Deep Features Learning For Face Recognition

Face recognition is a significant biometrics identification technique that aims to automatically determine the identity of a given face with facial features.With the de-velopment of artificial intelligence techniques,face recognition has been paid more and more attention by academia and industry.The key challenge of face recognition is to learn the effective feature representations that are discriminative and robust to intra-person variations.More recently,deep convolutional neural network(CNN)based rep-resentations,which have learned face representations with the depth structure of CNN by layer-by-layer nonlinear mapping,have achieved a state-of-the-art performance on LFW.However,unconstrained face recognition is still far from being a solved prob-lem.A key limitation exists in the available public domain datasets(such as LFW):the faces lack full pose variation.Under the new IARPA Janus Benchmark A(IJB-A)face recognition dataset that includes real-world unconstrained faces with full pose and illumination variations,There still remains a large gap between automated systems and human performance on familiar face.In this paper,we focus on the face recognition problem in dataset included faces with full pose variations,and exploit the activations from intermediate convolutional layer as the local deep features(LDF)for face recognition.In the statistical modeling phase,several techniques may be used to aggregate the LDFs into a face representations.Firstly,we utilize activations from intermediate convolutional layer as the local deep features(LDF)setting.Since the activations corresponding to the receptive field of input face image,the D-dimensional feature maps corresponding to a spatial unit as the feature vector in a spatial unit.The LDFs extracted from all spatial unit in convolu-tional layer output may be robust to the global image translations and rotations.When compared to several methods that utilize a deep CNN for multi-patch feature extraction which still need to handle errors introduced by automatical face detection and facial landmark localization algorithm,the LDFs extraction may be more efficient.Secondly,we proposed a Fisher vector(FV)coding method to implement the ef-fective aggregation of LDFs,which exploits the activations from successive convolu-tional layers to derive a probabilistic model.More specifically,it extracts activations of one convolutional layer as local deep features,and use feature maps of the succes-sive convolutional layer as indicator maps to guide the LDF features to be assign into several visual units defined by each feature map.A Single Gaussian model can be used to fit the LDF feature's distribution,and a Gaussian mixtures model(GMM)based on CNN(CNN-GMM)can be derived,which improve the modeling capability of tradi-tional GMM for high-dimensional LDF feature.Experimental results demonstrated that CNN base face FV representation significantly improve the performance.In addition,we proposed a mixture factors analyzers based fisher vector coding for face recognition.The proposed method firstly take a full covariance GMM model derived from the CNN activations,and then a MFA models can be obtained using this full-covariance GMM.The FV representation extracted by MFA model further improved the performance of face recognition system.Thirdly,in order to effectively reduce the face representation dimensional while implement aggregation of LDFs,we proposed a total variability modeling(TVM)method utilizing different output layers of CNN for face recognition.The high-dimensional face Gauss super-vector can be compressed into a compact total variation factors(iVector),which maintain performance advantage while greatly reducing the dimension of face representations.Finally,we proposed a cross-layer bilinear CNN to effectively combine CNN multi-layer information,which cover both front-end LDF feature extraction and back-end face representation learning.And the back propagation(BP)algorithm can be used to syn-chronization optimization for two modules.Compared to the face global deep feature representation extracted from fully connected layer(FC)of CNN,the proposed face representation has achieved significant performance improvement.