Research On Face Recognition Of Low-quality Image Based On Deeplearning

The development of deep learning has greatly promoted the progress of face recognition technology.But when the image quality is degraded because of illumination,distance,weather or other factors,the facial feature is vague and the accuracy of face recognition is greatly reduced.So face recognition is still faced with many technical challenges,and the research of low image quality face recognition based on deep learning is of great significance.In order to solve the problem of face recognition algorithm performance degradation and recognition accuracy degradation due to image quality,this thesis studies the problem of low image quality face recognition based on deep learning.The main research contents are divided into three parts:First of all,this thesis carefully studies the theory of deep learning,common network models and training skills.The field of deep learning mainly includes the design of the network structure,the backpropagation algorithm,the loss function and the training of the network model.Convolution neural network is a common network structure of image recognition,which mainly includes convolution layer,pool layer,full connection layer,activation function and loss function.Gradient descent method is a commonly used method of back-propagation algorithm.This thesis focuses on the gradient descent method and its improved algorithms.The training skill of network models is an essential part of improving the performance of network models.This paper focuses on the research of parameter initialization and batch normalization.Secondly,for improving image quality and improving face recognition accuracy,this thesis proposes an image enhancement algorithm: G-Log method.On the basis of analyzing the influence of distance,illumination on the quality of face image,and based on the research of image degradation model,the G-Log algorithm is proposed.This algorithm uses a nonlinear transformation to change the pixel distribution of images,so as to enhance image contrast and restore image details,and finally improve the accuracy of face recognition to a large extent.Finally,on the basis of a large number of experiments,this thesis designs a face recognition system for small data sample and low image quality face recognition.In order to solve the problem of small amount of data,this paper uses the data enhancement module to expand the existing face data.After analyzing various convolution neural network structures in detail,and combines the experimental results and the characteristics of this topic,this paper uses VGGFace network model to extract face features.The classification module has analyzed a variety of classification algorithms in detail.The experiment shows that the SVM classification results are more stable and the classification accuracy is higher.In summary,based on the G-Log image enhancement algorithm proposed by this paper,this thesis designs a face recognition system for the low quality and small sample data set: G-Log algorithm for image enhancement,data enhancement module for data expansion,VGGFace network module for face feature extraction,SVM classification.And finally face recognition accuracy has been further enhanced.