Available online at
http://www.iaeme.com/IJARET/issues.asp?JType=IJARET&VType=11&IType=10 ISSN Print: 0976-6480 and ISSN Online: 0976-6499
DOI: 10.34218/IJARET.11.10.2020.102 © IAEME Publication Scopus Indexed
FACE RECOGNITION BASED ON
CONCATENATION OF SPATIAL DOMAIN
FEATURES
Raveendra. K
Research Scholar, Dept. of ECE, GAT, Bangalore, Karnataka, India And Assistant Professor Dept. of ECE, GEC, K R Pet, Karnataka, India
Dr. Ravi. J
Professor, Dept. of ECE, GAT, Bangalore, Karnataka, India
ABSTRACT
Face recognition is one of most popular application in the electronic security system to identify a person as it is contactless and non-invasive. However, it is a big challenging task to achieve the better accuracy due to variations in intensity, illumination, orientation, different pose and facial expressions. To handle these constraints effectively, we propose a hybrid domain based face recognition using Asymmetric Region Local Binary Pattern (ARLBP) and Histogram oriented gradient (HOG) techniques. The preprocessing has been carried out on all the images to extract the face region by removing the background and resizing to 100x100. The face features are extracted using ARLBP and HoG techniques for different face databases. The obtained features are fused by concatenation and compared with trained set of features using Euclidean distance classifier. The Performance was evaluated by measuring False Acceptance Rate (FAR), False Rejection Rate (FRR), Total Success Rate (TSR) and Equal Error Rate (EER) for L space k, JAFFE and YALE Databases. It is found that L space k database has the least equal error rate among the databases studied and also we have achieved a better recognition rate for YALE database compared to the other state-of- the-art methods.
Keywords: Face recognition, recognition rate, HoG, ARLBP, Euclidean distance. Cite this Article: Raveendra. K and Dr. Ravi. J, Face Recognition Based on Concatenation of Spatial Domain Features, International Journal of Advanced Research in Engineering and Technology, 11(10), 2020, pp. 1051-1065
1. INTRODUCTION
Biometrics is a reliable and secure authentication system since it is convenient for users and difficult to steal. The term Biometric is originated from Greek word Bio means life and metrikos means measure meaning measurement of life [1]. It authenticates the person depending on his/her unique characteristics. Biometric traits are categorized as behavioral or physiological based on the characteristics of an individual. Behavioral characteristics are based on the behavior of a person includes signature, keystroke and gait recognition system. Whereas physiological characteristics are based on the shape of the body which includes face, finger, iris, palm, hand vein and DNA.
In fingerprint biometric each individual person should place his/her finger in proper position and angular orientation, where as in iris recognition system each individual should capture his/her iris template by using a special expensive device called ophthalmoscope. In order to overcome these hurdles face recognition is used because of universality, reliability and user friendly [2].
Face recognition is a one of the widely acceptable biometric, with a dynamic area of research covering various disciplines like image processing, pattern recognition and computer vision. With the advancement in the computer technology and popularity of artificial intelligence face recognition is extensively used in the latest security applications such as banking, identification of terrorist in mob, surveillance in crowd mobile phones, entrance security etc. which requires reliable authentication.
Face recognition process involves three phases namely 1) Face detection and Normalization, 2) Feature extraction and 3) Classification. During the last three decades, there is a enormous improvement in face recognition. Even though several algorithms have developed, face recognition with high accuracy remains a challenging task due to various conditions such as variations in intensity, illumination, orientation, lightning, occlusion, different pose and facial expressions [3].
This paper presents a unique face recognition system with Asymmetric Region Local Binary Pattern (ARLBP) and Histogram oriented gradient (HOG) techniques using Euclidean distance classifier. The face region is obtained using Viola Jones algorithm by removing the background and resized to 100x100. The features of the face are extracted using two descriptors ARLBP and HOG for different datasets and are fused by concatenation. The resultant final feature vector of training and testing dataset were compared using the Euclidean distance classifier. The estimated output with the trained and test feature, resulting in high accuracy.
The proposed research article is organized as follows: Section 2 confers the literature review of previous works with different feature extraction methods, classifiers and databases. Section 3 explains methodology of the proposed work with the databases used for the study, the descriptors used for feature extraction, and classifier for identification. Section 4 describes the algorithm. In section 5 result analysis is made for different databases and is compared with the performance of the existing algorithms. Section 6 Conclusion.
