NEURAL NETWORK FACE IDENTIFICATION DAVID MALANIK, ROMAN JASEK Department of Informatics and Artificial Intelligence Tomas Bata University in Zlin Nad Stráněmi 4511, 760 05 Zlín CZECH REPUBLIC {dmalanik;jasek}@fai.utb.cz Abstract: - The identification of peoples represents the actual problem of this World. The problem is the methods for contact-free or extremely user-friendly methods. The potential solutions of this problem are based on the method, that doesn t require the direct contact. This paper deals with the face identification method. These methods have a marginal problem with the quality of the tested sample. The quality might be repaired with the neural network. Key-Words: - Neural networks, Hopfield, Face detection, Identification, Biometric. 1 Introduction The problem of identification represents very important issue of the present World. There is potentially danger which flowing from the current World situation. The identification systems might detect any people identities inside the specific area. The place might be an airport or some government building. The problem of the contact-free face identification for a long range is the quality of the identification samples. The next potentially problem might be represented by the different age of the identification people. These factors could limit the commonly used comparison methods [1]. 2 Problem Formulation The problem of the face identification is inside the quality of the picture and the age of source sample. These both problems are based on the pictures. The scanning methods have their own limits in quality of pictures and the second problem might be in patterning of the source data. The captured picture must be adapted to specific position and size. The identification figure or picture must be in the correct direction and must have the same size. The resizing and direction correcting is quite difficult. The marginal role plays the problem of quality. The contact free identification based on the face identification is popular in the places with the high movement or high density of people [2]. The best examples might be the public part of airports, a government meetings and a public celebration. 2.1 Quality of scanning/capturing There are many factors, which limited the quality of captured picture of the face. The commonly described factor is the factor of environment [3]. The external environment might be affected by nature. The rain or the fog might produce the less quality of the face picture. The authentication method must repair these pictures before the authentication method might continue with the verification or identification of these people. The picture with the upper described deformation represents the Fig. 1. Fig. 1 - low quality face picture The low quality picture might provide the false identification of approved people or the positive identification of non-certificated people. These ISBN: 978-1-61804-028-2 129
false identifications might affect the security steps. The face identification without any repair functions miss affects the identification. This paper describes the one the possible methods for reconstruction face identification picture with the self-repair neural network. Next parts will describe the procedures for improvement of the picture quality. The paper includes the test results from described solution. 3 Neural network reconstruction and identification The neural network represents the huge potential of availabilities, which might help us with many problems inside the security branch. There are many types of networks. Some of these networks have very specific abilities. The most interesting ability is the package with selfrepaired function. The one of the most popular self-repaired neural networks is The Hopfield neural network [4,5]. The basic structure of this network is on Fig. 2. face from the gray-scale or monochromatic picture [7]. The typical characteristics are the measurement of the eye position, position of nose, the size of nose, lips etc. Each of these measure-based methods required the picture of searched face [7]. Each of these algorithms deals with the quality of source with other methods and principles. 3.2 Hopfield network construction The neural network for the picture reconstruction is based on the Hopfield neural network. The testing samples were pictures of 6 different faces. There were some woman s faces and some man s faces captured from different positions. These faces are shown on Fig. 3. Fig. 3 - Source faces Fig. 2 - The Hopfield network structure [5] The network tries to repair the damaged source to one of the known originals. The Hopfield network might repair only the previously learned samples. 3.1 The human face The human face represented the complex biometric system for identification. The usability is done with the easy contact-free identification methods. For the user identification based on the specific face characteristics are used many methods [6]. The biometric system can detect and identify the The neural network contained 2500 neurons. Each neuron represented the one pixel of source image. The next step was learned the network with these 6 samples. The result from the learning has this description "Continuous Hopfield model for pattern vectors with 2500 components. Created 2011-5-4 at 15:35. Type of neuron: SaturatedLinear. 3.3 Face reconstruction The next step was producing the images with the lower quality. There was used a noise for simulation the nature problems such as the rain, the fog and the limited visibility. The level of noises was changed in interval <0.20;0.95>. The testing procedure contains the learned neural network (network was learned with the samples of faces). These experiments were ISBN: 978-1-61804-028-2 130
