ABSTRACT
This research proposal presents face reconstruction and recognition technology in the three dimensional format with the help of artificial neural networks. This technology combines geometric interpretation, detailed survey and analysis of streams of data in order to process static and dynamic images to produce neural images using advanced computing system. 3D face reconstruction done with the help of interconnected circuit of artificial neurons is followed by the step of matching and verification against familiar faces. 3D face reconstruction and recognition technology finds extensive applications in a number of fields including forensics and ensuring secure transactions. The process is complex, systematic and helps to produce high quality images and information from random images. The emergence of 3D face reconstruction and recognition technology from two dimensional recognition system is attributed to advances in science and software technology.
Key Words: Back-propagation, neural, dynamic, reconstruction, recognition static
Introduction
Three Dimensional (3D) Face recognition is a developing and interesting concept that finds application in forensic science, data encryption and security. Its’ an advanced technique over 2D face recognition since it helps in accurate assessment.
The system for face recognition constitutes of two components namely the hardware and software. The input can be in the form of digital static images (pictures) or dynamic images (video frames). Face identification is carried out for identifying face by processing inputs from devices like video cameras or biometrics system and is used to check and recognize identity of people with the help of camera and 3D scanners
Efficient face recognition can be done by advanced techniques of processing fixed or static images and dynamic images. Owing to recent and incessant developments in software technology, several complex and accurate verification methods of recognition of facial traits and characteristics have come up. 3D face recognition uses more distinctive features of the face such as curves of eye socket, nose and chin for identification. The system needs to be well equipped with all the advanced facilities to be able to correctly identify faces by considering the three dimensional geometry of a sample face and verify it against the given set of parameters. 3D face identification can also be used in darkness and can recognize a subject with different view angles with the ability to recognize up to 90 degrees.
Three dimensional face reconstruction
3D face reconstruction and recognition is a step wise process that involves detection, alignment, measurement, representation, matching and verification. This complicated procedure is carried out with the help of neural networks. Neural network refers to system of artificial neurons designed to replicate the functionality of nervous system where neurons that transmit the signals sent through external agencies and transport it to the brain for their processing so as to extract meaningful information.
Human face comprises of about eighty characteristic parameters which includes the space between the eyes, width of the nose, shape of the zygotic bone and jaw width, height of the eyehole etc.
Figure 1: Geometric representation of the essential facial characteristics in humans
The unique combination of these parameters represents the person. These details are inserted in the database and are used to compare with the person’s facial characters for identification. Using neural network the system can be trained to recognize face from the picture by comparing it with the elements in the data set.
Normalization:
A three-dimensional face image comprises of a layout of three-dimensional co-ordinates that can be generated through structured lights or laser scanners. These images are taken from various directions, viewpoints and depths such that each image bears unique address. All the co-ordinates are transformed into a common system. This process is known as normalization and is a necessary aspect of the face recognition system. Thereafter, through the application of similarity measure, face matching and recognition are carried out using the Euclidean distances among feature vectors .
Pre-processing of images
For the pre-processing of images obtained through the camera, connectionist solution serves a good application. Therefore, it is quite understandable that there is high probability of the images getting blurred or blocked by some other object that comes in between. Therefore, the images obtained with the help of a camera need to be processed before they can be sent for recognition. Images can be processed through low, medium or high level processing methods.
Low level processing
Processing can be high level, middle level or at low level. Low level processing involves the techniques of edge detection, line fitting, thresholding, image smoothing and image intensification. In the first technique of edge detection, when an image is captured its edges are sharpened so as to enhance their visibility. Two-dimensional edge detection technique facilitates the recognition of the edges of the objects in a picture. Firstly, the difference in intensities of the local regions of an image are computed and recorded. The difference zones form the border between different objects or scene parts. The two most popular methods used for edge detection include the gradient and Laplacian methods .
In the gradient edge-detection method, the maximum and minimum in the first derivate of the image are computed for detecting the edges in the image. This method uses three kinds of filters, namely Robert, Prewitt and Sobel. However, the laplacian method uses the Mars-Hildreth filtration technique . This method detects edge of the image by noticing the zero-crossings found in the second-derivative of the image. The changes in intensity of the image found in the raw image are computed at varying scales and degrees. When this image gets filtered at an appropriate scale it gives the second-derivative of a Gaussian. It has been observed that these primary filters do not depend upon the orientation once the basic conditions get satisfied.