2. LITERATURE SURVEY
Chandan Singh et al., [4] have proposed an optimal face recognition algorithm by fusing the complementary features extracted from LBP or Local Ternary Pattern (LTP) and Zernike Moments (ZM).The ZM is capable of representing high quality global information of the image in spite of invariant to image rotation and noise whereas LBP/LTP extracts local features which are insensitive to illumination. The performance of this hybrid algorithm was analyzed on FERET, Yale and ORL databases.
Jyothi Ravikumar et al., [5] introduced the convolution based face recognition technique, in which the LL sub-band and HOG matrix are convolved to obtain the final features. In this method first LL sub-band was extracted by applying 2D-DWT on the face image, then on this sub-band HOG is applied to capture the HOG coefficients. The so extracted features are convolved to get the final set of features. The performance of this method is evaluated on the ORL, YALE, JAFFE and L Spacek database using Euclidean Distance to compare the features. Taif Alobaidi and Wasfy B. Mikhael [6] have presented the model in which two transform domains are used for feature extraction. The preprocessed image is transformed to Wavelet by applying DWT, which is then translated to Cosine domain by applying Discrete Cosine Transform (DCT). Distinct set of coefficients of DCT are taken to form the DCT feature matrix, the rest residual was transformed back to wavelet domain by applying inverse DCT.DWT is applied on these transformed residual to obtain other feature matrix. The performance of the model was evaluated on ORL and YALE database by using ED to compare the features.
Yang Zhong and Haibo Li [7] implemented the Block Matching method for face recognition instead of LBP. This method is simple as well as powerful in dealing the images with spatial shift and is radically distinct when compared to LBP. This method was evaluated on FERET and CMU-PIE dataset to validate its effectiveness comparing to LBP.
Poonam Sharma et al., [8] have introduced the technique for recognizing the pose -invariant faces, in which the statistical coefficients of invariant features captured by applying the curvelet transform are fed into the curvelet neural network. The performance of this method was tested on the FERET , CMU-PIE and LFW databases.
Zied Bannour Lahaw et al., [9] developed the robust face recognition methods in which preprocessing were performed with Discrete Wavelet Transform (DWT) to decompose the images into sub-bands. The Low-Low (LL) sub-band of the decomposed image is used for the feature extraction as this band contains the significant information. The feature extraction was done using PCA, Independent Component Analysis (ICA) and Linear Discriminant Analysis (LDA).The effectiveness of this approach was carried out on AT&T database using Support Vector Machine (SVM) classifier.
Xinfang Cui et al., [10] have implemented the Local Dominant Orientation Feature Histograms (LDOFH) algorithm for recognizing the face image. In this technique dominant orientation and the respective energy value is acquired by computing the gradient amplitude and direction of each pixel in a local patch. To minimize the redundancy and to get low-dimensional discriminative feature the Local mean-based nearest-neighbor discriminant analysis (LM-NNDA) build on PCA is employed. The effectiveness of this algorithm was tested on AR, IMM databases.
Ross P. Holder and Jules R. Tapamo [11] proposed the improved Gradient Local Ternary Patterrns (GLTP) in the view of recognizing the facial expression. First the face features are extracted by applying GLTP which are found to be large as they contain features with redundant information. This redundant information was reduced by the use of accurate Scharr gradient operator and PCA for reducing the dimensionality as a result of which the size of the feature vector was reduced, this gives rise to improvement in accuracy and efficiency.
Kaushik Sett et al., [12] have designed the algorithm in which spatial and temporal features are fused to develop the model for face recognition. Here the spatial features are extracted using PCA and the temporal features are extracted using DWT. The experiment was conducted on ORL dataset using different classifiers for comparing the fused features.
Nani Nurul Fatihah et al.,[13] have proposed the algorithm for face recognition using Local Binary Pattern (LBP) and nearest neighborur classifier. Feature extraction was done using the LBP and classification of training and testing features was carried out using k- Nearest
Neighborur and radius nearest neighbor. The performance of the algorithm was tested on AT&T, JAFFE and Yale databases. It is observed that the recognition accuracy of the Yale dataset is low.
