Research On The Key Technology Of Automatic Face Recognition Under Unrestricted Scenes

As an efficient non-touched biometric identification technology,automatic face recognition is one of the hot research topics in the field of artificial intelligence and computer vision,which has broad application prospect in the scenes of national defense,public security,bank transaction and other important fields.However,complex interference factors such as illumination,posture,expression,occlusion and low resolution in unrestricted scenes will lead to the recognition rate or real-time performance of most automatic face recognition system can not meet the practical application requirements.In this context,this dissertation makes a deeper research for three key technologies in automatic face recognition,i.e.face detection,facial feature point location and face recognition,which aim at improving the recognition rate and real-time performance of automatic face recognition system under unrestricted scenes.The solution with strong robustness and high real-time performance is proposed for each key technology.The main contributions of this study are summerized as followings:1.Under unrestricted scenes,the collected face images are easily suffer from many complex interference factors,which results to low detection rate of many classical face detection algorithms.In this dissertation,a multi-scale aggregate channel features extraction algorithm is proposed to address above issue.Based on three color features in aggregate channel features,the multi-scale Gabor histogram features are introduced to enhance descriptive ability of skin color,texture and scale properties in face images.The comparison experimental results on the FDDB face database show that the proposed feature extraction algorithm has strong robustness to various complex interference factors,and can achieve accurate and efficient face detection under unrestricted scenes.2.The changes of face pose and expression in unrestricted scences will lead to low location precision for most facial feature point location algorithms.In order to solve this problem,a facial feature point location algorithm based on eyes first fitting is proposed in this dissertation.Firstly,the AAM facial feature point model is splitted into eyes feature point sub-model and residual feature point sub-model.Then the face position and deflection angle are used to update and optimization the shape parameters in inverse compositional algorithm.The eye region feature points are first fitted,and then the residual feature points are fitted only after the eye region feature points fitted successful.Because the deflection angle is integrated into inverse compositional algorithm,the localization accuracy can be greatly improved and the time consuming of facial feature point location can be effectively shortened.Finally,the comparison experiments on BOSTON database and camera vedio stream demonstrate the superiority of the proposed algorithm.3.In order to improve the efficiency of the face recognition under unrestricted scenes,a fast face recognition algorithm based local fusion feature and hierarchical incremental tree is constructed.This algorithm achieves fast face recognition from two aspects,i.e.feature extraction and classifier construction.Firstly,a multi block CSLBP is used to extract the local feature of the patch centered at each facial feature point.Then the hierarchical clustering algorithm is used to construct hierarchical incremental tree.This tree employs hierarchical clustering algorithm to construct and retrieve the hierarchical tree structure by the method of coarse-to-fine,and fast classification of face samples is realized by discriminant fusion strategy of local region.Finally,experimental results on several face databases and practical application of TV program recommendation system verify the accuracy and efficiency of the proposed algorithm.4.Far away from camera installing position and insufficient training samples are the important reasons for the low recognition rate under unrestricted scenes.In order to solve the problems,a low resolution face recognition algorithm with single sample per person based on unified adaptive convolutional feature and local collaborative and fusion representation is proposed.Firstly,a unified feature subspace projection model is realized by combining deep convolution neural network and spatial pyramid pooling layer.This model can project different resolution image into unified feature subspace,and output constant dimension convolution feature for different resolution face image.Then the face image is divided into several local blocks and the unified adaptive convolutional feature with intra-class variation is generated by local collaborative and fusion representation model.In the proposed representation model,the multiple-metric fusion discriminant method is used to realize the sample classification,which can make full use of the intrinsic relationship between local blocks and improve the accuracy of classification effectively.Finally,the proposed algorithm is compared with other excellent algorithms in several face databases.The experimental results show that the proposed algorithm has a higher face recognition rate than other algorithms.