CNN integrated with HOG for efficient face recognition

Abstract

Human faces in the video are subject to illumination variation, out-of-focus blur and pose variations during face recognition process in various applications. The proposed system aims to eradicate the problems mentioned above. This is done by utilizing Histogram of Oriented Gradients algorithm as a feature descriptor to detect faces. The training data is composed of still images and blurred images. For the system to learn pose variations, an additional dataset of artificially aligned images is fed by using Face landmark estimations algorithm. Convolutional Neural network is trained, and effective face recognition is obtained. Thus, can make surveillance applications work efficiently.