Quaternion Discrete Racah Moments Convolutional Neural Network for Color Face Recognition

In this paper, we propose a novel architecture called quaternion discrete Racah moments convolutional neural network (QRMCNN) for color face recognition to improve the recognition accuracy and to reduce the computational complexity. Quaternion Racah moments (QRM) can extract effective features from color images. The use of QRM as first layer aims to reduce the input matrix size of the proposed architecture in the earliest orders and hence enormously decrease the number of parameters and the computational time. Experiments are carried out on faces96 and FEI face datasets to demonstrate the efficiency of the proposed architecture in the training process. The obtained results indicate clearly that the QRMCNN outperforms previous methods in terms of recognition rate and GPU elapsed time.