VGG based Convolutional Neural Network for Disguised Face Recognition

Abstract

Nowadays, with the increase in the number of data, including image, signal and audio data, the need for computer systems and automated algorithms seems more than ever. The field that automates this operation is called machine learning, which is a subset of artificial intelligence. In recent years, new methods of deep learning have been proposed that are based on neural networks and have a large number of layers. Disguised face processing is one of the subsets of image processing that has many applications such as identifying criminals. Convolutional neural networks are among the deep learning structures that have been very successful in image processing. In this article we used Disguised Faces in the Wild (DFW) which contains approximately 11000 images from 1000 subjects. These images include normal faces and also face with different lightning, pose, background and different make up. We used batch normalization and dropout to overcome overfat. We used activation function RELU that leads negative numbers to 0 and its function is: max (0, x) which means that it would be never saturated in positive region. In addition, ADAM (adaptive momentum estimation) was used as optimization algorithm. Our proposed architecture has 16 layers including 14 convolution layers and two fully connected layers. Finally, the proposed method reached 84 percent accuracy.