dept. of Computer Science & Engineering,
Visvesvaraya Technological University,
Karnataka, India
Dr. Shubhangi D.C H.O.D of CS &Engineering,
Visvesvaraya Technological University,
Karnataka, India
Abstract— This paper proposes a face recognition and verification algorithms based on histogram equalization to
standardize the faces illumination reducing in such way the variations for further features extraction Illumination
compensation and normalization play a crucial role in face recognition. In this paper we are going to compare
histogram equalization algorithms. We claim that edge orientation is useful for face recognition. Three OLHE feature
combination schemes were proposed for face recognition: 1) encoded most edge orientations; 2) more compact with
good edge-preserving capability; and 3) performed exceptionally well when extreme lighting conditions occurred. The
proposed algorithm yielded state-of-the-art performance on AR, CMU PIE, and extended Yale B using standard
protocols. We further evaluated the average performance of the proposed algorithm when the images lighted
differently were observed, and the proposed algorithm yielded the promising results.
Keywords— Face recognition, illumination compensation, histogram equalization, extended yale database,Olhe
I. INTRODUCTION
FACE recognition remains a popular progress has been made, the difficulty posed by variations in illumination
or head orientation makes the task of general face recognition unsolved yet.
Biometrics consists of a set of automated methods for recognition or verification of individuals using physical or
behavioral characteristics of such; as face, fingerprint, signature, voice, etc. This technology is based on the fact that each
single person is unique and has distinctive features that can be used for identification Figure 1 shows the block diagram
of biometrics .
Fig 1. Block diagram illustrating different modules in a biometric system.
As shown in Fig. 2, different biometric modalities have been used in several applications based on its suitability
and user convenience. Fingerprint, face and iris are the most widely used biometric modalities. Other biometric
modalities that have gained sufficient popularity includes gait, hand geometry, palm-prints, and voice. In general, there is
no single biometric modality that is best for all applications. Moreover, in some applications, more than one biometrics is
used to attain higher security and to address failure to enroll. Such systems are called multimodal biometric systems.
Several factors have to be considered while designing any biometric system such as location, security risks, identification
or verification task, expected number of users and their characteristics. This section provides the review of some of the
most widely used and accepted biometric modalities.
Fig. 2. Biometric modalities that are widely used in several applications.
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 107
Face recognition has been a topic of active research since the 80‟s, proposing solutions to several practical
problems. Face recognition is probably the biometric method easier to understand, because we identify people by mainly
their faces. However the recognition process used by the human brain for identifying faces has not a concrete explanation.
Because it is now essential to have a reliable security systems in offices, banks, businesses, shops, etc. several
approaches have been developed, among them the face-based identity recognition or verification systems are a good
alternative for the development of such security systems.
In this paper we are going to compare the various histogram equilization algorithms They are:1)histogram
equalization 2) adaptive histogram equalization, 3)contrast limiting adaptive histogram equalization or CLAHE,4)local
histogram equilization and our proposed algorithms is 5) Oreinted local histogram equilization.
II. ALGORITHMS
On this section we present the different techniques used on the system proposed.
A) Histogram equalization:
The histogram manipulation, which automatically minimizes the contrast in areas too light or too dark of an
image, consists of a nonlinear transformation that it considers the accumulative distribution of the original image; to
generate a resulting image whose histogram is approximately uniform. On the ideal case, the contrast of an image would
be optimized if all the 256 intensity levels were equally used. Obviously this is not possible due to the discrete nature of
digital data of the image. However, an approximation can be achieved by dispersing peaks in the histogram of the image,
leaving intact the lower parts. This process is achieved through a transformation function that has a high inclination
where the original histogram has a peak and a low inclination in the rest of the histogram.
