Research On Face Recognition Algorithm Based On Deep Learning

Face recognition is one of the most active research fields in computer vision and pattern recognition.It has many practical applications and is widely used in military,finance,public security,and daily life.Accurate recognition of a person has become extremely important.In recent years,many excellent algorithms,structures and databases have been proposed to solve face recognition.But for Asian faces in the case of a small sample data set,more accurate face recognition is still a major challenge.For a complete face recognition system,there are three steps of image preprocessing,face detection and face recognition.This article makes improvements from these three aspects,expands the data set,improves the existing algorithm,proposes a new model,and obtains good test results:1)Aiming at the problem of insufficient network model training due to small sample data sets,data enhancement and network generation techniques are used to expand the sample data set.This article starts from the perspective of expanding the richness of existing face photo samples,and realizes the expansion of data samples,so that the face recognition model can be trained even when the samples are insufficient.The expansion of data samples not only uses data enhancement techniques,such as mirroring,cropping,and noise addition.In addition,the generation of confrontation networks is used to generate rich sample data through the confrontation generation network model,adding more sample data containing face information,in the hope that the network model can learn more features from the sample images.2)Aiming at the problem of inaccurate detection accuracy of overlapped and semi-occluded faces,the multi-task convolutional network algorithm model is improved to improve the accurate positioning of the face position.First,analyze the classic multi-task convolutional network model,and then add a multi-layer perception module to it,so that the network can obtain richer detailed information and edge information,and at the same time improve its loss function,and balance the class spacing and intra-class spacing.Finally,the non-maximum suppression technology is used to merge and omit multiple overlapping regions,which improves the accuracy and completeness of detecting face detection regions.Use the improved face detection algorithm to detect and align faces to obtain a picture size suitable for the face recognition network and prepare for subsequent recognition.3)Aiming at the problem that the face recognition model cannot obtain a good training effect when the data is insufficient,a siamese network model with feature extraction from the residual network is designed to improve the accuracy of face recognition.First,analyze and study the residual network structure,change its network model;make the improved structure form a siamese network,and design a brand new convolutional neural network.Use the richly generated and expanded face sample set for learning and training,and train a better network model.The model is analyzed by multiple sets of experiments,and compared with the existing algorithms,a better performance improvement is obtained,and a better recognition accuracy rate for Asian face data is achieved.