2017 2017 13th13th International International Conference Conference on Signal-Image on Signal-Image Technology Technology & Internet-Based and Internet-Based Systems Systems (SITIS)
Pyramid Binary Pattern for Age Invariant Face Verification Saroj Bijarnia
Preety Singh
Indian Institute of Technology, Delhi, India Email:
[email protected]
The LNM Institute of Information Technology, Jaipur, India Email:
[email protected]
the paper.
Abstract—Verification is a fundamental issue in many security based systems. Automatic face verification across aging is an important problem that has recently been added to the problem of face recognition. To address this problem, we propose the Pyramid Binary Pattern. This employs Local Binary Pattern along with its hierarchical information. The generated texture feature vector is reduced through Principal Component Analysis. Classification is performed with Support Vector Machine. Results show that our proposed method enhances verification accuracy. Index Terms—biometrics, age invariant, verification, support vector machine
II. R ELATED
Ramanathan and Chellappa [14] have performed face verification using a Bayesian age difference classifier. In this method, given pair of faces are first grouped as intra-personal and extra-personal pairs. The intra-personal pairs are further divided into four groups based on their age difference by selecting the maximum of their Posteriori probabilities. Experiments have been performed on a private database. An accuracy of 15% has been reported.
I. I NTRODUCTION Biometric systems based on palm, finger, iris etc. are intrusive data acquisition. For example, in iris and fingerprint, the person is supposed to put his/her eye and finger at a proper position. In case of face recognition user is unaware that he/she is being monitored be a camera. This makes it nonintrusive. However, there are number of challenges associated with biometric systems, related to aging. This issue has not received much attention compared to other sources which degrade performance of face verification like pose, expression, lighting [13]. Aging affects both shape (weight gain/loss) and texture (wrinkles, skin color) of face [14][3]. It is not an easy task to predict how a person will look like after a certain period of life because aging affects each person differently based on their habits, lifestyle, hereditary factors, stress etc. Age invariant face verification can be employed in finding missing children, passport photo verification, multiple enrollment detection and many more. However, it is not possible to update the image databases at regular intervals. Thus, the task of face verification across aging becomes all the more vital. In this paper, we use texture features to verify whether the query image actually matches the enrolled image in the database. We propose the use of a new variant of Local Binary Pattern (LBP) by adding hierarchical information of the face texture to the traditional LBP approach. This is henceforth referred to as the Pyramid Binary Pattern. Dimensionality reduction of extracted features is performed using Principal Component Analysis (PCA). For classification, Support Vector Machine (SVM) is employed. The rest of the paper is organized as follows:- Section II presents a brief survey of research being done in this area. Section III presents the proposed methodology. Section IV discusses the experiments and results. Section V concludes 978-1-5386-4283-2/17 $31.00 © 2017 IEEE DOI 10.1109/SITIS.2017.45
WORK
Ling et al. [8] have proposed a method using gradient orientation pyramid and discarding the magnitude information. FGNET [1] database has been used. They have reported an accuracy of 75.90%. Biswas et al. [4] have used location drift of facial features across age progression. They report that certain selected facial features larger coherence among images across aging. Du et al. [6] have proposed a method which employs the active appearance model (AAM) and gradient orientation pyramid (GOP). Experiments have been performed on FGNET and MORPH [15] databases. Accuracy reported is 75.98% and 85.73% respectively. Mahalingam and Kambhamettu [9] have proposed a graph based method. An age model for each image is learnt and a graph space for each face image is computed. The graph of the test image is matched with the training images. They have used the FGNET database and accuracy is shown to be 80%. III.
PROPOSED METHODOLOGY
We propose extracting features based on Local Binary Pattern (LBP). LBP is robust to changes in illumination and is computationally simple. The process flow is shown in Figure 1. Preprocessing of image is done, followed by constructing a Gaussian image pyramid. Each pyramid level image is divided into segments and feature vector of each segment is computed using Local Binary pattern (LBP). The feature vectors of each segment at each level are concatenated to form the Pyramid Binary Pattern (PBP). Dimension reduction is done by Principal Component Analysis [12]. For classification we employ Support Vector Machine [5]. 218
Fig. 2.
