2nd Progress Report Submitted

A Literature Review

Introduction

            Humans are very good at recognizing. It is their innate ability to recognize and distinguish [1]. Giving the same capability to machine is vastly researched and debated area under the computer vision domain which up to date has not found any complete solution. Using a unique, measurable characteristic of a human beings known as biometrics is always a good idea in this scenario. Biometrics can be classified in to two sub domains, namely physiological biometrics and behavioral biometrics. Iris-scan, face recognition, retina scan, fingerprint scan, and palm scan can be given as examples for physiological biometrics while voice scan, signature scan, and keystroke scan falls on to behavioral biometric sub domain.

            Among the biometrics fingerprint scans are the most commonly used verification method. The solutions which uses fingerprint scanners are of low budget when compared to the capital of other solutions. Fingerprint is a very unique feature of a human where there is almost zero percent chance of two people having the same thumbprint. Main disadvantage of this method is user having to participate actively for the verification process. Hardware used for iris scans, retina scans are costly when compared with other solution which puts the main focus in to face recognition. Face is an important part of who you are and how people identify you, except for the occasion you being an identical twin. Humans use their innate ability to identify and distinguish other humans, and without any trouble humans are capable of doing this up to a satisfactory level. Identification and verification are two important concepts which comes under face recognition. Verification means where the system compares the given individual with who that individual says they are, and gives a yes or no decision and identification means the system comparing a given individual to all the other individuals in a database and giving a ranked list of matches. Scientists began to work on using the computer to recognize human faces since the mid-1960s. There are many researches done in still image analytics and video stream analytics [2] up to date and facial recognition software’s have come a long way since then. Most of the simple human face recognition systems does four major tasks, namely image capturing, extraction, comparison and finally giving a yes or no answer for the question or giving. There are many types of classifications but [3] gives a good overview of above methods based on their approach of solving the face recognition problem.

Knowledge-based Top-down Model

            The concepts of knowledge based top down model is formed by converting the knowledge about human faces in to a set of rules.  So what forms a face, from the very basic level the answer to that question would be two eyes, a nose and a mouth. The relationship between the positions of eyes, mouth and nose in a human face is taken into consideration. The important factor is the symmetry of the placement of this components. But this is not always true, what happens if a person’s face is deformed, will the same set of rules identify a face even if it’s deformed. As said in the beginning of this document an additional verification step should be done to avoid false matches (FAR).

            The main problem in this approach is deriving a set of rules from the information we have about a human face. The strictness of the rule set will have a direct effect on the results of the system. If the rules are too strict, some faces might not be identified as faces increasing the false rejection rate of the approach and in contrast if the rule set is not well formed or not strict enough it might identify non-faces as faces. A knowledge based model can be explained using Yang and Huang [4] where they try to solve the problem by three levels of rules. The rules in the higher level are more generalized as to what looks like a face and rules in the ground level are established based on features of the face. This approach failed to show considerable results and later Kotropoulous et al. [5] developed a method which extends afore said by averaging intensities in each column of the image and by finding abrupt changes in the intensities. The approach was successful giving an 85.6% success rate.

Template Matching Approach

            In this type of approach a standard pattern of human face, most preferably a frontal image is stored first. The correlation values with several types of standard patterns is calculated for nose, mouth, eyes the face contours independently. Using Skin color information or skin pixel information is one approach to do this. Yang Z. et al. [6] has conducted experiments using this method on the FERET database and the study has shown impressive results with future enhancements. Successful extraction of patterns in a face can be difficult when it comes to real life scenario, the view of the face might not be frontal always. If any match should be found through this method, the template should be extracted from a same scale image which is not possible in many scenarios. This is the main disadvantage of this method. Many researchers [7], [8] have tried to minimize the dependency on scale by using sub templates and using deformable objects.