Histogram Normalization is one of the most commonly used methods for preprocessing. In image processing,
the idea of equalizing a histogram is to stretch and redistribute the original histogram using the entire range of discrete
levels of the image, in a way that an enhancement of image contrast is achieved. The most commonly used histogram
normalization technique is histogram equalization where one attempts to change the image histogram into a histogram
that is constant for all brightness values. This would correspond to a brightness distribution where all values are equally
probable. For image I(x,y) with discrete k gray values histogram is defined by i.e. the probability of occurrence of the
gray level i is given by:
Where i ∈ 0, 1…k −1 grey level and N is total number of pixels in the image. Transformation to a new intensity
value is defined by:
Output values are from domain of [0, 1].To obtain pixel values in to original domain, it must be rescaled by the
K−1 value. Fig.3 shows the histogram equalization.
The widespread histogram equalization cannot correctly improve all parts of the image. When the original
image is irregularly illuminated, some details on resulting image will remain too bright or too dark. These are most
commonly used techniques of histogram adjustment. HE is to create an image with uniform distribution over the whole
brightness scale and HS is to make the histogram of the input image have a predefined shape.
Fig 3. : An original image, its histogram, Linear histogram equalization from left to right.
B) Adaptive histogram equalization:
In histogram equalization, the goal is to obtain a uniform histogram for the output image, so that an “optimal”
overall contrast is perceived. However, the feature of interest in an image might need enhancement locally. Adaptive
Histogram Equalization (AHE) computes the histogram of a local window centered at a given pixel to determine the
mapping for that pixel, which provides a local contrast enhancement. However, the enhancement is so strong that two
major problems can arise: noise amplification in “flat” regions of the image and “ring” artifacts at strong edges.
The neighboring tiles are then combined using bilinear interpolation to eliminate artificially induced boundaries.
The contrast, especially in homogeneous areas, can be limited to avoid amplifying any noise that might be present in the
image approach we have taken the DCT coefficients in zigzag pattern as shown in Fig. 4.
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 108
The DCT is performed on the entire image obtained after processing the input face images by adaptive
histogram equalization and logarithm transform.
Fig 4. Block feature of DCT coefficients and their selection in zigzag pattern.
After applying adaptive histogram equalization, we employed logarithm transform [10] for further enhancement
of the image.
i) Discrete cosine transform:
The DCT is a popular technique in imaging and video compression, which transforms signals from the spatial
representation into a frequency representation. The forward 2D-DCT [6, 10, 16] of a M × N block image is defined as.
),()()(),( 1
0
1
0 yxfvuvuC N
y
M
x
N
vy
M
ux
2
)12(cos
2
)12(cos (2)
The inverse transform is defined as
),()()(
0
1
0
1),( vuCvu
v
N
u
Myxf
N
vy
M
ux
2
)12(cos
2
)12(cos (3)
where
x and y are spatial coordinates in the image block, u and v are coordinates in the DCT coefficients block. Fig.5 shows the
properties of the DCT coefficients in M × N blocks with the zigzag pattern used by JPEG [10] compression to process the
DCT coefficients. Although the total energy remains the same in the M × N blocks, the energy distribution changes with
most energy being compacted to the low-frequency coefficients. In our approach we have taken the DCT coefficients in
zigzag pattern as shown in Fig. 4.
The DCT is performed on the entire image obtained after processing the input face images by adaptive
histogram equalization and logarithm transform.
(a) (b) (c)
Figure 5. (a) Original Image with its histogram; (b) After histogram equalization; (c) After adaptive histogram
equalization.
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 109
C) Contrast Limiting Adaptive Histogram Equalization or Clahe:
The variants of Histogram Equilization (HE) Adaptive Histogram Equalization (AHE) and Contrast Limited
Histogram Equalization (CLHE).
AHE method divides the image into small contextual regions called „tiles‟ and HE is applied to all the regions.
The tiles are then stitched back using bilinear interpolation method. The AHE method increases the local contrast of the
image but it results in noise amplification. To limit the noise amplification, CLHE method clips the height of the
histogram based on threshold. Pisano et al. proposed CLAHE method for enhancing the low-contrast medical images
[16]. The Contrast Limited Adaptive Histogram (CLAHE) is a combination of AHE and CLHE. CLAHE divides the
image into small contextual regions called „tiles‟ and does histogram clipping based on threshold. The CLAHE method
enhances the local details of the image.
