Occluded Facial Recognition with 2DPCA based Convolutional Neural Network

Abstract

Face occlusions with glasses or scarf are quite common in the real-world scenes, or more seriously, terrorists often cover their faces with sunglasses or a mask to hide themselves from the cameras. Occluded facial recognition is, therefore, an important problem in surveillance & defense department. A system that can recognize faces with occlusions may need to be trained by a huge set of facial databases. To reduce the complexity of an occluded facial recognition system, this paper investigates the effects of the two-dimensional principal component analysis (2DPCA) in the initialization phase on image classification by the convolutional neural network (CNN). Our experiments show that 2DPCA can reduce the image dimension for training while keeping the accuracy rate comparing to using the whole images. Our results, at 0.001 learning rate, showed 81.91% accuracy with 120 eigenvectors for the AR database, and 99.95 % accuracy rate with 190 eigenvectors for the GTAV database.