Facial Age Estimation By Improved Label Distribution Learning Algorithms

In recent years,the computer science has been applied to the human biology.Facial images contain a lot of important information,such as age,facial expression,gender and identity.As the inherent biological characteristics,the age can be widely used in a variety of scenarios.Although traditional classification(or regression)algorithms can be used to estimate the facial age,they are unable to make full use of the correlation and the uncertainty among the age labels.Label ambiguity will improve the performance if it can be reasonably exploited.In order to utilize the label correlation,Label Distribution Learning(LDL)algorithm can achieve the state-of-the-art in facial age estimation applications.Every instance is assigned a discrete label distribution according to its ground-truth.The aim of the LDL is to learn the age label distribution of each instance with the Kullback-Leibler as the loss function.The model of traditional LDL is the maximum entropy model.Therefore,this thesis improves the performance of classical LDL from following two aspects:On the one hand,traditional LDL algorithm has limited description which is not applicable to all LDL problems.In order to improve the performance with this specific model,this thesis needs to improve the loss function.Alternatively,replacing this exponential part with a general function to approximate this parametric model of maximum entropy model can avoid the potential influence of the specific model.This will make the loss function more complex about the complicate optimization strategies and may lead to over-fitting.This general function is variable such as decision tree or the linear combination.So this thesis can use this general function to represent any specific function which constitutes an LDL model family.This method is called Label Distribution Gradient Boosting(LDGB)algorithm which is a combination of the boosting method and the label distribution learning.This LDGB can fit individual weighted regression function(base learner)to realize the optimization step by step.On the other hand,traditional LDL algorithm uses the hand-crafted features such as BIF or HOG.Deep Learning has its natural advantages in feature learning.Therefore,we can combine the Deep Convolutional Neural Network(CNN)with the LDL to improve the limitation of the traditional hand-crafted feature learning.This End to End method is called DLDL which is not only achieving robuster performance than existing classification(or regression)methods,but also effectively relaxing the requirement for large amount of training images.There are six chapters in this thesis.Definition,state-of-the-art and open research problems of facial age estimation as well as label distribution learning are introduced in Chapter 1.From Chapter 2 to Chapter 5,traditional facial age estimation algorithms,traditional label distribution learning algorithms and proposed algorithms(LDGB and DLDL)are introduced with their detailed experiment results.Finally,we conclude this thesis in Chapter 6.