Research On Deep Learning Method And Applications For Face Recognition Technology

Along with the rapid development of bioinformatics in recent years,biometrics has been rapidly integrated into all aspects of human social activities.Commonly used biological information,including palm,fingerprint,iris,face,gait and so on.Among them,the face information with natural,easy to detect and other unique features,making it more suitable for covert detection,public security and other fields.With the development of hardware and computing capacity,data size,learning algorithm and feature extraction model have gradually become the important factors of face recognition task,and even the whole pattern recognition field.The current the deep learning technology for face-recognition,still have some problems in the fields below: 1)the data size is too small and the diversity is not enough;2)the learning algorithm is not optimized for training data;3)the depth feature extraction is sensitive to the preprocessing of face images.In view of the above problems,this paper starts from three points for improving the face recognition algorithm: the establishment of large-scale data sets,the more effective learning algorithm for specific data sets,and the more stable features for verification.The main work of this paper is as following:(1)Aiming at the contradiction between the diversity and richness of existing open datasets,the paper presents a large-scale face dataset and its complex sample construction method.This paper presents a large face dataset PDLFace,which combines three public datasets and three commercial datasets.The dataset has a strong diversity in race,skin color,lighting,accessories,attitude and so on.Furthermore,an IUSM is proposed to improve the face verification performance in the biased feature space without affecting the generalization performance of the model in the full feature space.(2)Aiming at the problem that the training data sets are not specially optimized for the training data set,and the model is susceptible to the dirty samples in the training data,a robust face recognition learning algorithm based on depth learning framework is proposed,which is called Soft-Triplets.After analyzing the limitations of traditional triplets selection strategy,a novel selection strategy based on triples loss is proposed.By introducing two adjustable variables,the strategy can adjust the fitting ability of the model to difficult samples,and adjust the generalization ability of the model to the conventional samples.By adjusting the hyperparameters in the model,this method can effectively reduce the influence of the dirty samples in the data set on the generalization ability of the model.(3)Aiming at the problem of deep facial feature's sensitivity to the preprocessing,a robust method for face detection is proposed.In this paper,we propose a method based on supervised training for linear weighting of multiple sliding window features.This method makes use of the binomial feature of face verification task and supervises training optimization based on Fisher discriminant criterion of mass data to make the similarity distribution of two categories more separate.In this paper,an iterative optimization method is proposed to improve the training efficiency and effectiveness.The method has the advantages of strong stability,small amount of fusion computation and simple realization in practical application.