Research On Face Feature Extraction And Location Algorithm Based On Transfer Learning

Face feature extraction and localization is one of the classic problems in computer vision,and it is the basis for solving many face related problems.Accurate face feature point location results are very important for many visual tasks,such as face recognition,3D face rebuild,face expression analysis and face pose estimation.In recent years,great progress has been made in the research of face feature point localization.The accuracy of constraining face feature points has been able to meet the practical application requirements,but the non-constrained face features extraction(that is,the face image has posture,expression,occlusion,illumination,age and makeup changes)still has inaccurate positioning,low precision and other issues.This thesis mainly researches the face feature extraction and localization algorithm based on transfer learning.Firstly a face feature extraction and localization algorithm based on deep convolutional neural network is designed to realize sparse face feature point localization.And then we transfer the network to the method of the local binary feature(LBF)for our dense face extraction and localization.This thesis has completed the following aspects:(1)A face feature extraction and localization algorithm based on deep convolutional neural network is proposed to accomplish sparse face feature extraction and localization.A deep convolutional neural network is used to construct a three-layer cascaded network structure,which makes the ReLU activation function have faster convergence under the same network structure and applies it to sparse facial feature point location.And We compared the L.Liang method,the M.Valstar method and the Luxand method,and selected the average error rate and the failure rate as the evaluation index.We achieved an ideal result with an average error rate and failure rate of less than 2%.(2)A transfer learning algorithm is designed to complete the feature extraction and localization of dense faces.Transfer learning can use pre-trained knowledge to accelerate the training process of deep convolutional neural networks.The deep convolutional neural network is combined with transfer learning,and the face feature extraction and localization algorithm based on deep convolutional neural network is designed to transfer to the local binary feature method.We propose two face feature point extraction and localization algorithms based on transfer learning,which are direct transfer methods and cascade fine tuning transfer methods.Using transfer learning,this method can transfer the CNN to extract the local features of the face,and then connect each feature point extracted separately to perform a global constrained regression,and finally accurately predict the feature points of the dense face.(3)Through experiments,the method proposed by the thesis compared with the mainstream facial feature point localization methods(such as: RCPR,ESR,SDM,CFAN,DeFA,LBF,CFSS,3DDFA,LAB,Wing-Loss).In comparison with individual algorithms(such as RCPR,ESR,SDM,CFAN,3DDFA),this method is in a leading position in localization accuracy.In the feature extraction process,the features extracted from the LBF are reduced from the original 100K+ dimension to the 4352 dimension.Compared with the current state of the art algorithm Wing-Loss,the EFFECTIVE is more than doubled in the case of less than half of iterations.It is proved that the transfer learning algorithm proposed in this thesis has achieved good application results in face feature extraction and localization.