Local Octal Pattern: A Proficient Feature Extraction for Face … · 2014-04-08 · Nithya J1,...
Transcript of Local Octal Pattern: A Proficient Feature Extraction for Face … · 2014-04-08 · Nithya J1,...
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 139
Available Online at www.ijcsmc.com
International Journal of Computer Science and Mobile Computing
A Monthly Journal of Computer Science and Information Technology
ISSN 2320–088X
IJCSMC, Vol. 3, Issue. 4, April 2014, pg.139 – 146
RESEARCH ARTICLE
Local Octal Pattern: A Proficient Feature
Extraction for Face Recognition
Nithya J1, Suchitra S
2, Dr. S Chitrakala
3
1,2Computer Science and Engineering, Easwari Engineering College, India
3Computer Science and Engineering, Anna University, India
Abstract— This paper presents a novel and efficient face image representation based on Local Octal Pattern (LOP) texture
features. The standard methods viz., the local binary pattern (LBP), the local ternary pattern (LTP) and the local tetra pattern
(LTrP) are able to encode with a maximum of four distinct values about the relationship between the referenced pixel and its
corresponding neighbors. The proposed method calculates the diagonal, horizontal and vertical directions of the pixels using
first-order derivatives. Thereby, it encodes eight distinct values about the relationship between the referenced pixel and its
neighbors. The performance of the proposed method is compared with the LBP, the LTP and the LTrP based on the results
obtained in terms of average precision and average recall on PubFig image database.
Keywords— face image retrieval; local binary pattern (LBP); local ternary pattern (LTP); local tetra pattern (LTrP); CBIR
I. INTRODUCTION
A. Motivation
The prodigious growth of camera devices, enable people to freely take photos to capture moments of life. By estimation more
than 60% of those photos cover human faces. The importance and the sheer amount of human face photos make manipulations
(e.g., search and mining) of large-scale human face images a really important research problem and aid many real world
applications. The designing aspects of a face retrieval system take account of noise, illumination, poses and expressions, low cost
for processing and eliminating misclassified face images. Thus, an efficient Face Image Retrieval System (FIRS) development is
required to automatically search the more relevant image from the large database.
Facial recognition system automatically identifies or verifies a person from a digital image. One of the methods is by
comparing selected facial features from a query image and a facial database. Automated FIRSs [1] typically involve finding facial
landmarks (such as the center of the eyes) for aligning, normalizing the appearance of face, selecting a suitable feature
representation, learning perfect feature combinations, and developing precise and scalable matching approaches.
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 140
Fig.1. Major steps in automatic face recognition as in [1]
The prominent step in FIRS is feature extraction whose efficiency depends upon the methods being implemented for extracting
features for a given face image. Color, texture, shape and spatial properties are the major features extracted from the face image.
There is difficulty in achieving a general representation for face images from the extracted feature, due to the fact that the
photographs are taken in different conditions such as illumination, orientation, pose, expression and aging. A comprehensive and
extensive literature survey on face detection, extraction and recognition is presented [2] – [5].
To leverage the representation of the image, the local octal pattern is calculated based on the direction of pixels using
horizontal, vertical and diagonal derivatives. By incorporating this method, we build a large-scale content based face image
retrieval system by taking advantage of the direction of the pixels. In order to evaluate the performance of the proposed system,
we conduct extensive experiments on public datasets named Pubfig [6].
B. Related Works
The local binary pattern (LBP) has been a powerful feature for texture classification and retrieval. The LBP as described in [7]
and [8] divides the face image into multiple regions for extracting the LBP feature distribution. Finally, the extracted feature is
converted into feature vector. Unlike other LBP feature extraction methods, genetic based LBP [9], attempts to reduce the size of
the feature extraction based on genetic and evolutionary computing (GEC). GEC evolves LBP extraction that has been randomly
distributed.
The forms of the LBP cannot passably deal with the presence of variations in images due to facial expression, pose,
illumination, etc. This problem has been addressed with introduction, the local ternary pattern (LTP) [10]. The proposed method
tackles the different lighting conditions by combining local feature extraction, distance transform based matching, kernel-based
feature extraction, illumination normalization and multiple feature fusion. LTP solves the sensitivity to noise to some extent by
introducing a third state. But such minimal pixel difference can be overcome by noise. To address this issue, relaxed LTP [11]
introduces uncertain state to encode the minimal pixel value.
The LBP and the LTP are able to encode images with only two and three different values. However, the LTrP [12] has been
able to encode four different values. LTrP supports varied applications like CBIR, fingerprint recognition due to its integration of
direction of pixels with respect to center pixel. The LBP, the LTP, and the LTrP translate the relationship between the center pixel and the referenced pixels based on the
distribution of edges. The maximum directions which can be coded is only four. Thus, the performance of these methods can be improved by separating the edges in more than two directions. This has led to proposal of a novel method, referred to as local octal pattern (LOP) for content based face image retrieval.