Intensity changes in an image can also be caused due to discontinuities in the surface structure, reflectance or illuminated boundaries which are spatially localized. These zero-crossings are not independent and there are certain rules devised for combining the zero-crossings of the image into a description of the image called the raw primal image.
There is a possibility of some noise disturbing the image quality. This noise can be considerably reduced with the help of the image smoothing technique. Image smoothing is best used when random noise is present, either caused by poor image capture equipment or by over-compression of the image. The kinds of filter to be used for smoothing can be grouped into various categories such as Box Filter, Gaussian Filter, Median Filter and Bilateral Filter.
The box filter is the simplest of all the image smoothing methods. Box filter smoothens the image by equally weighting a rectangular neighbourhood of pixels. The output pixel has a value equal to the average or mean of all the pixels located around it.
Gaussian filter is the most useful image smoothing technique. In this the output pixel (x,y) is weighted using the normal distribution method i.e. the local values carry weight. Here the sum obtained after each input point is convolved with the Gaussian kernel gives the output array.
The median filter replaces a pixel value by the median of its neighbourhood. It runs through every element of the input i.e. the image in this case and will perform the replacement of pixels.
The function of these filters is to smooth the image obtained for better understanding. However, at certain times they also dissolve away the edges of the image which deteriorates the quality of the image. The bilateral filter is responsible to avoid the dissolving of the image. Similar to the Gaussian filter, the bilateral filter also takes into consideration the weighted value assigned to the neighbouring pixels. At the same time, the bilateral component also takes into account the difference in intensities of the neighbouring pixels and the one under consideration .
The image intensification method is used to enhance the intensity of the image’s brightness. If an image is, for example, too dark to identify detail, a non-linear boost can be used. It boosts all intensities but special emphasis is laid on the lower intensities more than the higher ones. This process is known as Contrast Stretching. It is carried out using the expression shown below.
fx=x1y
Next, the Histogram Equalisation further facilitates in the image intensification process. The output image uses all intensity values and the exact number of pixels is assigned to each gray level.
The line detection technique plays a very important function to compensate for the loss of information about the image when an object is partially hidden by some other object e.g. a house is partially hidden by a tree. In these cases edge detection does not prove to be very useful and hence line detection method needs to be applied.
In order to ascertain the flow of the edges and check if they would join if extended, the Hough Transform method is used. It uses the following expression.
A straight edge is given by the formula (y = mx + c), where m is the slope and c is the y intercept. The method edge(x,y) is a function that returns 1 if there is an edge point at (x,y), 0 otherwise. Once the line segments have been found, the image can be split into the regions with similar patterns such as brightness.
Thresholding can be used to obtain binary images from the grey scale images. It chooses some pixels as foreground and the remainder as background. The threshold value is chosen to highlight certain features. The grey-tone values that lie above or below the threshold become foreground pixels. The threshold values provide a useful measure to filter the desirable elements from the undesirable ones based upon a set of predefined criteria.
Medium Level Processing
Medium Level Processing constitutes
- Colour segmentation
- Shading and Texture
- Depth and motion
Colour segmentation:
Colours find significant application in complex pictures by helping in object detection or higher image resolution. Histograms and fillers are used to extract homo-coloured regions in an image. A hue histogram is used to divide the image in chromatic region while an intensity histogram works in case of achromatic region. To recover over-segmentation, region growing techniques are used.
Shading:
Shading is a technique of mid-level processing which is derived from the fact that all objects are characterized by a natural albedo. The term ‘albedo’ refers to the ratio of total reflected illumination (spectral or diffuse) to the incident illumination. It must be remembered that all objects darken with distance.
Texture:
The third feature i.e. texture of a picture enables the segmentation of a picture into areas of interest. It uses either structural approach where texture is a set of primitive texels in some repeated or regular relationship. Alternatively in statistical approach, texture is a quantitative measure of the way the intensities of the region are arranged. Other important features of mid-level processing are depth and motion.
High Level Processing
High Level Processing includes object recognition which may be content based, relationship mapping, affine mapping and image distance measures. Content based recognition utilizes feature like colour percentages, layout, composition, texture and shape. This feature finds application in micro-electrics for board, favour and dye-inspection.
Iris identification is an intelligent technology that uses the picture of a person’s iris to create a digital code in order to ensure security of transactions. Firstly the eye is scanned by video cameras and during this process the 266 different characteristics of iris are measured. Using demodulation the unique ‘Iris Code’ is calculated. Finally identification is done by using statistical independence.