            There has been new development in this area after the introduction of 3D template matching techniques which could tolerate facial variations due to varying lighting conditions, different head poses and facial expressions. General steps towards 3D face modeling and matching consists of face model building using video or different still images of a person and comparing the 3D model itself or a synthetic 2D image generated using the system with a given image.

An approach based on 3D morphable model was discussed in [9] where 1200 real images of six people were tested. The system showed more than 90% accuracy surpassing global face recognition systems. A scheme based on the analysis by synthesis framework was proposed by Lu et al. [10] where a number of synthetic face images are generated with appearance variations from the aligned 3D face model. These synthesized images were later used to construct an affine subspace for each subject. Results were evaluated using images taken of 10 subjects over a span of five weeks and it outperformed the PCA based model (without synthesizing). Using the advantage of accumulation of multiple frames in video Park and Jain [11], proposed a way to compensate for low resolution, poor contrast and non-frontal poses. The proposed scheme was tested on CMU’s Face in Action (FIA) video database which consists of 221 subjects and 3D modelling of non-frontal images showed 40% performance over the traditional non-frontal way.

Appearance based Methods

Neural Networks

            Neural networks are systems of program and data structures which approximates the operation of human brain. This concept can be effectively used in to the domain of face recognition to classify and identify faces through extensive training. Reason behind choosing neural network approach to face recognition problem can be justified using the ability of neural networks to remarkable derive meaning from complicated and imprecise data. Neural networks have the capability of adaptive learning and self-organizing which gives it edge over some other methods described in this section.

            In this approach the system is trained to capture patterns in images of human face assuming that such is a highly structured group of patterns which could be detected using well defined boundaries. Training the system includes feeding it with face and non-face patterns. The ability of neural networks to adaptively learn which means learn to do a task by itself will take care of the problem from that point onwards. The advantages is that there is no need to define the features of the human face, the trained system will be capable of doing this by its own. Yet this cannot be treated as a complete solution to the problem because of the following shortcomings. There is no particular measure on up to what degree the system should be trained and how many face and non-face images should be used. Apart from that there is a high overhead on training the system since it takes a lot of time and effort.

            Perhaps the most significant work in neural network based method was by Rowley, Baluja and Kanade [12] in which they have proposed a method to detect faces which are invariant to the rotation up to any degree. A bootstrap algorithm is for training the networks, which adds false detections into the training set eliminating the difficult task of manually selecting non-face training examples. This method was validated using three sets of test data which consisted of entirely different images which were used to train the system. The results observed were impressive and it revealed that their method most heavily relied on the eyes, then on the nose, and then on the mouth in human face.

            Latha, Ganesan, Annadurai [13] has proposed a method using Back Propagation of Neural Networks (BPNN) and Principal Components Analysis (PCA) which was evaluated using 200 images from the Yale database. A preprocessing step was carried out to normalize the image to improve the robustness against illumination changes and occlusions. This method has shown an accuracy of over 90%.

Support Vector Machines (SVM)

            SVM is a learning technique which was developed by V. Vapnik and his team at AT & T Bell Labs. It is new paradigm for training polynomial, neural or Radial Based Functions (RBF) Classifier [3]. In this methods answers to linearly constrained quadratic problems is found using a decomposition algorithms which guarantees the global optimality.

            Optimizing equations of the quadratic form can be overwhelming since the space is too dense and the memory requirement tends to grow with the square of the number of data points. Osuna et al. [14] developed an algorithm which decomposes the problem to achieve global optimality. They have successfully demonstrated the applicability of SVM to face recognition by evaluating their concepts against a data set of 50,000 data points. The two test sets used A, B consisted of 313 high quality and 23 mixed quality images respectively and their results yielded at 97.1% and 74.2%.

            SVM is not fully effective when there are occlusions, which causes missing entries in feature vectors Hongjun and Martinez [15] has proposed a method by defining a criterion to minimize the probability of overlap. There is no rule that only one method should be used to solve the problems in this domain, Gadekar and Suresh [16] have shown a collaborative approach of feature extraction, PCA and SVM which provides an effective solution.

