Research On Face Recognition Based On Deep Learning Under Non-limiting Conditions

In recent years,the accuracy of face recognition has been improved by leaps and bounds with the introduction of deep convolution neural networks(DCNN).Various related applications,such as face recognition attendance machine,candidates authentication,face recognition payment,face classification inquiries,have begun to gradually put into use,the effect is significant.However,there are many problems in the face recognition technology of non-qualified condition in realistic scene,for example pose,illumination,shelter and so on.The recognition accuracy is also rapidly declining with the increase of the data scale and the difficulty of recognition.The current solution is to learn as many scenes as possible by increasing the size of the training database in complex scenarios.However,most large databases(one million)are held by private companies and not public.Due to the marked information is too small,the accuracy of public large scale databases can not be better guaranteed,it will impact on DCNN Training.In order to ensure the accuracy of the training database and to solve some of the pose problems,this thesis has made the following contributions:This thesis first get to know some theoretical knowledge of deep learning,summed up the current mainstream deep learning open source framework and face recognition of open source projects as well as domestic current mainstream face recognition related companies and solutions.The performance of the current mainstream face detection algorithm is analyzed by comparing the error,the detection effect and the stability of the large-scale data.According to the actual situation of this thesis,the evaluation mechanism of detection error is put forward to reduce the error.Finally we chose the synthesis effective algorithm for large-scale data processing.This thesis proposes a normalized image algorithm based on key point mapping.Now large-scale database accuracy can not be guaranteed,due to noise and other issues,a data cleaning method based on multi-angle evaluation is proposed.Calculate the similarity in each image and other images in the same category,count the number of images that are not similar to the image,then the image is cleaned up by a certain amount.The validity of the method of clearing data is verified by various experiments.Experiments show that the accuracy of the database training model in the LFW dataset has been improved,the accuracy rate of 99.17%has been obtained with a smaller training set,the Youtube dataset achieved an accuracy of 93.54%.In order to solve the problem of non-limiting pose face recognition,generate a positive image using image correction based on a 3D face model.A new method is proposed to generate a new feature vector by combining the feature of the front composite image extraction with the original image feature.The synthetic front image can provide features that the original image does not have.There is also information loss,but the characteristics of the original image has a high reference value.Therefore the new features have more comprehensive characteristics.Experiments show that the new feature vector can effectively improve the face recognition rate,we corrected 15 pairs of 50 pairs LFW error matching pairs,SWJTU-MF DB also achieved remarkable results.