Micro21 cell identification is another example of intelligent technology that reduces time and efforts spent in tedious microscopic reviews. Raw images are fed to an automated microscope. With the aid of a central computing system neural images are generated and then transmitted to the review station. From here stream of data is collected and organized to form reports.
Neural networks
Connectionist solutions comprise of artificial neural networks. Neural network is a simplified representation of the brain done artificially so as to imitate its functioning mechanism. Human brain is composed of numerous neurons that carry the signals sent through external agencies and transport it to the brain, where these signals get processed into useful information. Similarly the neural networks involve several units that have the ability to measure the strength of their inter-connections .
Neural networks have been successfully applied in the areas of face recognition, speech recognition, detection of grammatical errors in simple sentences etc. The data obtained or the signal has an associated level of strength or weight that determines its character. This signal can represent either positive or negative value depending upon the nature of activation. This variation in strengths distinguishing one signal from another imparts a fuzzy character to it .
Neural networks offer considerable flexibility and are capable of facing the real-time challenges. In such occasions where the inputs have certain defects or noise, this method offers the possibility to correct the input and impart it a usable form .
Figure 2: Face Recognition using neural network
The neural network algorithm known as the back-propagation algorithm can be successfully used in a variety of applications including face recognition. It makes use of the gradient descent method to sort out the minimum of the error function in weight space. The function of the combination of the weights so obtained is to minimize the error function. It is regarded as a solution of the learning problem. The continuity as well as differentiability of the error function has to be compulsorily ensured. This is so because in the learning problem the gradient of the error function is computed at every iteration step .
With the help of the backpropagation the chain rule is reduced to the recursive computation. The values obtained from earlier computations are stored by the network of computing units and thus they act as the underlying data structure. The network is run backwards and the nodes are labelled with the backpropagation error. Therefore, by applying the concept of dynamical object evaluation by a network of computing members and backpropagation as an inversion of network dynamics, the tedious redundant work gets simplified greatly .
Face recognition mechanism and numerical evaluation
Face Identification mechanism can be carried out with the help of two kinds of approaches which are symbolic solution and connectionist solution. The symbolic solution involves a combined effort of the various disciplines of cognitive science, artificial intelligence together with an efficient human-computer interface. Instead of numbers, the symbolic systems manipulate symbols . Symbols depict crisp values and therefore statistical methods can be incorporated for symbolic method.
The data set must contain records of all the faces against which comparison is to be carried out. The parameters are calculated using the distance formula, given below.
2x2-x12+y2-y12+z2-z1
Where, (x1,y1,z1) and (x2, y2, z2) are the coordinates of the two end-points.
The data so obtained then needs to be normalized i.e. corrected in manner such that it gains a value within the interval 0-1 using the following expression.
Xn=X-XminXmax-Xmin
Where, X represents the value that needs to be normalized while Xmin and Xmax are the minimum and maximum values of X.
Figure 3: Facial parameters
The software Abrosoft FaceMixer can be used to successfully carry out the task of face recognition. Another software known as the Neuroph Studio can be applied for training the neural network for conducting face recognition task.
Figure 4: Window of Abrosoft FaceMixer performing the task of face identification with the help of coordinates.
The data about the depth in the face is a great source of detailed information. The information about the facial curvature can be considered as one of the most important personal information. With the help of the Principal Component Analysis using surface curvature it can be ensured that the actual dimensions of the image can be reduced without any significant loss of the original information. From the maximum and minimum curvatures obtained, the eigenface is recognized. Cascade architecture of the fuzzy neural network is used to classify the faces. This is due to the reason that it can efficiently guarantee a high recognition rate and parsimonious knowledge base.
Figure 5: Flowchart representing the steps in face recognition system
Three-dimensional point cloud registration and matching can be carried out using a variety of processes. The Iterative Closet Point (ICP) Algorithm is used for three-dimensional point cloud registration. The ICP algorithm is used to generate the registration strategy that can be used for data normalization. The Principal Component Analysis can be used to generate the central axis of the face model followed by the use of the ICP algorithm for refining the registration. Euclidean distance is also used for the matching process. The ICP algorithm can be used for face alignment as well as for face matching. To carry out the recognition tasks the back-propagation neural networks need to be constructed.
Figure 6: Back-proportioning in neural networks
Different search loops can be run for each of the segments of the face. In the given situation, the captured images should first be suitably processed using the various techniques to obtain a clearer and more understandable version. Thereafter, this data needs to be compared with the existing database. The search for the subsequent code shall begin only after a successful match for the previous feature has been achieved. Also, after each step the comparing database will become shorter eliminating the sets that went unmatched. Therefore, as the database reduces in size, the steps will become faster with time.