Eigenface Approach

            Turk and Pentland [17] demonstrated that significant improvements can be achieved by first mapping the data into a lower dimensionality space this is known as classical eigenface approach. Later a probabilistic visual learning model which is based on density estimation of high dimensional space using an eigenspace decomposition was developed by Moghaddam and Pentland [18]. The latter showed better results when compared to the first approach, but there method was only demonstrated using the images which were taken upright (localized).

Statistical Approach

            A probabilistic model for object recognition using local appearances was proposed by Schneiderman and Kanade in [19]. This is significantly different from the appearance based methods [12], [13] which model the object on full global appearance. In simple the latter methods models the full face of the person. The statistical approach suggests that the local appearances and local patterns are more unique when it comes to human face. The intensity patterns around the human eyes are different from the intensity patterns of the cheek, hence providing a suitable criteria to identify a human uniquely. To represent this, statistics of local appearances need to be modeled.

            The Hidden Markov Model (HMM) is a clustering algorithm based on high order statistics. Rajagopalan et al. [20] presents two schemes in which the first scheme approximates the unknown distributions of the face and the face-like manifolds wing higher order statistics (HOS). An HOS-based data clustering algorithm is also proposed. In the second scheme, the face to non-face and non-face to face transitions are learnt using a hidden Markov model. The training set consisted of 2004 ”face” patterns, 4065 ”face-like” patterns and 6364 additional ”non-face” patterns and HOS scheme showed slightly better results than HMM scheme.

            Independent Component Analysis (ICA) can also be treated as a statistical model which reveals the underlying hidden factors. The information describing a face may be contained in both linear as well as high-order dependencies among the image pixels. These high-order dependencies can be captured effectively by representation in ICA space. Linear Discriminant Analysis (LDA) or Fisherfaces was another statistical approach developed by Sir. R.A. Fisher in 1963. LDA is a robust mathematical model which often produced good classification as same as some complex model. Many studies have been done to compare the benefits of ICA and LDA. A good comparison of the behavior of ICA and LDA over PCA method can be seen in the work of Delac et al. [21]. Further Lu and Plataniotis [22] have identified a novel separability criterion which is called as Maximum Separability Clusters (MSC) to prove that these methods can be used with large data sets. Draper et al. [23] conducted tests in view of finding the superior out of PCA and ICA methods. He used two ICA architectures, two ICA algorithms, and three PCA distance measures only to come to a conclusion that comparisons between PCA and ICA is complex.

Information Theory Approach

            Kullback-Leibler divergence is symmetric measure of the difference between two probability distributions H₀ and H_1. This is applied in Information Theory approach of face recognition where H₁ denotes the event being the template is a face and H₀ denoting template being a non-face. From the training sets Most Informative Pixels (MIP) are extracted to improve the Kullback relative information [24]. A window is passed over the image and distance from far space (DFFS) is calculated. If the DFSS to the face cluster is lower than the DFSS to the non-face cluster it is assumed that a face is within the image. 

Bottom-up Feature Model

            The researchers have been trying to find feature invariant for the face detection. Robustness to poses, viewpoints and changes in illumination had to be achieved for successful face detection. They were inspired by the idea that humans are able to identify faces even when there is large scale of pose variations, view point changes and changes in lightning. Plus point was this method was the ability to detect faces even the poses were different but it had its shortcomings where illumination played a big role. The features of the face were badly affected by the false detection of edges due to shadows etc. making the perceptual grouping algorithms useless.

             In the early stage the work of Govindaraju et al [25], [26] suggested that a face could be formed in terms of edges of the frontal image of the face. They used the Marr-Hildreth operator to detect edges of the image which was later processed and optimized using a thinning algorithm. The links between edges were connected depending on the proximity and the orientation of the edges. Even though it showed over 70% results on a 50 image data set, all the images used were frontal making it less preferable to address a real life scenario.