Gaussian image pyramid
B. Construction of Gaussian Image Pyramid Gaussian image pyramid [2] is constructed by using bicubic interpolation [7]. This is done upto two levels by scaling each level by half, yielding a 3-level pyramid. In the Gaussian image pyramid, the lowest level (level 0) is of the highest resolution and highest level (level 2) is of the lowest resolution. Computation of texture pattern at each level of the pyramid adds to information so that verification accuracy can be enhanced. Example of image pyramid can be seen in Figure 2. For a given image, I(p), where p = p(x, y) indicates pixel locations, we define the pyramid of image I(p) as: I(p(x, y); 0) = I(p(x, y)) I(p(x, y); l) = [I(p(x, y); l − 1) φ(p)] ↓2 l = 1, 2
(1)
where φ(p) is a Gaussian kernel and represents the convolution operator. ↓2 denotes decrease in size by half (using bicubic interpolation), l denotes number of pyramid level (here, we go till two levels). This considers 16 pixels (4 × 4) neighborhood for image resampling.
Fig. 1.
C. Feature Extraction
Flowchart of Proposed Methodology
The Pyramid Binary Pattern (PBP) is now determined by first computing Local Binary Pattern (LBP) [11] for each segment at each level. LBP is preferable because of its two properties: tolerance against illumination conditions and is computationally simple. It is applied in face recognition tasks successfully.
A. Image Preprocessing In this step, we preprocess the image to extract face region from the image using the following steps:
1) Local Binary Pattern: For any pixel in an image, a sequence of binary digits is obtained by comparing the pixel value with its eight neighboring pixels. LBP for a given image can be created as follows: 1) For each pixel in an image, consider a 3 × 3 neighborhood. Threshold the center pixel intensity with each of its neighbors in a clockwise direction.
1) Alignment of face is done first. This is done using the center of eye locations. Using the two eye locations, a 2D affine transformation is performed to fix the angle and distance between the two eyes. 2) The face detector proposed by Viola and Jones [16] is now used as a face region detector.
2) Let the central pixel intensity value be c and neighboring pixel intensity value be x. If the center pixel’s value is greater then the neighboring pixel’s value, then 1 generated, otherwise 0 is generated. This produces an
3) The face region is cropped from the whole image. 4) The RGB image is converted into grayscale image.
219
TABLE I C ONFUSION MATRIX
Positive Actual Class Fig. 3.
Predicted Class Negative
Positive
True Positive (TP)
False Negative (FN)
Negative
False Positive (FP)
True Negative (TN)
Illustration of Local Binary Pattern
D. Classification For face verification task we are using Support Vector Machine as a classifier. Cross validation is used for optimal parameter selection for cost and gamma parameters using 10fold cross validation. We model face verification into a binary class problem by making intra-personal (pair of feature vectors of same person) and extra-personal pairs (pair of feature vectors with a different person). E. Evaluation Metrices
Fig. 4.
A confusion matrix (refer Table I) helps analyze the results of classification. Here, True Positives (TP) and True Negatives (TN) represent the correct predictions. False Positives (FP) and False Negatives (FN) are incorrect predictions. Accuracy is measured as the percentage of predictions that are correct and can be computed by Equation 3.
Divide in segments
eight-bit binary number for each pixel value. 1; if s(c) = 0;
c≥x c 17. This subset consists of roughly frontal images of 269 people.
3) Generated eight-bit binary number is converted to a decimal number using weights assigned to neighbors. An example of this is shown in Figure 3. 4) This process is repeated for all pixels. A histogram is computed using the decimal values associated with each pixel. This generates the Local Binary Pattern (LBP). 2) Computation of Pyramid Binary Pattern: The Pyramid Binary Pattern uses texture information at different levels of the image pyramid. 1) Divide each generated pyramid level image into equal size segments. Example is shown in Figure 4.
For each compute image we pyramid LBP as discussed in Section III. Pyramid LBP gives 3072 texture descriptors for an image. PCA is applied and we choose 190 principal components accounting for 95% data variance. Using these principal components we generate 680 intra-personal pairs and 2000 extra-personal pairs randomly. Training and testing sets are created in the ratio 70:30 respectively.
2) Compute and concatenate the LBP of the four segments for finding the LBP of one level of the pyramid.
On performing face verification, we achieve an accuracy of 92.24%. The same set of experiments is done by employing the regular LBP to achieve 91.1% accuracy. Though the improvement is not very significant but indicates that hierarchical information is contributing to the system performance. This information can be exploited in further work to add more levels of pyramid and incorporate additional texture information.
3) Repeat for each level of image pyramid. 4) Concatenate LBP of all pyramid levels to form the Pyramid Binary Pattern (PBP). 3) Dimensionality Reduction: For dimensionality reduction of feature vectors formed using Pyramid Binary Pattern, we use Principal Component Analysis [12]. PCA transforms feature sets from a higher dimensional space to a lower dimensional space.
