Researches On Face Attribute Estimation Based On Deep Convolution Neural Network

With the development of social informatization,the collection and utilization of facial attribute' information is increasingly gaining popularity in people's daily life.People's face can reflect a variety of attributes,including identity,gender,age,physical characteristics,dressing and so on.Among them,identity is the most important part of facial attributes,and there have been many researches on face recognition and verification.Traditional face recognition is the identification of certain face,which has been almost fully developed.This paper mainly focuses on three specific facial attribute estimation tasks,including age estimation,kinship verification and facial multi-attribute estimation.This paper also proposes specific solutions to their existing problems at the present.The main work and contributions are as follows:1,during the research of age estimation,we propose refining label distribution based on deep learning to solve the problems from insufficient training data and from the different rate of change of facial appearance at different ages.Assuming that the label distribution obeys discrete Gaussian distribution,we use the deep convolution neural network(DCNN)and back propagation(BP)algorithm to learn the appropriate distribution parameter for each age,so that we can better estimate the age of the face.2,during the research of kinship verification,it is difficult to distinguish whether there is kinship between different relatives' face image pairs.In order to solve this problem,margin adaptive triplet loss algorithm is proposed.The margin of every kind of kinship is constantly updated during training,to suit for the current network state.This allows more data to be used for training,so as to achieve the goal of pulling relatives closer and pushing people with no kinship farther away.In addition,two strategies for updating margin are proposed,and the DCNN framework is used in experiments.3,during the research of facial multi-attribute estimation,we noticed that the attribute should be detected from their natural location instead of the entire space domain,so we put forward the position-squeeze and excitation(PSE)block.The PSE block compresses the spatial information of the corresponding position of different channels into a position descriptor,and then uses a sigmoid function to capture network dependence on different locations of feature space.PSE block can help the network learn more useful features from the natural location of the attribution.Finally,the end-to-end learning is carried out by using the DCNN to build deep multi-task learning framework.