            A better method was proposed by Yow and Cipolla [27]  which consisted of two stages. Important features called as “interest points” were extracted using the raw data from the image as first step. At the final stage of processing the extracted interest points were grouped together using Gestals principles. The points were labeled using the knowledge gathered from the training data set. Each grouping was then evaluated using Bayesian network. Results obtained by this method was in the reach of 94%.

            Using human skin color to distinguish features of the human face was also a good idea. The studies showed that difference of the appearance is due to the change of intensities rather than the colors [3]. Identifying skin like pixels in an image and grouping them together using component analysis or clustering was done in this approaches.  More recent studies in color based segmentation [28] and color invariant methodologies like Color SIFT [29], [30] has widen the use of such methods.

            The turning point in human face recognition was due to the introduction of a technique called Scale Invariant Feature Transform or SIFT. Perhaps this is considered as the most influential paper in computer vision [31]. Lowes work on this area is considered as outstanding. He introduced a concept called scale invariant features, the specialty of these features were they were invariant to scale, translation, rotation and partial illumination. The so called features or interesting points were so special that they could be identified even if the image is scaled or rotated. SIFT algorithm consists of 4 phases namely, scale space extrema detection, keypoint localization, orientation assignment and building a key point descriptor. Each interest point will be described using this feature vector. This method could identify faces even the faces were partially occluded [32]. Lowe came up with a more efficient method to optimize the search time using nearest neighbors of points which was far more efficient than exhaustive searches [33]. This method even tolerates affine transformation up to some degree, but when it comes to face detection and face recognition more tolerance on affine transformation is not needed.

            An efficient method for face recognition and retrieval using Local Binary Patterns (LBP) and SIFT was proposed by Tayade and Bansode [34]. Their system took an image as input, filtered it and represented in sparse matrix deriving SIFT and LBP features. After that a transition matrix was computed using the inner distances and inner context to retrieve the results. This system identified images at an accuracy of 81.25 %.

            Mohamed Aly compared the performances of Eigenfaces, Fisher Faces and SIFT algorithms using Nearest Neighbor Algorithm on two standard set of databases, The AT & T database and the Yale database. His results have clearly shown the superiority of SIFT over the other algorithms [35]. Amit Kr. [36] came up with an improved solution which uses SIFT and Knearest neighbor classifier in which he compares the recognition rates of five different face recognition algorithms including SIFT. His method tops the table with a success rate of 97.1% where SIFT alone has only achieved 96.3%. Results were also verified using standard ORL database. There are many improvements done to the SIFT algorithms and there are versions such as Fast Approximated SIFT and Very Fast SIFT. Grabner [37] has done a comparison between SIFT and a more developed version of SIFT named fast approximated SIFT in terms of computational cost and speed. Study by Alwarin [38] demonstrates another method called VF-SIFT. He proves that by using a suitable tradeoff on up to what degree the features are matched will result in significant decrease of computational time, sometimes giving 1250 times speed over the normal method. Both of this results sets gives a hint to us that further improvement of this method will be highly likely to in future.

            Speed Up Robust Features (SURF) is also method which uses the power of feature descriptors [39]. It was a novel scale and rotation-invariant detector and descriptor. The main advantage of SURF is its speed, it approximates or even outperforms previously proposed schemes with respect to repeatability, distinctiveness, and robustness which was a progress. The reduction of the computational time allowed SURF to match features across bigger data sets. This improvement was gained through the use of concepts like integral images and interpolation. A simple Hessian matrix was used for the approximation which uses box type convolution filters initially proposed by Viola and Jones [40]. These box type convolution filters are known as Haar cascades or wavelets. SURF has a 64 dimension feature vector which gives the same results as SIFT which uses 128 dimension vector but with a reduced cost of computation. The accuracy of the hessian detector for the application of camera self-calibration and 3D reconstruction is also discussed in the same article.

