Research On Multi-pose Face Recognition Based On Depth Learning

In recent years,the rapid development of computer technology,especially the greatly improved performance of computer hardware,has provided the necessary support for the development of deep learning.At present,automatic face recognition technology has been widely used in many fields,such as public security,online dating,finance,education and so on.However,the face images collected in the natural environment often have problems such as illumination,expression,occlusion,pose and so on.These factors will lead to a sharp decline in the accuracy of face recognition,resulting in a lack of application.In most application scenarios,face recognition technology can only be used under strictly limited conditions.The pose problem has always been a technical problem in the face recognition research,a workable idea is to learn the characteristics that do not change with the change of pose,Then the pose's image is corrected,and then the feature is extracted and verified in a unified attitude space.This paper focuses on the multi pose face recognition,and the following work is done:1.In the multi-pose face correction method based on deep learning proposed by Zhenyao Zhu et al,pose correction network model using only 3 Convolution layers facial feature extraction,feature extraction leads to the problem of insufficient,puts forward the improved pose correction model Wild-B.Based on the deep learning open source framework(Caffe),the Wild-B is constructed and implemented,and then the Multi-PIE face database is used to train the model.After that,a comparative experiment is done to verify that the improved facial pose correction model has better facial image features.The face image based on Wild-B correction is more favorable for improving the accuracy of face recognition.2.Considering that training a face recognition model directly with corrected images may bring new redundancy or cause information loss,and the original face image always contains facial feature information the most original,the most comprehensive,has great reference value for face recognition,so the proposed weighted fusion method for face image feature after the original image and Wild-B correction to extract,produce the combination of new features,and then used for training.3.We build the combination feature of deep learning network training,and test and analyze it on the LFW face dataset,and determine the weight of the combination feature.Then we further verify it on the YTF video face dataset.The accuracy of recognition model trained by Wild-B corrected face image is 98.09% on LFW,and 91.05% on YTF,respectively,which is 0.32% and 0.46% higher than that of initial model.The accuracy rate on LFW and YTF after fusion is 0.74% and 1.41% higher than that of the initial model.The experimental results show that the Wild-B face image after correction to a certain extent to solve the recognition pose brings the problem of inaccurate,and the better characteristics form obtained by feature fusion,which further enhance the recognition accuracy,show that the method has a certain reference value for the study of multi-pose face recognition.