This paper is about the facial recognition system technology which has been an ongoing process for 50 years and desired excellence has finally been achieved through Artificial Neural Networks. The field was previously governed by systems like principal component analysis, linear discriminant analysis, independent component analysis and the elastic bunch graph matching method which have been briefly described along with their drawbacks. The older systems had a consistent problem of not being able to recognize faces with accuracy due to factors like low light, facial hair, glasses or expressions on the face which was a huge drawback not letting the technology advance to live video facial recognition. On the contrary, ANN is a modern system based on the neural structure found inside brain which process information of a highly complex nature in an efficient and effective manner. As the new system ANN has emerged to be better than the earlier systems it is necessary to go through all the earlier and newest concepts to ascertain which system would be the most appropriate for the future advancements.
Background of Facial Recognition Systems 3
Principal Component Analysis 3
Disadvantages of PCA 4
Linear Discriminant Analysis 4
Drawbacks of LDA 5
Independent Component Analysis 5
Drawbacks of ICA 6
Elastic Bunch Graph Matching Method 6
Drawbacks of EBGM 7
Artificial Neural Networks and the Way Forward 7
Neural Networks 7
Artificial Neural Networks in Facial Recognition 8
How ANN is better than Other Systems 9
Over the past few decades, humans have made exponential advancements in technology in every field to come up with innovations that help us at every step and have changed our life to a large extent. Certain technologies are enhanced to make our lives more secure and comfortable, and the field of information technology has played a major role in developing such technologies. Facial Recognition is one such technology, which has been in a development stage for a long time since the beginning of research on facial recognition concepts in 1960s and now, the facial recognition systems have advanced much further, breaking older boundaries and setting new standards of prowess in facial recognition technologies. Facial recognition technology has helped in maintaining a highly accurate record of people, in taking better photo graphs, in identifying and searching people in a security perspective, using devices in a customized manner, in 3D modeling for making animations better, for successful plastic surgeries in a medical perspective and the extent of its utility has advanced much further after its advent.
From an information technology perspective, developing facial recognition technologies has been a gargantuan task as the real world is constituted of jumbled objects with little symmetry to be found. Identifying a face and extracting it as an image for analyzing and recognizing it through a database of faces is an extremely intricate process. Just being able to detect a face isn’t enough and for the technology to become practical, systems should also be able to recognize the faces. The last 50 years have showed us various applications of this technology and in one of its most important field of utility: law enforcement, the accuracy of these systems have been low at first due to an underdeveloped database and basic algorithms, but recent systems and algorithms to process and recognize images have optimized the accuracy and made the use of facial recognition systems much more plausible. In order to determine the rate of success and the dependability of facial recognition systems, there is a need to observe and analyze various systems in practice to ascertain which system would be an appropriate choice for future development.
Background of Facial Recognition Systems
The first facial recognition systems required the user to manually locate the required facial features in a face to accurately identify a face. The features required to locate exclusively were eyes ears, nose and the mouth. After recording these features the system took a reference point into account and measured the distance and ratios of the features with this point. This system was enhanced in the 1970s to record more features from a face like the thickness of lips, color of hair, etc. and these features were measured manually which made the process very slow (GOLDSTEIN, HARMON, & LESK, 1971). In these systems, similarities caused error at times and require manual intervention. During the 80’s, there was also an introduction of the technique which took principle component analysis into account and proved that the features of a face could be coded in less than 100 values which was considered a breakthrough at the time (Sirovich & Kirby, 1987).
Principal Component Analysis
For modern systems, the concept of principal components analysis (PCA) broke new ground and paved way to more complex systems that allowed the use of algorithms to decode face structures. This technique developed in 1991 used eigenfaces to detect face patterns from images. This system later helped to develop real time facial capture techniques (National Science and Technology Council, 2006). PCA analysis allowed to reduce the range of data by compressing it and allowed to formulate eigenfaces. Eigenfaces are orthogonal components formed with a face structure that discards irrelevant information. The data obtained from facial images are stored in a one dimensional array and requires the stored image data of the frontal part of the face in order to perform an accurate analysis (Kanti & Sharma, 2014).