            Panchal [41] has compared these two methods in terms of speed and the number of interest points detected. Tests was carried out using two images where SIFT detected 892 and 934 interest points while SURF only detected 281 and 245. But the number of feature points matched were 41 and 28, resulting in 1.543 s and 0.546 s computational time. The summary was that even though SIFT detected much more interest points it was not really needed. The results were not verified using a standard data set which included more complex scenes to come to a conclusion on what method is better over the other.

            In his Masters thesis Guerro [42] compared the performance of three different algorithms namely SIFT, SURF and Fast [43].  The results confirmed the above mentioned observation where SIFT feature detection was more far ahead of SURF but it was less effective since SURF did the same matches with less number of matched points. The study was carried in civil engineering background where the most of the images contained more edges. The images which had more textures were easily detected by SIFT and SURF. He suggested that FAST corner detector is a better detector when compared to SIFTs Difference of Gaussian (DOG) and SURFs box filter based methods.

            A Comparison of SIFT, PCA-SIFT and SURF by Jaun and Gwon [44] using K-Nearest Neighbor (KNN) and Random Sample Consensus (RANSAC) has showed that SIFT presents stability in most of the situations but was slow when it comes to computing. SURF was the fastest one with good performance as same as SIFT. PCA-SIFT showed its advantages in rotation and illumination changes. The work was impressive since it has evaluated the performances of these algorithms against aspects such as time, scale, rotation, blur, illumination and affine transformation, and results are as tabulated below.

Algorithm	Time	Scale	Rotation	Blur	Illumination	Affine
SIFT	Common	Best	Best	Best	Common	Good
PCA-SIFT	Good	Common	Good	Common	Good	Good
SURF	Best	Good	Common	Good	Best	Good

Table 1: Comparison of SIFT, PCA-SIFT & SURF

A Comparative Study of SIFT and its Variants by Wu et al. [45] compares the relative performances of SIFT, PCA-SIFT, CSIFT, GSIFT [46], ASIFT [47] and SURF over factors such as scale, rotation, illumination, blur, affine transformation and time cost. The results obtained are tabulated below to get a comprehensive idea about how these algorithms perform under different circumstances.

Algorithm	Scale & Rotation	Illumination	Blur	Affine Transformation	Time Cost
SIFT	Best	Good	Better	Good	Better
PCA-SIFT	Better	Better	Better	Good	Better
GSIFT	Good	Best	Best	Good	Better
CSIFT	Best	Better	Good	Better	Good
SURF	Common	Common	Common	Common	Best
ASIFT	Good	Common	Common	Best	Common

