Optimization of Convolutional Neural Network Structure for Biometric Authentication by Face Geometry

The article presents development of the methodology of using a convolutional neural network for biometric authentication based on the analysis of the user face geometry. The need to create a method of the structural parameters of convolutional neural network adaptation to the expected conditions of its use in a biometric authentication system is postulated. It is proposed to adapt the convolutional neural network structural parameters based on the maximum similarity to the process of recognizing a human face image by an average user considering peculiar properties of computer input and display. A group of principles for optimization methods is formulated by combining this assumption with the generally accepted concept of a convolutional neural network constructing. The number of convolution layers should be equal to the number of the person image recognition levels by an average user. The number of feature maps in the n-th convolutional layer should be equal to the number of features at the n-th recognition level. The feature map in the n-th layer, corresponding to the j-th recognition feature, is associated only with those feature maps of the previous layer that are used to build the specified figure. The size of the convolution kernel for the n-th convolutional layer should be equal to the size of the recognizable feature on the n-th hierarchical level. Based on these principles, a method of the structural parameters optimization of a convolutional neural network has been developed. Advisability of these principles use has been proved experimentally.