No-Reference Face Image Quality Assessment Based On CNN

Since modern society is an information society,all areas are involved in fast and effective automatic authentication.Currently biometric identification technology is considered as the most reliable way for authentication.Face is the most ideal biological characteristic for the identification.It has extensive application in some areas such as the national security,the judiciary,customs and insurance.However,In most cases,face image collection is non-contact and uncontrollable,and it is easily affected by light,angle,cover and expression,which could have some negative effect on identification.If the quality of face images can be controlled,the system detection and recognition performance can effectively be improved.So it is very important and necessary for us to research face image quality assessment.In this thesis,we focus on the research of no-reference face image quality assessment based on CNN.Firstly,this thesis describes the background and significance of face image quality evaluation,summarizes the face image quality assessment method,and the main types of face distortion.It emphatically analyzes the current method of no-reference face image quality assessment.Secondly,the convolutional neural network is summarized.The thesis analyzes the theoretical basis of convolution neural network,introduces the algorithm of convolutional neural networks for no-reference image quality assessment and the feasibility analysis of no-reference image quality assessment.Finally,CNN is build on Caffe platform.After optimized,it is used to extract deep features of images.And I design and train multi-class support vector machine classifier(SVMs).The assessment scores can be obtained from the outputs of SVMs.The test images assessment score distribution and the results that I compare no-reference face image quality assessment based on CNN with Learning To Rank show that the assessment method proposed in this thesis concurs with our instinctive perceptual system.