Table 2: Comparison of SIFT and it’s Variants

References

[1]	R. S. Sandhu and P. Samarathi, "Access Control: Principles and Practice," IEEE Communication Magazine, pp. 40-28, 1994.
[2]	W. Zhao, R. Chellappa, A. J. Phillips and A. Rosenfield, "Face Recognition: A Literature Survey," ACM Computer Journal, vol. 35, no. 4, pp. 399-458, 2003.
[3]	M. Yang, A. Ahuja and D. Kriegman, "A Survey on Face Detection Methods," 1999.
[4]	G. Yang and G. S. Huang, "Face Recognition in Complex Backgrounds," Pattern Recognition, vol. 27, no. 1, pp. 53-63, 1994.
[5]	C. Kotropoulos, A. Tefas and I. Pitas, "Frontal Face Authentication using Morphological Elastic Graph Matching," Image Processing, IEEE Transactions, vol. 9, no. 4, pp. 555-560, 2002.
[6]	Z. Jin, Z. Lou, J. Yang and Q. Sun, "Face detection using template matching and skin-color information," in International Conference on Intelligent Computing, Hefei, 2005.
[7]	T. Sakai, M. Nagao and S. Fujibayashi, "Line extraction and pattern detection in a photograph," The Journal of the Pattern Recognition Society, vol. 1, no. 3, pp. 233-236, 1969.
[8]	A. Yuille, P. Halinan and D. Cohen, "Feature Extraction from Faces using Deformable Templates," International Juornal of Computer Vision, vol. 8, no. 2, pp. 99-111, 1992.
[9]	J. Huang, B. Heisele and V. Blanz, "Component-Based Face Recognition with 3D Morphable Models," Springer, Berlin, 2003.
[10]	X. Lu, H. Rein-Lien, A. K. Jain, B. K. P. and B. K. P., " Face recognition with 3D model-based synthesis," in International Conference on Biometric Authentication volume 3072 of Lecture Notes in Computer Science, Verlag, 2004.
[11]	U. Park and A. K. Jain, "3D Model-Based Face Recognition in Video," Proc. International Conference on Biometrics,, pp. 1085-1094, 2007.
[12]	H. A. Rowley, S. Baluja and T. Kanade, "Rotation Invariant Neural Network-based Face Detection," in Computer Vision and Pattern Recognition, 1998. Proceedings. 1998 IEEE Computer Society Conference on, Santa Barbara, CA, 1998.
[13]	P. Latha, L. Ganesan and S. Annadurai, "Face Recognition using Neural Networks," Signal Processing International Journal, vol. 3, no. 5, pp. 153-160, 2009.
[14]	A. Osuna, R. Freund and F. Girosi, "Training Support Vector Machines: an Application to Face Detection," in Conference on Computer Vision and Pattern Recognition, Puerto Rico, 1997.
[15]	H. Jia and A. M. Martinez, "Support Vector MAchnes in Face Recognition with Occlusions," The Department of Electrical and Computer Engineering, Ohio State University, Columbus, OH.
[16]	A. D. Gadekar and S. S. Suresh, "Face Recognition Using SIFT-PCA Feature Extraction and SVM Classifier," IOSR Journal of VLSI and Signal Processing, vol. 5, no. 2, pp. 31-35, 2015.
[17]	M. Turk and A. Pentland, "Eigenfaces for Face Recognition," Journal of Cognitive Neuroscience, vol. 3, no. 1, pp. 71-86, 1991.
[18]	B. Moghadam and A. Pentland, "Probabilistic Visual Learning for Object Detection," in International Conference on Computer Vision, Cambridge, MA, 1995.
[19]	H. Schneiderman and T. Kanade, "Object Detection Using the Statistics of Parts," International Journal of Computer Vision, vol. 56, no. 3, pp. 151-177, 2004.
[20]	A. N. Rajgopalan, K. S. Kumar, J. Karlekar, R. Manivasakan, M. M. Patil, U. B. Desai and P. G. C. S. Poonacha, "Finding faces in photographs," in Sixth Annual Conference on Computer Vision, Mumbai, 1998.
[21]	K. Delac, G. Grgic and S. Grgic, "Independant Comaparative Study of PCA, ICA, and LDA on the FERET Data Set," Wiley Periodicals, Inc., London, 2006.
[22]	J. Lu and K. N. Plataniotis, "Boosting Face Recognition on a Large Scale Database," in International Conference on Image Processing, Rochester, 2002.
[23]	A. Bruce, A. Draper, K. Baek, M. S. Bartlett and J. R. Beveridge, "Recognizing faces with PCA and ICA," Computer Vision and Image Understanding, vol. 91, pp. 115-137, 2003.