In Contrast Limited Histogram Equalization (CLHE), the histogram is cut at some threshold and then
equalization is applied. It improves the local contrast of image. In CLHE, since the peak of the histogram in each tile is
clipped, the amount of over amplification is avoided.
Contrast Limited Adaptive Histogram Equalization (CLAHE) [27] method uses the concepts of AHE and CLHE.
It uses small contextual regions called „tiles‟ and the peaks of histogram are clipped. It increases the local contrast of the
image and the noise amplification is avoided. The disadvantage of CLAHE method is it gives unsatisfactory results when
there is unbalanced contrast and increased brightness [54]. Thus, here Enhanced CLAHE method is proposed by
combining CLAHE with threshold technique.
i) Thresholding:
Thresholding is the one of the simplest, computationally faster method which is used in image processing for
image segmentation . Given image „f‟, the object can be extracted from its background with threshold „T‟ and create
image „g‟ using the Eq.(4).
The above method is called as global threshold. However if image has noise or illumination, global threshold
would not give good results and variable thresholding has to be applied.
Steps of CLAHE:
1) First, create an illumination invariant face image by combining CLAHE and thresholding. The image obtained is
termed as Enhanced Contrast Limited Adaptive Histogram Equalization (Enhanced CLAHE).
2) Second, the performance of the CLAHE and Enhanced CLAHE methods are compared. The face recognition of
the two methods is tested using the three public databases AR, Yale and ORL. The train database and test
database are created for testing the two methods. The images in the test database are matched with images in
train database using Fuzzy K Nearest Neighbour classifier. The face recognition accuracy rate percentage is the
number of correct matches in test database divided by the total number of images in test database. The face
recognition accuracy rate percentage of CLAHE and the proposed method „Enhanced CLAHE‟ are computed
for three public face databases AR, Yale and ORL.
3) The efficiency of the proposed method is compared with CLAHE by using the subspace projection method. The
subspace projection method „fisher face‟ is used here to extract the features from CLAHE and Enhanced
CLAHE images. The feature extraction step is done in both train and test database images. The features in test
database images are matched with features of train database image using cosine similarity measure. The face
recognition accuracy rate percentage, Equal Error Rate (EER), Half Total Error, False Acceptance Rate (FAR)
at 1%, 0.1% and 0.01% are calculated for both CLAHE and Enhanced CLAHE methods.
The results for Enhanced CLAHE in AR face database image are shown here. The results of pre-processing
steps as discussed in section 3.4.1 are shown in Fig.6(a)-Fig.6(d).
Fig.6. a) Colour image b) Gray c) Average filter d) Contrast Limited Adaptive histogram Equalization (CLAHE).
D) Local Histogram Equalization:
An extension to HE known as Local Histogram Equalization (LHE) has been introduced. The main idea of LHE
is to define a local transform function for each pixel based on its surrounding neighbouring pixels.
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 110
Fig 7. The basic concept of LHE operation
Generally, LHE uses a small window to define a Contextual Region (CR) for the center pixel of that window.
The relationship between the image and the window is illustrated in Fig. 9. Only the block of pixels that fall in this
window is taken into account for the calculation of CDF. Thus, as the window slides, the CDF is modified.
The CDF is the main contributor for the HE transform function. Hence, in LHE, the transform function of a
pixel is depending on the statistics of its neighbors in CR. Because the transform function changes as a response to the
changes in the contents of CR, LHE is also popularly known as Adaptive Histogram Equalization (AHE). Examples of
LHE methods are Non-Overlapped Block HE (NOBHE) , Block Overlapped HE (BOHE) and , Interpolated Adaptive
HE (IAHE) , Weighted Adaptive HE (WAHE) , Contrast Limited Adaptive HE (CLAHE) , Variable Region Adaptive
HE(VRAHE) , Local Information HE (LIHE) , Spatio- Temporally Adaptive HE (STAHE) , Partially Overlapped Sub-
Block HE (POSHE) , Conditional Sub-Block Bi-HE (CSBHE) [39], and Multiple Layers Block Overlapped HE
(MLBOHE) .