Liwei Li and Haibin Xie [14] have introduced the face recognition system in which the Binary Gradient Pattern (BGP) features are cascaded. In this method face features are extracted by applying BGP to the original image multiple times, these are merged in series resulting in a rich texture information which in turn enhances the discriminability and robustness. The effectiveness of the algorithm was tested on Yale and extended Yale B database using the nearest neighbhour classifier.
Sujay S N and H S Manjunatha Reddy [15] proposed a novel for face recognition using extended LBP and multilevel Support Vector Machine (SVM) classifier. In this technique LBP features are extracted at different degrees and final feature was acquired by the application of histogram technique. The performance parameters were measured for Yale and FERET databases using multiclass SVM classifier for matching the test images with the trained images.
3. PROPOSED MODEL
The proposed work consists of two different spatial texture pattern extraction techniques for extracting facial image features. The system incorporates Asymmetric Region Local Binary Pattern (ARLBP) technique to extract the set of feature and merged with Histogram Oriented Gradient Features (HoG) to extract another set of face image texture features in the spatial domain. The extracted features from both the techniques are fused by concatenation to get the final features. Matching is done using the Euclidean distance classifier. The block diagram of the proposed model is as shown in Figure.1.
Figure 1 The block diagram of the proposed model
3.1. Training Face Database
3.1.1. L-Spacek Database
The L-spacek database is considered for algorithm development because of its large variations in lightning, different orientations and expression. The database has 113 male persons containing 20 face images for each person. The system is trained for first 63 persons out 113
considering first 10 images per person for recognition. The recognition rate is calculated by testing the system with eleventh image of the 63 persons.
3.1.2. JAFFE Database
JAFFE Database is considered because of its identical/similar face images of all the persons. The database has 10 persons containing 22 images per person making 220 images. The system is trained for first 6 persons out of 10 persons and considering first 10 images per person for recognition. The FRR and TSR are calculated by testing the system with 13th image from same of 6 persons. The FAR is calculated based on the remaining 4 persons out of 10 as out of database.
3.1.3. YALE Database
Yale Database is considered because of its variations in intensity, illumination change, different pose and facial expressions. The database has 15 persons containing 11 images per person making 165 images. The system is trained for first 10 persons out of 15 and considering first 7 images per person for recognition. The FRR and TSR are calculated by testing the system with 9th image from same of 10 persons. The FAR is calculated based on the remaining 5 persons out of 15 as out of database.
3.2. Preprocessing
The proposed work adopts Viola Jones algorithm [16] to detect the face from the complex background. This method uses Haar like features to extract features of both face and non-face regions. Haar features are small kernels having different shapes and scales which are used to identify the presence of the feature in the given image. Any redundancies of the obtained features were eliminated using Adaboost learning algorithm. Adaboost is machine learning algorithm which identifies best among all the features, but these features are called weak classifier, and to construct strong classifier by linearly combining weak classifier. Finally the cascade classifier contains strong classifier at different stages are used to detect the face in the given image. The Region of Interest (Face) is obtained by cropping and resized to 100x100 for all database.
3.3. Asymmetric Region based Local binary Pattern (ARLBP)
The basic Local Binary Pattern (LBP) operator of size 3x3 neighborhood has the limitation of capturing dominant features at the higher scale texture analysis, to overcome this limitation ARLBP operator is used.
The ARLBP is scalable and is capable of extracting dominant features at higher scales by considering higher values or rounded average intensities of the sub regions around the central pixel value. This will result in reducing the loss of texture information and increases the discriminative ability compared to LBP [17]
Figure 2 shows ARLBP operator consisting of eight different sub regions around the central region which are labeled as 𝑅𝑖{𝑖 = 1,2,3 … … . .8}. The sizes of the regions 𝑅1, 𝑅3, 𝑅5 and 𝑅7 are varying in both horizontal and vertical directions. Whereas the sizes of regions 𝑅2 and 𝑅6
varies in vertical directions and 𝑅4 and 𝑅8 varies in horizontal direction, the central region is
of fixed size of 1x1, whereas the sizes of the sub regions change the size of the operator. In general, ARLBP operator with (2m+1) x (2n+1) size consists of four nxm, two 1xn, two mx1 and central region is of 1x1 rectangular sized regions where m and n are width and height of the region. Keeping the value m=1 and n=1, the ARLBP operator is equivalent to basic LBP operator. The ARLBP operator is applied to extract the features by considering 5x5 window assigning m=2 and n=2 and average value of the pixel is calculated and rounded off to the nearest integer value.
