A new Rule to Constrain Convolution Neural Network Architecture in Face Recognition System

Abstract

Face recognition (FR) system is an essential part of a biometric security system, which runs faster than other security methods and done remotely. One of the most important techniques that used in FR is the convolutional neural network (CNN). Traditionally, the choice of the CNN architecture is achieved by experimental trails. In this paper, a new approach is proposed to build a mathematical model that helps to select a proper architecture. This model is built from the experimental results by applying different architectures on the well-known dataset (Vggface2). By changing the class number, image number and convolutional layer number, where the accuracy of each case is recorded. Finally, the proposed model is evaluated on the sets of datasets (Essex. FEI, Caltech), where the accuracies of (99.13, 98.51, 97.78) respectively, are achieved. The evaluation results proved that the proposed model is an efficient for many types of small and middle scale dataset.