Figure 1 PCA Method for Facial Recognition
Disadvantages of PCA
When faces are analyzed using PCA, there is a volatility in the image data of same people under influence of different kinds of light trajectories. This brings a level of inaccuracy in the data which hinders the dependability of this method. Changes in the position of the face and the expressions of a person also have an impact on the accuracy of the analysis which hinders facial recognition in a street view or surveillance in an external environment (Mahajan & Kaur, 2013).
Linear Discriminant Analysis
The Linear Discriminant Analysis relies on a statistical perspective on the classification and addition of new samples in the database relying on the already classified training samples. This technique generates facial graphs based on certain fiducial points which are a part of the graph that gully covers the face on the basis of these points. LDA manages to distinguish a class of images that overcomes the limitations of PCA by increasing the ratio of the determinant regarding the project samples. This statistical approach doesn’t give a direct result but the closest class of data. As it doesn’t deal with the whole database but the samples which gives faster results and a higher accuracy.
Figure 2 Face Classes Using LDA
Drawbacks of LDA
The major drawback of this method of analysis is that it may come across the problem of small sample size which causes problems in recognition due to the singularity of the within class scatter matrix. In certain cases, the face image data that is processed for storing, turns out to be more than usual due to a highly illuminated subject which causes variation in the face image pattern data. PCA outperforms LDA when sample size is small as LDA uses class discrimination while usually LDA should outperform PCA in other cases (Martinez & Kak, 2001).
Independent Component Analysis
Independent Component Analysis (ICA) is a statistical method of analysis used for facial recognition which utilizes underlying components from the statistical data lying in multiple dimensions. ICA performs better than existing system in cases where there are problems regarding illumination of subject and varying facial orientations (Bhele & Mankar, 2012). The search for a non-Gaussian component makes the ICA system unique. The similarity between ICA and previous methods is that is also derives a linear representation of data. When taking basic images into account, ICA displayed better performance than PCA. Like in every method, ICA also assigns independent features on a face but in ICA, the analysis begins after transforming the face into a vector. Just like the recent methods which combine two methods to achieve better results, optical correlation technique was taken up with ICA which gave a robust correlation (Bartlett, Movellan, & Sejnowski, 2002).
Figure 3 ICA Weight Matrix A = WI -1 Image Synthesis
Drawbacks of ICA
The drawback of ICA is just that it remains to be an exclusively statistical tool which is not enough for current requirements. ICA was superior in comparison to its statistical counterparts but with the arrival of superior systems it lost relevance as it could not improvise by itself like the other systems (Jafri & Arabnia, 2009).
Elastic Bunch Graph Matching Method
The Elastic Bunch Graph Matching (EBGM) Method is based on dynamic link structures. The concept of EBGM relies on the fact that facial structures cannot be always put into a linear manner and quantified for a statistical analysis to achieve perfect results. If a linear approach is taken, than a vast array of nonlinear elements both in picture and video mediums like lights, shadows, posture and expressions are not taken into consideration which hampers the accuracy of the process. The framework of a face is transformed into an elastic grid of a dynamic link structure. The nodes on a facial structures are recognized as Gabor jets which help in the detection of additional shapes and surfaces which change the behavior of the pixels. This complex process is achieved by the replication of the processes occurring in the visual cortex region (National Science and Technology Council, 2006).
Figure 4 Elastic Bunch Graph Mapping
Drawbacks of EBGM
The EBGM methods requires a landmark localization of accurate manner which can only be generated by the combination of PCA and LDA processes which tends to make the complete process time consuming.
Artificial Neural Networks and the Way Forward
The most advanced face recognition technologies are based on the mechanism of our brain and nervous system. As the face recognition systems advanced with time, the linear element got eliminated as faces could now be traced without linear concepts. The neural networks made to trace facial structure have a basis in the neural pathways of a human nervous system which help in the flow of signals through the brain. These structures contain nodes which transmit data in the same way as neurons do. The most vital feature of this system is that these neuron structure can learn on their own which is similar to the concept of artificial intelligence. The types of learning that occurs in a neural network is error based learning, memory based learning, or supervised (Programmable) learning. The similarity of this neural structure to neurons is that it has a process of learning and that it can store memory through connecting nodes (Szeliski, 2010).