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 141
Fig.2. Calculation of the LBP and the LTP
II. FEATURE DESCRIPTORS
A. LBPs
The LBP operator thresholds the each pixel with the center pixel gc value in 3 X 3-neighborhood. It encodes 1 for gray value
greater than gc and 0 for gray value less than gc. Fig. 2 illustrates the basic LBP operator based on [13].
∑ (
)
(1)
where gc is the center pixel gray values, gp is the gray value of its neighbors, P is the number of neighbors, and R is the radius of
the neighborhood.
B. LTPs
As introduced in Tan and Triggs [10], LTP is an extension of LBP. Unlike LBP, it uses a threshold constant to threshold
pixels into three values rather threshold the pixels into 0 and 1.
The gray values when equal to the threshold are quantized to zero, those above the threshold are quantized as +1, and those below
the threshold are quantized as -1. Fig. 2 shows the calculation of LTP.
(
)
| |
C. LTrPs
Murala et al. [12] proposed LTrP that encodes the relationship between the center pixel and the referenced pixel based on the
direction of the center pixel gc. The horizontal and vertical directions are calculated using first-order derivatives.
The center pixel’s first-order derivatives (gc), along vertical (gv) and horizontal (gh) direction can be formulated as:
(4)
(5)
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 142
and the direction of the center pixel can be formulated as
(6)
From (6), the possible direction is 1, 2, 3, or 4, for each center pixel and finally the image is converted into four values.
Fig.3 and Fig.4 shows the calculation of LTrP. More details about LTrP can be found in [8].
Tetra Pattern = 3 0 3 4 0 3 2 0
Pattern1 = 00000010; Pattern2 = 10100100; Pattern3 = 00010000
Fig.3. Example of LTrP
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 143
Fig.4. Calculation of tetra pattern for the center-pixel direction “1”
III. PROPOSED METHOD
In this section, we first outline our scalable content-based face image retrieval system, and then we explain the proposed
method: Local Octal Pattern (LOP).
A. System Overview
Fig. 7 shows the overall system representation. First, system detects the location of the faces using Viola-Jones face detector [14] for all image in the database. For each detected face image, we will extract the facial components of face viz., two eyes, nose tip, and two mouth corners. A 5X7 grid is defined at each detected component, where each grid is a square patch [15]. Thereby a total of 175 grids are extracted from five components. A LOP feature descriptor is computed for each patch. Every descriptor is quantized into codewords using sparse code as described in [16], after extracting local feature. Many techniques in information retrieval can be applied in the storage of the inverted index [17]. When a query image arrives, it will go through the same procedure to obtain sparse codewords and use these codewords to retrieve images in the index system.
B. Local Octal Pattern
LOP encodes the relationship between the center pixel and the referenced pixel based on the direction of the center pixel gc.
The diagonal, the horizontal and vertical directions are calculated using first-order derivatives. An 8-bit tetra pattern for each
center pixel is obtained. From the calculated directions, patterns are divided into four parts. Finally, three binary patterns are
generated for each direction. Thus, a total of 12(4X3) binary patterns are obtained.
As described by Guo et. al.[18], using sign and magnitude components will extract more useful information. Thus, the 13th
binary pattern by using the magnitudes of diagonal, vertical and horizontal first order derivatives has been included.
Fig. 5 illustrates the flowchart of the proposed image retrieval system and Fig. 6 shows the algorithm of LOP feature
descriptor.
Fig.5. LOP System Framework
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 144
Fig.6. LOP Algorithm
IV. EXPERIMENTAL RESULTS
For experiment, images from PubFig database [6] have been used. This database consists of facial images of people taken in
different conditions. The performance of the proposed method is measured in terms of average precision and average recall.
The precision is defined as follows:
The recall is defined as follows:
where A = Number of relevant images retrieved
B = Number of relevant images not retrieved
C = Number of irrelevant images retrieved
6 𝐼 𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼45𝑜
𝑔𝑐
7 𝐼45𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼 𝑜
𝑔𝑐
8 𝐼45𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼 𝑜
𝑔𝑐
𝑀𝐼1 𝑔𝑝
𝐼 𝑜 𝑔𝑝
2 𝐼45𝑜 𝑔𝑝
2 𝐼 𝑜 𝑔𝑝
2
𝐿𝑃 𝑝− ∗ 𝑓
𝑃
𝑝=
𝑀𝐼1 𝑔𝑝 𝑀𝐼1 𝑔𝑐
Step4: Calculate the 8-bit tetra pattern for each
center pixel, and three binary pattern are generated
from them
Step5: Calculate the histograms of binary patterns
Step6: The magnitude of center pixel can be
calculated as
Step7: Construct the feature vector
Step8: A given query image and images in the
database is compared and a set of similar images is
retrieved
𝐼 𝑜 𝑔𝑐 𝐼 𝑔 𝐼 𝑔𝑐
𝐼45𝑜 𝑔𝑐 𝐼 𝑑 𝐼 𝑔𝑐
𝐼 𝑜 𝑔𝑐 𝐼 𝑔𝑣 𝐼 𝑔𝑐
𝐼 𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼 𝑜
𝑔𝑐
𝐼 𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼 𝑜
𝑔𝑐
𝐼 𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼 𝑜
𝑔𝑐
𝐼 𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼 𝑜
𝑔𝑐
5 𝐼 𝑜 𝑔𝑐 𝑎𝑛𝑑 𝐼45𝑜
𝑔𝑐
Step 1: Load the image and convert it into
grayscale.
