Feature extraction and fusion for face recognition systems using pre-trained convolutional neural networks

Abstract

Recently, face recognition applications achieved promising results by using Convolutional Neural Network (CNN). CNN has the capability to extract features automatically from images and does not need to extract hand-crafted features as traditional algorithms. Feature fusion aims to provide improvements of data validity for both traditional algorithms and deep learning algorithms. In this paper we propose a feature fusion approach for face recognition, the approach performs fusion at the feature level by applying two pre-trained CNNs AlexNet and ResNet-50. Firstly, extracting the feature from both pre-trained CNN AlexNet and ResNet-50 separately. Secondly, fuse the feature maps learned from AlexNet and ResNet-50. Finally, a Support Vector Machine (SVM) classier is used for the classification task. Experiments are conducted on the following datasets: FEI face, GTAV face, ORL, F_LFW, Georgia Tec Face, LFW, DB_Collection, demonstrate the effectiveness of the proposed approach. In addition, the fusion of the two CNN based models AlexNet and ResNet-50 lead to significant performance improvement. In particular, the fusion approach achieves accuracy in range (96.21%-100%) on all datasets.