Since illumination generally contributes low-frequency component of images, and edges contributes high-
frequency component of images, some works tried to compensate the illuminated low-frequency image component. We
define the generalized LHE operator as:
(5)
where ξ, η is the relative position of the anchor point to the pixel to be processed, HWI is the input image whose
dimension is W-by-H, and HWI is the histogram-equalized image with the same dimension. The typical LHE which
uses the k-by-k local window can be denoted as 0,0
kL , since the anchor point is exactly the pixel to be processed itself. If
the pixel to be processed is brighter than all the neighbouring pixels around it, it will have a large intensity value after the
local histogram equalization, and vice versa. We can make LHE „oriented‟ by changing anchor positions. Fig. 10 shows
nine LHE operators using 3-by-3 windows.
Fig 7. Illustration of LHE
(a)
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 111
(b)
Fig 8. (a) Nine generalized LHE operators using a 3-by-3 window. (b) Illustration of anchors (blue circles) and the
corresponding pixels to be equalized (red „X‟s) on local search windows.
Comparison of the facial image in Fig. 9 after applying aO3 and
0,0
3L Note that the edges are enhanced using the
proposed aO3 .
E) Oriented Local Histogram Equalization:
In this section, we introduce a novel method of facial image preprocessing algorithm called Oriented Local
Histogram Equalization (OLHE). OLHE is similar to local histogram equalization (LHE), but it captures the orientation
of edges while LHE does not.We begin with a brief review on LHE. For each pixel on an image, we perform the
histogram equalization on the local w-by-h window centering on the pixel using.
(5)
where x is the pixel intensity value, cd f (x) is the cumulative distribution function of the histogram of the pixel intensities
in the w-by-h window, cd fmin is the minimum intensity in this window, and L is the desired number of output gray levels.
Typically a square window is used, and we define k ≡ w = h. We call the center of the k-by-k window the anchor. For
LHE, the anchor point is the pixel to be processed itself.
The eight operators with {ξ, η} other than {0, 0} are „oriented‟, and they are dubbed as the Oriented Local
Histogram Equalization operators (OLHE operators) in this paper. The following gives the formal definition of the
OLHE operators:
where k is an odd number. Note that according to our definition, there will always be exactly eight OLHE operators no
matter what the value of k is. Given an image I , OLHE produces 8 images, The images are referred to as the OLHE
images. We apply the OLHE operators to some basic image elements, such as edges and gradients, and Fig 10. shows
the results. In Fig. 3, the first row shows the input images, and the second row shows the corresponding results after
applying OLHE operators or the LHE. Some conclusions can be drawn from Fig. 10. 1) For ideal edges [Fig. 10 (b) and
(c)], LHE operator and all OLHE operators have exactly the same response. 2) For color gradients, LHE operator yields
uniform responses regardless of the directions of color gradients.
The OLHE operators, on the other hand, capture the directions of the color gradients. For example, O3 yields
the strongest response when it is applied to the image with a strong diagonal gradient that gradually intensities from
upper-left to lower-right, as what is shown in Fig. 10 (e). It can be also seen that operators
3O and
3O yield the second
strong responses, since the diagonal color gradient has components in both horizontal and vertical directions. 3) For a
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 112
pixel, if all the intensities of all the pixels within a local window centered at its upper-left location are smaller than the
intensity of this pixel, O_ 3 will have the strongest response. It is hence O_ 3 captures the diagonal gradient that
intensifies from upper-left to lower-right, or diagonal edges span from upper-right to lower-left.
Likewise, the rest of the seven OLHE operators capture gradients and edges of different 8 directions.
