Research On Face Recognition Based On Multi-task Cascaded CNN And Metric Learning

With the continuous progress of society and the urgent requirement of rapid and effective automatic authentication,biometric recognition technology has developed rapidly in recent decades.Face recognition is a kind of biometric recognition technology based on face feature information.It is superior to other biometric techniques in reliability and stability and is less likely to be infringed.The main feature extraction methods in traditional face recognition algorithms are principal component analysis(principal component analysis,PCA),local binary pattern(Local Binary Pattern,LBP),Gabor transform,HOG direction gradient histogram feature and so on.These feature extraction algorithms have good experimental results,but these algorithms use artificial design features,not only involved in subjective factors,but also under the condition of big data,it is difficult to extract face features,resulting in a decline in generalization ability.Experience has shown that traditional algorithms can only be applied to small face databases.Nowadays,in the era of big data,there are thousands of faces to be recognized,as well as the faces of each person are different,complex and diverse,furthermore the deep convolution neural network has the ability of nonlinear description.For this reason,deep learning is introduced into the field of face recognition technology.However,face recognition technology in today's society faces many challenges,mainly as follows: Due to illumination,occlusion,posture,expression,scale and other non-restrictive complex scenes,the face intra-class gap increases,and the inter-class gap is narrowed;The computational complexity is high and the convergence rate of the model is slow.In view of the above problems,this paper has done some research:Firstly,research on face Detection algorithm.In the front end of the face recognition system,we use various image preprocessing(face detection and alignment as part of image preprocessing)to correct the pose and position of the face,and remove the part of the background environment of the face image.The size of each face image is unified so that the intra-class distance is shortened and the degree of differentiation between classes and within classes is increased.The experimental results show that the recognition rate of face detection and alignment is improved compared with the original dataset,and the effect of using mtcnn algorithm to detect face alignment is better than that of DLIB.Secondly,research on face recognition algorithm.In the loss function of face recognition model,the method of metric learning(usually Euclidean distance)is used to calculate the inter-class distance.The L2 and triple loss function are combined with back BP propagation to train and optimize the face recognition model.Then embedding learning is used to reduce the feature dimension of the model output.Experimental results show that the face recognition model can distinguish features by metric learning.Thirdly,a comparative study of two kinds of metric learning.Due to the slow convergence rate of the model with triple loss training,in order to reduce the large amount of calculation of triple selection,the paper proposes the combination of softmax loss function and center loss function to train and optimize the face recognition model by backward BP propagation.Compared with triplet loss,the center loss can achieve the same experimental effect and is easy to realize.