Figure 5 Basic Architecture of an Artificial Neural Network
The neural network structure are also vital in executing complex functions like preprocessing of the image, feature extraction in recognition systems associative memory and recognition of patterns. Pattern recognition which one of the most important features in the process of facial recognition is easily possible through neural networks because of their ability to master nonlinear input output relationships of complex nature (Kanti & Papola, 2014).
Artificial Neural Networks in Facial Recognition
Artificial Neural Networks (ANN) are a bundle of nonlinear algorithms utilized for the extraction of the features from faces in the process of facial recognition. They are also used for classification of the images while storing new sets of data into a database. This complex information processing system facial recognition through complex procedures and advanced methods of analysis. Furthermore, ANN algorithms help in alignment of faces while viewing a photo or video for recognition and also normalization of image so that it can be put into a normal position if it is rotated or tilted. Perceptrons are the technology associated with ANNs that help in channeling desired behavior. Due to the utilization of perceptrons, the features that are detected get increasingly invariant and global in nature. The functions of the facial recognition system in ANNs are aligned in figure 5 and help in getting an idea how images go through the system.
How ANN is better than Other Systems
The ANN system is better than all previous algorithms due to the capacity to handle complex tasks easily and in the area of facial recognition, it can identify faces easily through large classes and samples of data. The faces which are registered through neural networks can be identified even if the person is wearing glasses, or facial hair, ornaments, caps, partial masks or even in low light conditions and if the person is being expressive through his/her face. The function is achieved with a higher rate of accuracy than any other system in the past like PCA, EBGM, LDA and ICA. This system learns on its own and communicates within itself like a neuron which makes the identification and classification process faster than any other system. ANN eliminates the drawback of every previous system and is an improvement upon the aspects of all the systems (Le, 2011).
As it has been observed through the analysis of the older systems in the area of facial recognition and the latest Artificial Neural Network algorithm, it can be ascertained that the latest system proves to be better than the old systems by improving upon old systems and removing the drawbacks of the old systems effectively. The way to the future has been paved further through the advent of neural networks and the utility of the facial recognition systems have been improved by ANN. This does not mean that the old systems have become obsolete but ANN can be further improved by integration with older systems like PCA, LDA, EBGM, etc. to further enhance the effectiveness of ANN networks and improve on functionality and accessibility of ANN. There has been research conducted on the fusion of older systems and ANN algorithm which show promise (Mahajan & Kaur, 2013) (Kanti & Papola, 2014). It can be said with conclusive proof that ANN has certainly changed the scenario of Facial Recognition systems and improved the utility of its applications in security systems.
Bartlett, M. S., Movellan, J. R., & Sejnowski, T. J. (2002). Face Recognition by Independent Component Analysis. IEEE Trans Neural Netw, 1450-1464.
Bhele, S. G., & Mankar, V. H. (2012). A Review Paper on Face Recognition Techniques. International Journal of Advanced Research in Computer Engineering & Technology, 339-346.
GOLDSTEIN, A. J., HARMON, L. D., & LESK, A. B. (1971). Identification of Human Faces. PROCEEDINGS OF THE IEEE (pp. 748-760). IEEE.
Jafri, R., & Arabnia, H. R. (2009). A Survey of Face Recognition Techniques. Journal of Information Processing Systems, 41-68.
Kanti, J., & Papola, A. (2014). Smart Attendance using Face Recognition with Percentage Analyzer. International Journal of Advanced Research in Computer and Communication Engineering, 7321-7324.
Kanti, J., & Sharma, S. (2014). Automated Attendance using Face Recognition based on PCA with Artificial Neural Network. International Journal of Science and Research, 291-294.
Le, T. H. (2011). Applying Artificial Neural Networks for Face Recognition. Advances in Artificial Neural Systems.
Mahajan, A., & Kaur, P. (2013). Face Recognition System using EBGM and ANN. International Journal of Recent Technology and Engineering, 14-18.
Martinez, A. M., & Kak, A. C. (2001). PCA versus LDA. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 228-233.
National Science and Technology Council. (2006). Face Recognition. Washington D.C.: National Science and Technology Council.
Sirovich, L., & Kirby, M. (1987). A Low Dimensional Procedure for the Characterization of Human Faces. Journal of the Optical Society of America, 519-524.
Szeliski, R. (2010). Computer Vision: Algorithms and Applications. Springer.