For a system where the images are perfectly clear, the traditional statistical method of searching and comparing the each data would serve the purpose. However, in a real-life situation where there are several possibilities of images getting blurred or any hurdle obstructing the process of capturing images. Even after several attempts of processing the image may not have perfect clarity to a degree such that its comparison with the database elements could be carried out with ease. In such a situation, the incorporation of fuzzy logic in the search-mechanism can enable enhanced search-ability. The subsequent quantization of the elements according to certain threshold values can ensure much higher degree of accuracy. However, artificial intelligence solutions like neural networks offer various methods to rectify the errors that may crop up in real-time. Artificial intelligence improves the efficiency of the system remarkably as compared to the traditional method. Furthermore, neural network techniques reduce the need for external supervision to the minimum enhancing the ease of workflow.
Applications
The applications of the face recognition technology can vary from the governmental to commercial. Commercial uses range from face recognition technology uses in banking, voter verification, residential security to healthcare and gaming industries. On the other hand, government uses include face recognition technology for immigration, legislature, prisons, security or counterterrorism etc. The surveillance images are compared to the existing database of known terrorists. The ranked list of matches so obtained is then derived. It facilitates the identification and recognition of the terrorists in a speedy and efficient manner .
Face recognition technology provides robust detection capabilities to aid forensic department. It helps to augment the effectiveness of available forensic data by
- Incorporating operators as an integral system of the system’s process to enhance images
- Correcting for pose and alignment\
- Encoding local features that can be used by the algorithm.
Facial recognition capability is also beginning to appear on smart phones of police officers introducing a new domain known as ‘mobile facial recognition technology’.
Figure 7: Face recognition using the biometric methodology
Conclusion
It can thus be seen that 3D face reconstruction and recognition provides details of higher accuracy than its 2D predecessor. This technology by assembling scientific and software technologies helps to process random images using advanced computing systems. Random digital images of the face whether static or dynamic can be processed using artificial neural networks to produce high quality images. Face and facial expression recognition is done with the help of geometric interpretation of facial parameters like sockets under eyes, face texture etc. Computing systems simplify the procedure as they utilize advanced neural networks to process the raw images. Inspite of its potential abilities in providing advanced security during transactions and personal identification, it faces certain operational challenges like illumination, head pose and occlusion.
Bibliography
A., S., 2008. Using 3D pose alignment tools in forensic applications of Face Recognition. Biometrics: Theory, Applications and Systems,, 2nd(IEEE International Conference), pp. 1 - 6.
Bagchi, T. P., 1993. Taguchi Methods Explained-Practical Steps to Robust Design. s.l.:Prentice Hall.
Blackburn, D. M., 2001. Face Recognition: The Technology and Its Applications. [Online] Available at: http://simson.net/ref/biometrics[Accessed 14 August 2013].
Chafik, S., Anuj, S. & Mohamed, D., 2006. Three- Dimensional face recognition using shapes of facial curves. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(11), pp. 1858-1863.
Chang, K. I., Bowyer, K. W. & Flynn, P. J., 2006. Multiple nose region matching for 3D face recognition under varying facial expression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(10), pp. 1695-1700.
Hildreth, E. & Marr, D., 1980. Theory of Edge Detection. [Online] Available at: http://rspb.royalsocietypublishing.org[Accessed 1 July 2013].
Image Processing, n.d. Smoothing Images. [Online] Available at: http://docs.opencv.org[Accessed 1 July 2013].
Jain, D. C. & Pawar, V. P., 2012. A Novel Approach For Recognition Of Human Face Automatically Using Neural Network Method. International Journal of Advanced Research in Computer Science and Software Engineering, 2(1).
Juang, C. F., 2004. A hybrid of genetic algorithm and particle swarm optimization for recurrent network design. IEEE Transactions on Systems, Man and Cybern., 34(2), pp. 997-1006.
Kennedy, J. & Eberhart, R., 1995. Particle swarm optimization. IEEE Conference on Neural Networks, Volume 4, pp. 1942-1948.
Kim, J., Choi, J. & Turk, M., 2005. Effective representation using ICA for face recognition robust to local distortion and partial occlusion. IEEE transactions on pattern analysis and machine intelligence, 27(12), pp. 1977- 1981.