Figure 3 shows the averaging of all the sub regions. Example: The average value of the sub region R1 is calculated by adding all four pixel values such as 5, 3, 2 and 6 and divided by 4 to
get the average value of the sub region R1. In the sub region R3 the average value is 3.75 which
all the sub regions. By averaging the sub region values of the 5x5 matrix of Figure.2 is reduced to 3x3 matrix as shown in Figure.4. The final feature vector is obtained by comparing the neighbor pixel value with central pixel value using relation 𝐴𝑅𝐿𝐵𝑃(𝑥𝑐, 𝑦𝑐).
𝐴𝑅𝐿𝐵𝑃(𝑥𝑐, 𝑦𝑐) = ∑8 𝑆(𝑎𝑖− 𝑎𝑐)2𝑃
𝑖=1 (1)
Where 𝑎𝑖 is the average gray values of the regions 𝑅𝑖{𝑖 = 1,2,3 … … . .8} and 𝑎𝑐 is the central pixel value. The function 𝑆(𝑎𝑖 − 𝑎𝑐) can be defined as
𝑆(𝑎𝑖− 𝑎𝑐) = {1 , 0, 𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒(𝑎𝑖− 𝑎𝑐) > 0 (2)
When central pixel value is 3 and pixel value of the neighbor i.e. R1 is 4 which is higher
than the central pixel value 3 hence is assigned the binary value 1, where as in R6 the average
pixel value is 2 which is less than the central pixel value hence is assigned binary value 0 and so on. The binary code 11111011 is constructed in clockwise direction and converted into equivalent decimal value of 251 and this value is considered as an ARLBP feature value and same procedure is carried out for the entire face image. The histogram descriptor values are calculated on obtained ARLBP features by moving the window on the entire face image as shown in Figure.5.
3.4. Histogram Oriented Features
In a pattern recognition applications, feature extraction techniques like Histogram oriented Gradients (HOG) plays a very important role since it extracts the crucial information even from the distorted images. Hence it is well applicable for Face biometric systems. The HOG feature extraction process is basically identifying the change in intensity of the pixels in the local regions and further, calculating the magnitude and orientation of pixels towards which the change occurs. Finally, it represents each local regions in an image in terms of group of histograms.
Figure 6 A part of the image depicted using uniform sized cells.
Figure 7 Representation of HOG feature vector where H (B11) depicts cell Histogram
The obtained feature vector is considered as final ARLBP feature vector for matching. Each portion of the image is divided into several blocks as shown in Figure 2 containing 9 cells, the HOG feature extractor takes each block arranges them in the form of vector as shown Figure 3.
The face detected regions using Viola Jones Algorithm have been resized for fixed dimension of 100 x 100. Resizing process is done to make sure that the number of m-HOG features generated for the succeeding images are uniform. Gradient computation is performed after resizing the image which involves computing the gradient values by applying 1D centered point discrete derivative mask in both vertical and horizontal directions. Figure 8 shows the face image divided into several blocks and block gradient represented using arrows. Specifically, this approach involves filtering the gray scale image with the following filter kernels
𝐷𝑥= [−1 0 1] And 𝐷𝑦= [−1 0 1]′
So, given an image 𝐼, we obtain the x and y derivatives using a convolution operation: 𝐼𝑥= 𝐼 ∗ 𝐷𝑥 And 𝐼𝑦= 𝐼 ∗ 𝐷𝑦
Then the magnitude of the gradient is given by: |𝐺| = (𝐼𝑥2+ 𝐼𝑦2)0.5
And angular orientation of the gradient is given by: 𝜃 = 𝑡𝑎𝑛−1(𝐼
Figure 8 The face image divided into several blocks and block gradient represented using arrows same block represented as numbers.