[24]	A. Pentland, B. Moghaddam and T. Starner, "View-based and modular eigenspaces for face recognition," in IEEE Conference on Computer Vision & Pattern Recognition, Seattle, WA, 1994.
[25]	V. Govindaraju, "Locating Human faces in photographs," International Journal of Computer Vision, vol. 19, no. 2, pp. 126-146, 1996.
[26]	V. Govindaraju, D. B. Sher, R. K. Srihari and S. Srihari, "Locating human faces in newspaper photographs," in Computer Vision and Pattern Recognition, 1989. Proceedings CVPR '89., IEEE Computer Society Conference, San Diego,CA, 1989.
[27]	K. C. Yow and R. Cipolla, "Feature-Based Human Face Detection," Image and Vision Computing, vol. 15, no. 9, pp. 715-735, 1997.
[28]	Y. Tayal, R. Lamba and S. Padhee, "Automatic Face Detection using Color Based Segmentation," International Journal of Scientific and Research Publications, vol. 2, no. 6, 2012.
[29]	A. E. Abdel-Hakeem and A. A. Farag, "CSIFT: A SIFT Descriptor with Color Invarint Characteristics," in IEEE Computer Society Conference on Computer Vision and Pattern recognition, New York, 2006.
[30]	A. Singh, S. K. Singh and S. Tiwari, "Comparison of Face Recognition Algorithms on Dummy Faces," The International Journal of Multimedia & Its Application (IJMA), vol. 4, no. 4, pp. 121-135, 2012.
[31]	D. Lowe, "Object Recognition from Local Scale-Invariant Features," in International Conference on Computer Vision, Corfu, 1999.
[32]	Y. Cai, "Invariant Local Features for Face Detection," Submitted for Publication, Vancouver, 2008.
[33]	D. Lowe, "Distinctive Image Features from Scale-Invariant Keypoints," International Journal of Computer Vision, vol. 60, no. 2, pp. 91-110, 2004.
[34]	S. R. Tayade and S. Bansode, "An Efficient Face Recognition and Retrieval Using LBP and SIFT," International Journal of Advanced Research in Computer and Communication Engineering, vol. 2, no. 4, pp. 1769-1773, 2013.
[35]	A. Mohammed, "Face Recognition using SIFT Features," California Institute of Technology, California, 2006.
[36]	G. Amith Kr and Twisha, "Improved Face Recognition Technique using SIFT," in International Conference on Advances in Engineering & Technology, Chandigarh, 2014.
[37]	M. Grabner, H. Grabner and H. Bischof, "Fast Approximated SIFT," in Asian Conference on Computer Vision, Hydrabad, 2006.
[38]	F. Alhwarin, D. Ristic-Durrant and A. Graser, "VF-SIFT: Very Fast SIFT Feature Matching," Springer-Verlag, Berlin, 2010.
[39]	H. Bay, A. Ess, T. Tuytelaars and L. Van Gool, "Speed -Up Robust Features (SURF)," Elsevier, Amsterdam, 2008.
[40]	P. Viola and M. J. Jones, "Robust Real-Time Face Detection," International Journal of Computer Vision, vol. 57, no. 2, pp. 137-154, 2004.
[41]	P. M. Panchal, S. R. Panchal and S. K. Shah, "A Comparison of SIFT and SURF.," International Journal of Innovative Research in Computer and Communication Engineering, vol. 1, no. 2, pp. 323-327, 2013.
[42]	M. Guerrero, "A Comparative Study of Three Image Matcing Algorithms: Sift, Surf, and Fast," Utah State University, Logan, Utah, 2011.
[43]	M. Trajkovic and M. Hedley, "Fast Corner Detection," Image and Vision Computing, vol. 16, pp. 75-87, 1998.
[44]	L. Juan and O. Gwon, "A Comparison of SIFT, PCA-SIFT and SURF," International Journal of Image Processing (IJIP), vol. 3, no. 4, pp. 143-152, 2009.
[45]	J. Wu, Z. Cui, V. S. Zheng, P. Shao, D. Su and S. Gong, "A Comparative Study of SIFT and its Variants," Measurement Science Review, vol. 13, no. 3, pp. 122-131, 2013.
[46]	E. N. Mortensen, H. Deng and L. Shapiro, "A SIFT descriptor With Global Context," in IEEE Computer Society Conference on Computer Vision and Pattern Recognition , San Diego, CA, 2005.
[47]	J. M. Morel and G. Yu, "ASIFT: A New Framework for Fully Affine Invariant Image Comparison," SIAM Journal on Imaging Sciences, vol. 2, no. 2, pp. 438-469, 2009.