Ramanathan et al. [14] have reported an accuracy of 15% using a private database. Ling et al. [8] report an accuracy
220
of 75.90% on FGNET database. Meng et al. [10] show an accuracy 47.83% on MORPH database. Mahalingham and Kambhanmettu [9] have used FGNET and have reported 80% accuracy. Du et al. [6] report an accuracy of 85.75% and 75.98% respectively on the MORPH and FGNET datasets. We see that our accuracy of 92.24% using Pyramid Binary Pattern outperforms these results.
[15] A. W. Rawls and K. Ricanek, Jr., “Morph: Development and optimization of a longitudinal age progression database,” Proceedings of the 2009 Joint COST 2101 and 2102 International Conference on Biometric ID Management and Multimodal Communication, pp. 17–24, 2009. [16] P. Viola and M. J. Jones, “Robust real-time face detection,” Int. J. Comput. Vision, vol. 57, no. 2, pp. 137–154, 2004.
V. C ONCLUSION Considering the impact of age on face verification system, it is very important to work towards an age invariant face verification system. In this paper, we propose Pyramid Binary Pattern, which is a variant of the Local Binary Pattern. We apply Principal Component Analysis on the extracted features and perform face verification. Results show that our method enhances system accuracy. The hierarchical information will be further increased in future experiments to explore effect on improvements in face verification. R EFERENCES [1] “FGNET Aging Database,” 2010. [Online]. Available: http://www.fgnet. rsunit.com [2] E. H. Adelson, C. H. Anderson, J. R. Bergen, P. J. Burt, and J. M. Ogden, “Pyramid methods in image processing,” RCA engineer, vol. 29, no. 6, pp. 33–41, 1984. [3] S. Bijarnia and P. Singh, “Age invariant face recognition using minimal geometrical facial features,” Advanced Computing and Communication Technologies, pp. 71–77, 2016. [4] S. Biswas, G. Aggarwal, N. Ramanathan, and R. Chellappa, “A nongenerative approach for face recognition across aging,” Biometrics: Theory, Applications and Systems, 2008. BTAS 2008. 2nd IEEE International Conference on, pp. 1–6, 2008. [5] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol. 20, no. 3, pp. 273–297, 1995. [6] X. C.-M. Du, Ji-XiangWu, “A method based on active appearance model and gradient orientation pyramid of face verification as people age,” Mathematical Problems in Engineering, pp. 11–16, 2014. [7] P. Getreuer, “Linear Methods for Image Interpolation,” Image Processing On Line, vol. 1, 2011. [8] H. Ling, S. Soatto, N. Ramanathan, and D. Jacobs, “Face verification across age progression using discriminative methods,” Information Forensics and Security, IEEE Transactions on, vol. 5, no. 1, pp. 82–91, 2010. [9] G. Mahalingam and C. Kambhamettu, “Age invariant face recognition using graph matching,” Biometrics: Theory Applications and Systems (BTAS), 2010 Fourth IEEE International Conference on, pp. 1–7, 2010. [10] C. Meng, J. Lu, and Y.-P. Tan, “A comparative study of age-invariant face recognition with different feature representations,” Control Automation Robotics Vision (ICARCV), 2010 11th International Conference on, pp. 890–895, 2010. [11] T. Ojala, M. Pietikainen, and D. Harwood, “Performance evaluation of texture measures with classification based on kullback discrimination of distributions,” Pattern Recognition, 1994. Vol. 1 - Conference A: Computer Vision amp; Image Processing., Proceedings of the 12th IAPR International Conference on, vol. 1, pp. 582–585, 1994. [12] K. Pearson, “LIII. on lines and planes of closest fit to systems of points in space,” The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, vol. 2, no. 11, pp. 559–572, 1901. [13] P. Phillips, J. Beveridge, B. Draper, G. Givens, A. O’Toole, D. Bolme, J. Dunlop, Y. M. Lui, H. Sahibzada, and S. Weimer, “An introduction to the good, the bad, amp; the ugly face recognition challenge problem,” Automatic Face Gesture Recognition and Workshops (FG 2011), IEEE International Conference on, pp. 346–353, 2011. [14] N. Ramanathan and R. Chellappa, “Face verification across age progression,” Image Processing, IEEE Transactions on, vol. 15, no. 11, pp. 3349–3361, 2006.
221