Choosing the structure of convolutional neural networks for face recognition

Evaluating the number of hidden neurons and hidden layers necessary for solving of face recognition, pattern recognition and classi cation tasks is one of the key problems in arti cial neural networks. In this note, we show that arti cial neural network with a two hidden layer feed forward neural network with d inputs, d neurons in the rst hidden layer, 2d+2 neurons in the second hidden layer, k outputs and with a sigmoidal in nitely di erentiable function can solve face recognition tasks. This result can be applied to design pattern recognition and classi cation models with optimal structure in the number of hidden neurons and hidden layers. In addition, we propose a new type of convolutional neural network, which is capable to extract most powerful features. The experimental results over well-known benchmark datasets shows that the convergence and the accuracy of the proposed model of arti cial neural network is acceptable. Findings in this paper are experimentally analyzed on ve di erent face datasets from machine learning repository.