Fig. 11 shows facial images under different illumination conditions after applying the eight OLHE operators. Note that
different operators emphasize edges and gradients of different directions.
i) Notes on OLHE Implementation:
To obtain the OLHE results, given an image, we need to apply 8 operators to obtain 8 OLHE images, and it
seems 8 times of LHE-like operations on the whole image are required at the first glance. However, since the histogram
equalization result of each local window actually provides the results of OLHE operators at different pixel locations, we
actually need only one LHE-like operation on the whole image.
Fig 12: Implementing OLHE efficiently.
Each histogram-equalized local window can update different OLHE images at different pixel locations. For
example, in Fig. 6 (a), assume there is a 5*5 local patch on the image, and the coordinate of the upper-left-most and the
lower-right-most pixel is {1, 1} and {5, 5}, respectively. After performing the local histogram equalization within the
3*3 window centering at {3, 3}, the results at {2, 2} is exactly O 3 {2, 2}, and the results of the rest operators can be
similarly obtained, as shown in this figure. It is thus performing the local histogram equalization on every pixel once
yields the 8 OLHE images. Note that the histogram equalization results of the 8 surrounding pixels are ignored in
traditional LHE method. It is thus OLHE encodes more information than LHE. Assume a W-by-H image is given and the
local window of size k-by-k is applied, it requires W · H · k2 histogram bin updates to compute the 8 OLHE images.
ii) Face Recognition Using OLHE:
A straightforward way to apply the OLHE to face recognition is to concatenate all the eight resulting image,
forming an image of size 8∗W ∗ H, given the input image of size W ∗ H.
We define such a new image as
(6)
where the superscript „c‟ stands for „cascaded.‟ The new image obtained using c
kO contains profound information on
edges, gradients and their orientations. As will be seen in our experimental results, yielded promising overall face
identification rates.
When lower feature dimension is desired, one can also combine the OLHE images to form the image that has
the same dimension with the input image. Several ways to reduce the dimension are possible. Maybe the most
straightforward way is to sum them up with equal weights. We define
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 113
(7)
where the superscript „a‟ stands for „average‟ where and the ∑ is the
matrix summation. This representation is more compact than )(IOc
k meanwhile, by summing up the edges of 8
orientations a
kO has very good edge-preserving capability.
Fig 12. (a) and (c) Two input images with limited variance in pixel intensities since many local regions. (b) and (d)
Corresponding output of OLHE.
III. EXPERIMENTAL RESULTS We evaluated the proposed algorithm on 3 public available data sets: AR , CMU PIE , and Extended Yale B . The
experiments on AR studied face recognition via OLHE using different feature combination schemes with different
window sizes. These experiments also compared the performance of OLHE to those of other algorithms. The
experiments on CMU PIE, and Extended YaleB reported the performance of OLHE under the standard protocols. These
experiments provided a direct comparison between OLHE and the state-of-the-art algorithms.
As this paper focused on illumination compensation, only frontal faces without variations in head pose or facial
expression were considered. All the images were cropped according to the eye centers using the ground truth, and resized
to 80×100 pixels. In all the experiments except TABLE V, only the uniformly frontally lit images were used as
enrollment in the gallery. Images with variations in illumination conditions were used as probes to be identified. To
perform the face identification, we used the Nearest-Neighbor classifier on the processed intensities.
A. Performance Evaluation Under AR Dataset
AR data set [28] included facial images from 126 individuals. Each person had two sets of frontal images shot at
different time. Each set contained 13 images, one was uniformly-lit face without facial expressions, three were shot with
facial expressions, three were shot under three different lighting conditions, and the rest six were shot with sun glasses or
scarfs. We used faces without facial expressions, sun glasses and scarfs. The uniformly-lit faces were used as enrollment
in the gallery, and the faces with lighting variations were used as probes to be identified.