Step2: Calculate the first-order derivatives in
diagonal (gd), vertical (gv) and horizontal (gh)
directions with respect to center pixel, gc.
Step3: Direction of each pixel is calculated by
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 145
Fig.8. Illustrates results of the LBP, LTP, LTrP and LOP with respect to average precision and average recall
Based on this figure, it is evident that the LOP significantly improves the retrieved results.
V. CONCULSION AND FUTURE WORK
In this paper, we have presented a novel approach named as LOP for face recognition. The LOP calculates the direction of
pixels using diagonal, vertical and horizontal derivatives.
The overall performance of LOP has been compared with the LBP, the LTP and the LTrP on facial images. The result has
improved significantly with respect to the average precision and the average recall.
Results can be further leveraged by incorporating human attributes (e.g., gender, race, and hairstyle) in a scalable framework.
REFERENCES
[1] U. Park and A. K. Jain, “Face matching and retrieval using soft biometrics,” IEEE Trans. Inf. Forensics Security, vol. 5, no. 3, pp. 406–415, Sep 2010.
[2] Yongzhong Lu, Jingli Zhou, Shengsheng Yu, “Survey of Face Detection, Extraction and Recognition” , Computing and Informatics, Vol. 22, 2003.
[3] Li Xianwei, Zhang Haiyan , “A Survey of Face Recognition Methods”, in Proc. 2nd International Conference on Computer Science and Electronics Engineering,2013.
[4] Rabia Jafri and Hamid R. Arabnia, “A Survey of Face Recognition Techniques”, Journal of Information Processing Systems, Vol.5, No.2, June 2009.
[5] V.Vijayakumari,“Face Recognition Techniques: A Survey”, World Journal of Computer Application and Technology, 41-50, 2013.
[6] N. Kumar, A. C. Berg, P. N. Belhumeur, and S. K. Nayar, “Attribute and simile classifiers for face verification,” in Proc. Int. Conf. Computer Vision,
2009.
[7] T. Ahonen, A. Hadid, and M. Pietikainen, “Face description with local binary patterns: Applications to face recognition,” IEEE Trans. Pattern Anal.
Mach. Intell., vol. 28, no. 12, pp. 2037–2041, Dec. 2006.
70
72
74
76
78
80
82
84
86
10 20 30 40 50 60 70 80
Pre
cis
ion
Number of Images
LBP
LTP
LTrP
LOP
70
75
80
85
90
10 20 30 40 50 60 70 80
Re
call
Number of images
LBP
LTP
LTrP
LOP
Nithya J et al, International Journal of Computer Science and Mobile Computing, Vol.3 Issue.4, April- 2014, pg. 139-146
© 2014, IJCSMC All Rights Reserved 146
[8] G. Zhao and M. Pietikainen, “Dynamic texture recognition using local binary patterns with an application to facial expressions,” IEEE Trans. Pattern
Anal. Mach. Intell., vol. 29, no. 6, pp. 915–928, Jun. 2007.
[9] Joseph Shelton, et. al., “Genetic Based LBP Feature Extraction and Selection for Facial Recognition”, Published in ACM, 2011.
[10] X. Tan and B. Triggs, “Enhanced local texture feature sets for face recognition under difficult lighting conditions,” IEEE Trans. Image Process., vol.
19, no. 6, pp. 1635–1650, June 2010.
[11] Jianfeng Ren, Xudong Jiang and Junsong Yuan, “Relaxed Local Ternary Pattern for Face Recognition”, 20th IEEE International Conference on Image
Processing, Sep. 2013.
[12] Subrahmanyam Murala, R. P. Maheshwari, and R. Balasubramanian, “Local Tetra Patterns: A New Feature Descriptor for Content-Based Image
Retrieval”, IEEE Transactions on Image Processing, vol. 21, no. 5, 2874-2886, May 2012.
[13] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution gray-scale and rotation invariant texture classification with local binary patterns,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 971–987, Jul. 2002.
[14] P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” in Proc. IEEE Conf. Computer Vision and Pattern
Recognit., 2001.
[15] Z. Wu, Q. Ke, J. Sun, and H.-Y. Shum, “Scalable face image retrieval with identity-based quantization and multi-reference re-ranking,” in Proc. IEEE
Conf. Computer Vision and Pattern Recognit., 2010.
[16] J.Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online dictionary learning for sparse coding,” in Proc. ICML, 2009.
[17] J. Zobel and A. Moffat, “Inverted files for text search engines,” ACM Comput. Surveys, 2006.
[18] Z. Guo, L. Zhang, and D. Zhang, “A completed modeling of local binary pattern operator for texture classification,” IEEE Trans. Image Process., vol. 19, no. 6, pp. 1657–1663, Jun. 2010.