Lacher, R. C., Hruska, S. I. & Kuncicly, D. C., 1992. Backpropagation learning in expert networks. IEEE Transactions on Neural Networks, 3(1), pp. 62-72.
Lee, Y.-H. & Han, C.-W., 2006. 3D Facial Recognition Using Eigenface and Cascade Fuzzy Neural Networks: Normalized Facial Image Approach. First International Computer Science Symposium in Russia, Volume 3967, pp. 457-465.
Lin, C.-J., 2008. 2D/3D Face Recognition Using Neural Networks Based on Hybrid Taguchi-Particle Swarm Optimization. Intelligent Systems Design and Applications, 2008. ISDA '08. Eighth International Conference on, Volume 2, pp. 307 - 312.
Lin, C.-J., Wang, J.-G. & Chen, S.-M., 2011. 2D/3D Face Recognition using Neural Network basedon Hybrid Taguchi-Particle Swarm optimization. International Journal of Innovative Computing, Information and Control, 7(2), pp. 537-553.
Lin, C. J. & Xu, Y. J., 2007. A self-constructing neural fuzzy network with dynamic-form symbiotic evolution. International Journal of Intelligent Automation and Soft Computing, 13(2), pp. 123-137.
Liu, C., 2004. Gabor-based kernel PCA with fractional power polynomial models for face recognition. IEEE Transactions on pattern analysis and machine intelligence, 26(5), pp. 572-581.
McCarthy, J 2007, ‘Applications of AI,’ What is Artificial Intelligence?, viewed 24 April 2013 <http://www-formal.stanford.edu>
Morelli, R, ‘CPSC 352 -- Artificial Intelligence,’ Introduction to computing, viewed 24 April 2013 <http://www.cs.trincoll.edu>
Neuroph, n.d. FACE RECOGNITION USING NEURAL NETWORK. [Online] Available at: http://neuroph.sourceforge.net/tutorials/FaceRecognition/FaceRecognitionUsingNeuralNetwork.html[Accessed 30 July 2013].
Niu, B., Shiu, S. C. K. & Pal, S. K., 2005. A Novel 3D Face Recognition Algorithm Using Template Based Registration Strategy and Artificial Neural Networks. Pattern Recognition and Machine Intelligence, Volume 3776, pp. 315-317.
Obaidat, M. S. & Macchiarolo, D. T., 1994. A multilayer neural network system for computer access security. IEEE Transactions on Systems, Man, and Cybernetics, 2(5), pp. 806-813.
Oravec, M. & Pavlovicova, J., 2004. Face recognition methods based on principal component analysis and feedforward neural networks. IEEE International Joint Conference on Neural Networks, Volume 1, pp. 25-29.
owlnet.rice.edu, n.d. Other Methods of Edge Detection. [Online] Available at: http://www.owlnet.rice.edu[Accessed 1 July 2013].
Rojas, R., 1996. The Backpropagation Algorithm. In: Neural Networks. Berlin: Springer-Verlag, pp. 151-184.
Senior, A. W. & Bolle, R. M., n.d. FACE RECOGNITION AND ITS APPLICATIONS. New York, USA: Yorktown Heights.
Shri, Y. & Eberhart, R. C., 1999. Empirical study of particle swarm optimization. Proceedings of the 1999 Congress on Evolutionary Computation, pp. 1945-1950.
Stanford Encyclopedia of Philosophy, n.d. Connectionism. [Online] Available at: http://plato.stanford.edu[Accessed 1 July 2013].
Stanford University, n.d. Symbolic Systems Program. [Online] Available at: http://symsys.stanford.edu[Accessed 1 July 2013].
Sun, T.-H. & Tien, F.-C., 2008. Using backpropagation neural network for face recognition. ScienceDirect, pp. 361-372.
V., V. K. & Kumar, G. S., 2009. Face Recognition using Probabilistic Neural Networks. Cochin University of Science and Technology.
Yang, MH, Kriegman, DJ and Ahuja, N 2002, ‘Detecting Faces in Images: A Survey,’ IEEE transactions on pattern analysis and machine intelligence, vol. 24(1). Like common
Yuen, P. C., Dai, D. Q. & Feng, G. C., 2004. Face recognition by applying wavelet subband representation and kernel associative memory. IEEE Transactions on neural networks, 15(1), pp. 166-177.
Zhou, D., Yang, X., Peng, N. & Wang, Y., 2006. Improved- LDA based face recognition using both facial global and local information. Pattern Recognition Letters, 27(6), pp. 536-543.
