ACCESS CONTROL BY FACE RECOGNITION USING NEURAL NETWORKS*
Dmitry Bryliuk and Valery Starovoitov Institute of Engineering Cybernetics, Laboratory of Image Processing and Recognition Surganov str., 6, 220012 Minsk, BELARUS E-mail: email@example.com, firstname.lastname@example.org A Multilayer Perceptron Neural Network (NN) is considered for access control based on face image recognition. We studied robustness of NN classifiers with respect to the False Acceptance and False Rejection errors. A new thresholding approach for rejection of unauthorized persons is proposed. Ensembles of NN with different architectures were studied too. Advantages of the ensembles are shown, and the best architecture parameters are given. The explored NN architectures may be used in real-time applications.
Introduction Access control by face recognition has the following advantages in comparison with other biometrics systems. There are no requirements for expensive or specialized equipment, a system may be built using a simple video camera and a personal computer. The system is passive. There is no need to touch something by fingers or palm, no need to say any word or lean eye to a detector. Any person just may walk or stay before the camera, and the system performs recognition. It is especially useful in everyday usage. Also it has advantages in different extremal or non-standard situations, when it is impossible or inconvenient to took other biometric characteristics, for example when catching criminals. The recognition performance of a simple face recognition system is not the best in comparison with other biometric-based systems, and such a system can be relatively easy deceived. But using a face thermogram or output of an infrared camera, the system can achieve very high recognition rate and robustness to deceiving. The face thermogram is strictly individual for every person, it does not change when lighting condition are changed, and it is impossible to deceive even by plastic operation. The ultra-high security access is based on face thermogram recognition. There are many works devoted to the face recognition problem. But most of them are oriented on obtaining higher recognition rate for some test face databases in the prejudice of identification robustness and stability for real applications. The rejections of misclassified or unauthorized persons are not studied well. We tested the ORL face database (www.cam-orl.co.uk/facedatabase.html). It has 40 persons and 10 different images for each person, 92x112 pixels with slightly varying lighting conditions, pose, scale, face expression and presence or absence of glasses. No attempts to normalize these images were made.
This work has been partially supported by project INTAS 00-626.
In work  the multilayer perceptron neural network was used. It has one hidden layer with number of hidden units varying from 60 to 80. The input of neural network was a set of discrete cosine transform coefficients. The first 30 coefficients from 10304 were used. The achieved recognition rate was from 94% to 97%. In work  the convolution neural network were used. It has sophisticated architecture for image recognition. The input of such network was whole image. Reported recognition rate was from 96% to 98.5%. The Pseudo-2D Hidden Markov Models (P2D-HMM) were used in work . Reported recognition rate was from 98% to 100%. All this works used ORL database. The first five images of each person were used for training, and the last five – for testing. There were no any attempts neither to estimate the reliability of classification, nor to develop the rejections for unauthorized persons or for unreliable cases for authorized persons. For neural networks and P2D-HMM’s there were used the maximum response rule, when a unit of the output layer (or particular model for HMM’s) with maximum value indicates recognized person. The main question is how reliable...
Please join StudyMode to read the full document