Fig 13. Results of the identification accuracy of LHE and OLHE on AR dataset using different kernel size. Fig. 13 shows
the face identification results using LHE and OLHE with different feature combination schemes c
kO and a
kO as well as
OLHE with Low-Intensity-Variation Pixel masking scheme (i.e. m
kO )
B. Comparison Between OLHE and State-of-the-Art Algorithms on PIE and Extended Yale B Dataset:
CMU PIE data set contained 41368 images of 68 individuals. Each person was imaged under 13 different poses,
43 different lighting conditions and 4 different facial expressions. We used the frontal faces under 21 different lighting
Tahseen et al., International Journal of Advanced Research in Computer Science and Software Engineering 6(12),
December- 2016, pp. 106-114
© 2016, IJARCSSE All Rights Reserved Page | 114
conditions with background lighting off.We randomly selected one single face for each person for enrolment, and the rest
20 faces were used for recognition. TABLE I gives the error rates of different algorithms in different data sets.
Table 1 : Different algorithm error rates
ALGORITHMS DATASETS(ERROR %)
AR
DATASET
CMU
PIE
EXTENDED
YALE B
HISTOGRAM EQUILIZATION 48.5 47.8 48.5
ADAPTIVE HISTOGRAM
EQUILIZATION
20.1
23.2 64.1
CONTRAST LIMITED
HISTOGRAM EQUILIZATION
17.8 0.36
78.0
LOCAL HISTOGRAM
EQUILIZATION
5.1
69.3 4.89
ORIENTED LOCAL
HISTOGRAM EQUILIZATION
4.1 0.1 1.0
IV. CONCLUSION
Histogram equalization is powerful method for image enhancement and it will increase the contrast of image.
The enhanced image will give the full dynamic range of histogram. However, histogram equalization process tries to
merge the adjacent gray levels together in order to force the uniformity of number of pixels in each appeared gray levels.
Consequently, the intensity saturation will be presented in darkness regions and whiteness region. Histogram equalization
assigns the intensity values of pixels in the input image such that the output image contains a uniform distribution of
intensities. It improves contrast and obtain a uniform histogram. This technique can be used on a whole image or just on
a part of an image..
We propose the Orientated Local Histogram Equalization (OLHE) that compensates illumination while
exploiting the edge orientations. We argue that the edge orientation is useful for face recognition. Three OLHE feature
combination schemes were proposed for face recognition: one encoded most edge orientations, one was more compact
with good edge-preserving capability, the other performed exceptionally well when extreme lighting conditions occur.
We also showed that LBP is a special case of OLHE, and OLHE is more effective than LBP for face recognition. The
computational complexity required by OLHE is relatively low compared with state-of-the-art algorithms such as LTP
that involves an additional chain of preprocessing or TVQI that requires solving a variational problem. We evaluated
OLHE on AR, CMU PIE, and Extended Yale B data sets, and showed the proposed algorithm outperformed the state-of-
the art algorithms in most cases.
REFERENCES
[1] Bob Cromwell, "Localized Contrast Enhancement: Histogram Equalization," paper on his web site at
http://www.cromwellintl.com/3d/histogram/Index.html.
[2] S. M. Pizer and E. P. Amburn, “Adaptive histogram equalization and its variations,” Comput. Vis. Graph.,
Image Process., vol. 39, no. 3, pp. 355–368, 1987.
[3] W. Zhao, R. Chellappa, P.J. Phillips and A. Rosenfeld “Face Recognition: A Literature Survey,” ACM
Computing Surveys, vol. 35, no. 4, pp 399-458, Dec 2003.
[4] H.Wang and N. Ahuja, “Facial expression decomposition,” in Proc.IEEE Int. Conf. Computer Vision, 2003,
vol. 2, pp. 958–965.
[5] Karel Zuiderveld, “Contrast Limited Adaptive Histogram Equalization”, Graphics Gems IV, 1994. A.S.
Georghiades, P.N. Belhumeur, and D.J. Kriegman. From few to many: Illumination cone models for face
recognition under variable lighting and pose. IEEE Trans. Pattern Anal. Mach. Intelligence, 23(6):643–660,
2001.
[6] X. Zou, J. Kittler, and K. Messer, “Illumination invariant face recognition: A survey,” in Proc. 1st IEEE Int.
Conf. Biometrics: Theory, Appl., Syst., Sep. 2007, pp. 1–8.