Figure 9 Shows the HoG features values for the training dataset considering 160 face samples The m-HOG returns the features in a row vector and comprising of 4356 features for each face image. All such row vectors for the subsequent images are stacked one below the other as shown in Figure 9 to create the trained dataset that would be used for feature matching.
3.5. Fusion and Matching
The features of ARLBP and feature of HoG are combined by means of concatenation to obtain the final feature set for matching. The final features are used for matching the test face image. Matching is performed by comparing test feature vector with trained set of feature vectors called feature space. Euclidean distance metric is uses to match the test feature vector with trained features.
3.6. Euclidean Distance (ED)
Euclidean Distance (ED), is used since its simple and effective. The Euclidean Distance 𝑑 for the pair of feature vectors is given by
𝑑 = √∑(𝑝𝑖− 𝑞)2 𝑁
𝑝𝑖 Feature vector from the Database images
𝑞 Feature vector from the Test Image 𝑁 Total Images in the Database.
4. ALGORITHM
Input: Image of Face.
Output: Recognized Face of Person.
Step 1: Face image has been selected from the database for reading.
Step 2: Face region is detected from the input image using Viola Jones Algorithm. Step 3: The detected face image is resized for 100 X 100 dimensions.
Step 4: Asymmetric Region Local Binary Pattern (ARLBP) is applied over the face image to extract the texture features
Step 5: HoG features are extracted where the histogram channels are uniformly spread over 0 to 180 degrees.
Step 6: The features from ARLBP and HoG are combined by concatenation to form final feature vector
Step 7: Repeat the above procedure for test image from step 1 to 6
Step 8: Test image features were compared with Database image features through Euclidean distance Classifier for matching.
Step 9: Image with more matching features were recognized as matching image otherwise not matching for the different threshold values.
5. RESULTS AND DISCUSSIONS
The performance of proposed face recognition system has been evaluated on L_Space K, JAFFE and YALE databases. The performance is evaluated by computing FRR, FAR, TSR and EER for different threshold values.
From the Table 1, it is found that by varying the threshold values from 0 to 0.8, the FRR has been decreased from 100% to 0% with increase in total success rate to 100%. Furthermore, FAR increases with increase in threshold value.
The FAR and FRR with respect to threshold values is shown in the Figure 10 for the database L Space K. It is found from the figure that FAR is very minimal with a value of 0 and FRR is also 0. Further, the equal error rate is 0% at a threshold value of 0.44. Hence, the threshold value 0.44 is considered as optimum value
Table 1 The variation of FRR, FAR and TSR with threshold values for L Space K database
TSR FRR FAR Threshold 0 100 0 0 0 100 0 0.04 0 100 0 0.08 0 100 0 0.12 0 100 0 0.16 0 100 0 0.20 27.19 72.11 0 0.24 60.16 39.73 0 0.28 81.62 18.43 0 0.32
93.85 6.29 0 0.36 96.72 3.14 0 0.40 100 0 0 0.44 100 0 0 0.48 100 0 0 0.52 100 0 0 0.56 100 0 0 0.60 100 0 0 0.64 100 0 37.42 0.68 100 0 83.51 0.72 100 0 91.87 0.76 100 0 100 0.80
Figure 10 FAR and FRR as a function of threshold value for the database L space k
Table 2 The variation of FRR, FAR and TSR with threshold values of JAFFE database
TSR FRR FAR Threshold 0 100 0 0 0 100 0 0.04 0 100 0 0.08 0 100 0 0.12 0 100 0 0.16 0 100 0 0.20 0 100 0 0.24 0 100 0 0.28 0 100 0 0.32 0 100 0 0.36 0 100 0 0.40 33.54 66.57 0 0.44 50 50 0 0.48 66.77 33.17 25.34 0.52 83.32 16.78 50.11 0.56
100 0 75.74 0.60 100 0 100 0.64 100 0 100 0.68 100 0 100 0.72 100 0 100 0.76 100 0 100 0.80
It has been observed from the Table 2 that the value of FRR decreases from 100% to zero whereas TSR increases to 100%. Further, it is observed that FAR tends to zero till the threshold value was up to 0.48 and later it increases with increases in threshold values.
