Design And Research On Face Recognition System Based On Deep Convolutional Neural Network

The rapid development of computer hardware and network have made solid foundation for computer vision and machine learning.Biometrics feature recognition as one of the hot research direction of computer vision have absorbed researchers' extensive attentions that from domestic and overseas,especially after deep learning have been proposed and used successfully.Face recognition technology as one of the important research direction has developed quickly for its high commercial value and wide application prospect.Although face recognition technology has been developed for decades,for its complexity and many technological difficulties have not solved,it's still has not meet the request of business application.In this thesis,we go alone with the pipeline of face recognition,research the key technology of face recognition in each stage and do some chooses and optimizations.A complete face recognition system is including image capture,image preprocessing,face localization detection,face landmarks detection and alignment,face principle feature extraction and recognition.To the image quality reduction for the influence of environment,equipment and etc,we deal with the problems with image normalization,denoising and image augment,and make it convenient for the face detection and face feature extraction later.In order to avoid the full area research of face detector,we combined with the face skin color extraction and segmentation in YCbCr color space algorithm and face detection algorithm based on Haar feature together and improve the speed of face detection.Transplanting the traditional face recognition algorithm under the embedding platform NVIDIA TK1 development board and many experiments and analysis are carried out.Aimed at face big data and recognition in complex environment,we have designed face recognition system that based on deep learning that can suite the face recognition under a large amount of face data and more complex environment.Aimed at the difficulty for normal researchers to acquire big face data,we have proposed a multi-scale samples segmentation data augment algorithm that can augment the dataset largely.Furtherly,for the accuracy of traditional network model in cross-age face recognition environment decreased largely,we have proposed face age feature progress model,and transfer the former deep convolutional neural model in face age dataset for further learning,Enhancing the robust of face recognition model.At last,we have integrated the hardware and software system together based on the optimization of algorithm in every stage.We also have made web development at front and server building at end and made our face recognition system can be visited in the internet.Besides,the server can collect the face big data and will provide powerful support for big data mining through sufficient data analysis.