Disguised Face Recognition With Deep Learning

With the rapid development of deep learning and computer vision,face recognition has achieved tremendous progress over the past decade.Now,the state-of-the-art face recognition can surpass the human being in the recognition accuracy.While newly-proposed algorithms continue to achieve improved performance,a majority of the existing face recognition systems are susceptible to failure under disguise variations.Because the publically-available disguised datasets has insufficient number of face images,disguised face recognition algorithms have to employ a two-stage training approach.In its first stage,a large-scale face dataset is used to train a deep neural network for generic face recognition.The second stage is to fine-tune the trained mode with the disguised face dataset.This paper foucuses on the disguised face recognition and the main contributions can be summarized as follows.(1)The state-of-the-art methods in the field of face recognition at home and abroad are reviewed.We first introduce the process of face recognition,with empasis on the loss function,network architecture and datasets.Then,we present the history,the main problems and challenges for disguised face recognition.(2)A novel two-stage training approach for deep disguised face recognition is proposed.The first stage of the new algorithm is to train a convolutional neural network for generic face recognition with the powerful additive angular margin loss.Then,this network is finely adapted to the DFW dataset using a novel joint loss,which combines the triplet loss with a novel loss for further controlling the distances over positive pairs.Extensive experiments over the DFW testing dataset show that the accuracy of the proposed method under the three test protocols of DFW reaches 60.48%,82.88% and 82.04% respectively under the condition that false acceptance rate is 0.1%.(3)A novel piecewise margin loss function is proposed by combining the Additive Cosine Margin Loss function and the Additive Angular Margin Loss function.The analytic results shows that it can make the network to converge rapidly at the beginning stage of training,and in the same time it can force the angle between the embedded feature vector and the identification weight towards zero,which is known to be desirable for achieving smaller intra-class distances and larger inter-class distances.Experimental results show that the use of the proposed loss function in training can achieve better accuracy on the LFW dataset and DFW test set,compared to the use of the Additive Angular Margin Loss function.(4)To remedy the overfitting problem for disguised face recogintion,we propose to combine the existing face datasets,including CASIA-Web Face and LFW,with DFW for refining the trained model.Experiments show that the prposed methold can alleviate the overfitting problem and improves the final performance.