The Research Of The Key Technology Of Face Recognition Based On Deep Learning

Face recognition,as one of the most representative biometrics,plays an increasingly important role in various fields of society.The improvement of face recognition performance has a greater impact on social development and satisfying people's needs.Positive effect.It is the key to improve the performance of face recognition by extracting the face feature representation with better discrimination and better generalization ability and the faster design and higher accuracy.Based on the theory of depth learning,this paper studies and improves the two steps of face feature extraction and classification recognition in face recognition.The main contents are as follows:(1)The feature extraction method in face recognition is studied.Through the comparison and analysis of several commonly used facial feature extraction methods,the convolution neural network in depth learning is selected as the face feature extraction method,which is different from the traditional method based on artificial features.It uses a large number of face image data with labels to train the convolution neural network model for feature extraction.The model is still well under unrestricted conditions,and the extracted facial features also have strong discriminative and generalization ability.(2)Improved network structure for convolution neural networks for face feature extraction.With a convolution layer with a smaller convolution kernel instead of a convolution layer with a larger convolution kernel,and two convolutions with asymmetric convolution cores instead of the convolution layers of the small convolution kernel,This not only reduces the amount of calculation and training time of the model,the extracted features also have a stronger expression.In addition,the traditional cornvolutional neural network output layer is only connected to the top of the feature extraction layer,so that the output layer only uses the representative of the overall,invariant,very little detail of the high-level abstract features.In this paper,more feature extraction layers are connected to the output layer,so that the local low-level features of the underlying layer and the middle layer can be fully utilized.Finally,the validity of the improved convolution neural network is verified by experiments.(3)A face verification algorithm based on joint Bayesian is proposed.The face verification process of general face recognition uses only simple distance measurement or learning type measurement method for face feature matching.In this paper,two methods are combined to perform face feature matching.First use the cosine distance to make rough decisions,which can reduce the time of feature matching.And then use the joint Bayesian fine decision,this can guarantee the accuracy of feature matching.Finally,the effectiveness of the algorithm is verified by experiments.