The variation of FRR and FAR as a function of different threshold values is shown in the Figure 11. It is found from the figure that EER for JAFFE database is 30% for a threshold value of 0.533
Figure 11 FAR and FRR as a function of threshold value for the database JAFFE
Table 3 The variation of FRR, FAR and TSR with threshold values for YALE database
TSR FRR FAR Threshold 0 100 0 0 0 100 0 0.05 0 100 0 0.1 0 100 0 0.15 0 100 0 0.2 0 100 0 0.25 12.8 87.2 0 0.3 12.2 87.8 0 0.35 25 75 0 0.4 50 50 0 0.45 62.1 37.9 0 0.5 75 25 0 0.55 87.7 12.3 0 0.6 100 0 28.87 0.65
100 0 57.35 0.7 100 0 85.11 0.75 100 0 100 0.8 100 0 100 0.85 100 0 100 0.9 100 0 100 0.95 100 0 100 1
From the Table 3, it is found that by varying the threshold values from 0 to 1, the FRR has been decreased from 100% to 0% with increase in total success rate to 100%. Furthermore, FAR increases with increase in threshold value.
The FAR and FRR as a function of threshold values is shown in the Figure 12 for the database YALE. It is found from the figure that the equal error rate is 8.7% at a threshold value of 0.13. Hence, the threshold value 0.13 is considered as optimum value.
Figure 12 FAR and FRR as a function of threshold value for the database YALE
Table 4 Comparison of proposed work in terms of recognition rate and Execution time for YALE database with existing techniques
YALE Dataset (70:30 Split) Recognition Rate
(%) Training Time (ms) [18] 60.50 - [18] 83.83 - [18] 87.25 - [19] 85.58 39 [19] 85.92 23 [18] 82.08 2210 [18] 83.75 1240 [18] 89.04 1680 [20] 85.13 - [21] (40 X 40) Resizing 93.33 31.3 [21] (64 X 64) Resizing 96.11 93.8 Proposed (40 X 40) Resizing 93.28 5354.5 Proposed (64 X 64) Resizing 100.0 5702.4 Proposed (100 X 100) Resizing 91.29 5913.3
In the proposed method we got the percentage recognition rate as 91.29 and the training time 5913.3 ms as compared to the existing techniques proposed by Weiwei Zong and Guang-Bin Huang [18], Weiwei Zong [19],Sam Yin Yee [20] and Khushwant Sehra [21] as listed in Table 4. It is observed that the recognition rate and training time is better compared to the existing techniques.
6. CONCLUSION
In this work, Face recognition using a hybrid technique combines both Asymmetric Region Local Binary Pattern (ARLBP) and Histogram oriented gradient (HOG) techniques for feature extraction technique and Euclidean Distance classifier has been proposed effectively for different databases. Here the face region of an image is obtained using Viola and Jones algorithm and resized to uniform dimensions of 100 x100. Final features were obtained by concatenating ARLBP using HOG feature extraction technique. Euclidean Distance classifier is used for matching. The parameters such as FRR, FAR and TSR were measured for different databases by varying the threshold values. Further, EER has been measured and compared. It is found that EER for L space k and YALE (for image size 64 X 64) databases are 0% and JAFFE database is 30% respectively. It is observed that EER is little more in JAFFE database due to similarity in face variations compared to the L space K database.
ACKNOWLEDGMENT
The research was supported by Visvesvaraya Technological University, Jnana Sangama, Belagavi – 590018, Karnataka, India.
REFERENCES
[1] Marcos Faundez-Zanuy, “Biometric Security Technology,” Encyclopedia of Artificial Intelligence, Vol. 1, pp. 262–264, Jan. 2008.
[2] Pattarakamon Rangsee, K B Raja and Venugopal K R, “ Modified Local Ternary Pattern Based Face Recognition using SVM,” International Conference on Intelligent Informatics and Biomedical Sciences, pp. 343-350, Nov 2018.
[3] Yi-Kang Shen and Ching-Te Chiu, “Local binary pattern orientation based face recognition,” IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 1091-1095, Aug 2015.
[4] Chandan Singh, Neerja Mittal and Ekta Walia, “Complementary feature sets for optimal face recognition,” EURASIP Journal Image and Video Processing, pp. 1-18, 2014.
