Research On Age Estimation Method Of Facial Images Under Local Blocking Condition

In recent years,the rapid development of computers and other science and technology has led to many areas of progress,the development of artificial intelligence,pattern recognition,image processing and other fields have brought many changes for people's lives,face images can reflect the human complexion,emotions,age and other important information.Therefore,the study of human face image is more and more popular with the majority of scholars.The age estimation of face image also shows prominent application prospect,which makes the question have more and more research value.The main purpose of this thesis is to study the method of age estimation of face images.The main work is to study the method of face image age estimation under non-occlusion and partial occlusion.Specific work as follows:Firstly,we choose texture features to extract in this thesis.LBP features and Gabor feature are common texture feature extraction methods.Because the face image itself is not highly differentiated by age,single feature can not distinguish between different ages.This thesis proposes a method of feature extraction based on multi-directional Gabor feature and uniform LBP feature for age estimation,The experimental results of single feature extraction and fusion feature extraction are compared and verify that the feature fusion is better than the single feature extraction.Secondly,in the training part of the age estimation model,this thesis first studies the SVM method.Based on the single feature and the fusion feature,this thesis designs the single-layer model of SVM for age estimation.For the small samples and the uneven sample distribution,in this thesis,a double-layer model based on fusion feature SVM is proposed to estimate the age.It is also verified experimentally that the age estimation of the fused bivariate model relative to the single-layer model can reduce the error of the age estimation to a certain extent.Thirdly,we apply deep belief networks to age estimation problem,and choose the parameters of deep belief networks by comparing several groups of experiments.Furthermore,we evaluate the age estimation effect of deep belief networks based on different feature extraction methods.In this thesis,we design a set of experiments to compare the age estimation based on SVM model and depth confidence network model under different training set numbers.It is verified that when the training set number is sufficient,the deep confidence network can effectively reduce the error of age estimation and improve the estimation accuracy.Fourthly,based on the above research,this thesis presents a method to extract the training multi-model based on local features,in view of the fact that there is a partial occlusion of face images in real environment,due to the reduction of the distinction of age features caused by the lack of local features.The fusion of LBP feature and gray level co-occurrence matrix feature proposed for the feature extraction of local occlusion face images.The experiment verifies that this method can reduce the error of the age estimation of partially occluded face images and improve the accuracy of the estimation.