Region based 3D face recognition using a convolutional neural network

Abstract

In this paper we are proposing a compact CNN model for expression insensitive 3D face recognition. 3D face recognition is a trendy interested area in computer vision and is applied in different real time applications. Lots of research work is going on in industry and academia in this area. Traditional machine learning approaches for 3D face recognition is now superimposed by deep neural networks and are trained using large amount of data. We are applying a region based 3D face recognition approach along with a fusion CNN. 15 sub regions are generated from the frontal face region and features are extracted from it. The features extracted from each region are fused using the fusion CNN. These facial features are rich in identification information and they are not represented using single features. Fusions of multiple features extracted from the 15 regions are combined. The set of values extracted from the features after preprocessing are given as input to the CNN. The fusion CNN uses the features from different layers and fuses them together for a prediction as shown in figure1. The lower layer and higher layer features are fused. The computation time of the proposed system is 3.24 s for preprocessing and 0.09 s for matching. The overall computation time is 3.33 s. The running time of our previous region based approaches [13][14] is around 6 .48 seconds and 12 seconds. It is evident that the computation time of our proposed approach stands good and can be applied in time critical security applications. The three major steps are preprocessing, deep feature learning and deep feature classification.