[5] Jyothi Ravikumar, A C Ramachandra , K B Raja and K R Venugopal, “Convolution Based Face Recognition Using DWT and HOG,” International Conference on Intelligent Informatics and Biomedical Sciences, pp. 327-334, 2018.
[6] Taif Alobaidi and Wasfy B. Mikhael, “Face recognition system based on features extracted from two domains,” International Midwest Symposium on Circuits and Systems, pp. 977-980, 2017. [7] Pengcheng Wei, Zhen Zhou, Li Li and Jiao Jiang, “Research on face feature extraction based
on K-mean algorithm,” EURASIP Journal on Image and Video Processing, pp. 1-9, 2018. [8] Ross P. Holder and Jules R. Tapamo, “Improved gradient local ternary patterns for facial
expression recognition,” EURASIP Journal on Image and Video Processing, pp. 1-15, 2017. [9] Zied Bannour Lahaw , Dhekra Essaidani and Hassene Seddik, “Robust Face Recognition
Approaches Using PCA, ICA, LDA Based on DWT, and SVM Algorithms,” International Conference on Telecommunications and Signal Processing, pp. 213-217, 2018.
[10] Xinfang Cui, Peng Zhou and Wankou Yang, “Local dominant orientation feature histograms for face recognition,” Applied Informatics, pp. 1-10, 2017.
[11] Ross P. Holder and Jules R. Tapamo, “Improved gradient local ternary patterns for facial expression recognition,” EURASIP Journal on Image and Video Processing, pp. 1-15, 2017. [12] Kaushik Sett, Pritam Das, Prosun Ghosh, Surya kanta Ghosh, Aritra Dey, Neelanjan Saha and
Suparna Biswas, “Face Recognition using Fusion of Spatial and Temporal Features,” Emerging Trends in Electronic Devices and Computational Techniques, pp. 1-4, 2018.
[13] Nani Nurul Fatihah, Gunawan Ariyanto, Asslia Johar Latipah and Dwi Murdaningsih Pangestuty, “Face Recognition Using Local Binary Pattern and Nearest Neighbour Classification,” International Symposium on Advanced Intelligent Informatics, pp. 142-147, Aug 2018.
[14] Liwei Li and Haibin Xie, “Face recognition algorithm based on cascading BGP feature fusion,” Chinese Control And Decision Conference, pp. 3676-3680, July 2018.
[15] Sujay S N and H S Manjunatha Reddy, “Extended Local Binary Pattern Features based Face Recognition using Multilevel SVM Classifier,” International Journal of Recent Technology and Engineering, Vol.8, Issue.3, pp. 4123-4128, sept 2019.
[16] Paul Viola and Michael J. Jones, “ Robust Real-Time Face Detection,” International Journal of Computer Vision, Vol. 57, No. 2, pp. 137-154, 2004.
[17] C. L. Shrinivasa Naika, Pradip K. Das , Shivashankar B. Nair, “Asymmetric region Local Binary Pattern operator for person-dependent facial expression recognition”, International Conference on Computing, Communication and Applications, pp. 1 – 5, Feb. 2012.
[18] Weiwei Zong and Guang-Bin Huang, “Face recognition based on extreme learning machine,” Neuro computing Vol. 74, Issue.16, pp. 2541–2551, Sept 2011.
[19] Weiwei Zong, Hongming Zhou, Guang-Bin Huang and Zhiping Lin, “Face Recognition Based on Kernelized Extreme Learning Machine,” Autonomous and Intelligent Systems. Lecture Notes in Computer Science, Springer, Berlin, Heidelberg, vol. 6752, pp. 263–272, 2011. [20] Sam Yin Yee, Taha H. Rassem, Mohammed Falah Mohammed and Suryanti Awang “Face
Recognition Using Laplacian Completed Local Ternary Pattern (LapCLTP),” Advances in Electronics Engineering. Lecture Notes in Electrical Engineering, Springer, Singapore, Vol. 619, pp. 315-327, Dec 2019.
[21] Khushwant Sehra , Ankit Rajpal , Anurag Mishra , and Girija Chetty, “HOG Based Facial Recognition Approach Using Viola Jones Algorithm and Extreme Learning Machine,” Computational Science and Its Applications, Lecture Notes in Computer Science, Springer, Cham, vol. 11623, pp. 423–435, 2019.