realized 10 times and the followed picture represents the outputs from the reconstruction. The Fig. 4 represents the very weak noise. The noise does not affect each sample. Some samples are still relatively clear and processing system might identify it. Fig. 6-40% of noise Fig. 4-20% of noise The reconstruction in the Fig. 7 shows, that the neural network might repair these damaged samples to the original state. The Fig. 5 represents the reconstructed picture from the neural network. These reconstructed samples are the same that the source samples. Identification is quite simple in this case. Fig. 7-40% of noise - reconstruction Fig. 5-20% of noise - reconstruction The next figure describes the samples with 40% of the noise. This level made marginal changes in some samples, especially the first and the second from the left in the second line. The identification of these samples is quite complicated, because samples lost some of their biometric characteristics. The Fig. 8 shows the medium level of the noise, the noise is at 60% level at this figure. The source samples are damaged and the human operator does not recognize the original samples. The pixel comparison method does not work, because the samples lost to much biometric identification markants [7]. The pictures are damaged. ISBN: 978-1-61804-028-2 131
repairs well, but there were false positive identification of sample. The second sample from the bottom right was identifying as the same that was second sample from the top left on the Fig. 11. This might provide the false positive identification of sample that does not approved for this area/work. Fig. 8-60% of noise The result from the neural network repair procedure is shown on the Fig. 9. The network can repair the original samples with extremely accuracy. The pictures were rebuilder to their original state. The identification system might analyse these samples and provide the positive or negative identification of selected peoples. The huge potential of the neural network selfrepair function is represented by these results. Fig. 10-80% of noise Fig. 11-80% of noise - reconstruction Fig. 9-60% of noise - reconstruction Next figures, starting with the Fig. 10 are not applicable for user identification. The level of damage is too high and perception of false positive identification rising marginally. The next figures represented only the ability of selfrepaired in extreme point of view. The Fig. 12 represents the potential limit for the neural network self-repair function. There are 2 damaged samples. The first sample from the top right is deformed and combined with other. The sample received some characteristics from the other one. So as the Fig. 13 shows, some samples are still reconstructing able and usable for identification. But the potential risk of false or positive identification is marginal. The Fig. 10 represents the samples with 80% of noise and the result from the neural network repairing. The less complicated samples were ISBN: 978-1-61804-028-2 132
Fig. 14-95% of noise Fig. 12-90% of noise Fig. 15-95% of noise - reconstruction Fig. 13-90% of noise - reconstruction The Fig. 15 represents the result; where is 3 maybe 4 false positive identifications. As it shown on the Fig. 15, there are 4 samples reconstructed to the one original. The figure contains 4 woman s faces with the same characteristics and 2 man s faces, but one of the man s samples is huge deformed. 4 Conclusion The previously shown results represent the ability and function of the neural network for the image processing and reconstruction. The reconstruction of the images; in these case face pictures, is important for the contact-free identification of people. The biometric identifications methods like the face identification represents the solution for the places where is impossible to identify each people with the biometric contact scanner. The main argument for the contact free identification might be the speed of identification. The camera just captures some pictures and system might detect faces and compares it with internal databases. The contact is not necessary. The controlled people could not know about this identification system. The system has one marginal issue. The issue is the quality of captured picture. This paper deals with this problem and showing the possibilities of usage the neural network for successfully image reconstruction and next processed identification of people based on the face identification. Samples represented the experiment, which had a specific purpose. The purpose was try to find the potentially limitation of neural network reconstruction and verification people with specific biometric characteristics such as their face. The contactfree identification represents the future of the ISBN: 978-1-61804-028-2 133
people identification [7]. These methods are addicted on the pictures and their quality. The neural network reconstruction might represent the image pre-processing subsystem for these identification methods. Acknowledgement: The research work was performed to financial support of grant reg. No. IGA/57/FAI/10/D and by the European Regional Development Fund under the Project CEBIA-Tech No. CZ.1.05/2.1.00/03.0089. References: [1] TIPTON, Harold F; KRAUSE, Micki. Information security management handbook. 6th ed. Boca Raton : Auerbach, 2007. 3231 s. ISBN 0-8493-7495-2 [2] GRAHAM, Benjamin; DODD, David L. Security analysis : principles and technique. 6th ed. New York : McGraw-Hill, 2009. 766 s. ISBN 978-0-07-159253-6 [3] BITTO, Ondřej. Šifrování a biometrika, aneb, Tajemné bity a dotyky. Vyd. 1. Kralice na Hané : Computer Media, 2005. 168 s. ISBN 80-86686-48-5 [4] ZELINKA, Ivan. Umělá inteligence aneb úvod do neuronových sítí, evolučních algoritmů. 1. vyd. Zlín : Univerzita Tomáše Bati, Fakulta technologická, 2005. 127 s. ISBN 8073182777. [5] ZELINKA, Ivan; OPLATKOVÁ, Zuzana; ŠENKEŘÍK, Roman. Aplikace umělé inteligence. Vyd. 1. Zlín : Univerzita Tomáše Bati ve Zlíně, 2010. 151 s. ISBN 978-80-7318-898-6 [6] MEYERS, E. and WOLF, L. Using Biologically Inspired Features for Face Processing. International Journal of Computer Vision. 76, 1 (Jul. 2007), pp. 93-104, DOI: 10.1007/s11263-007-0058-8. [7] MALTONI, D. and ECCV 2004 International Workshop. Biometric authentication : ECCV 2004 International Workshop, BioAW 2004, Prague, Czech Republic, May 15th, 2004 : proceedings. Springer, ISBN: 9783540224990. ISBN: 978-1-61804-028-2 134