Facial Attribute Estimation And Age Synthesis

Generally,face analysis includes studying of identity feature,age feature,expression fea-ture,facial attribute feature,gender attribute,race attribute,etc.Recent identity feature methods lack the fine-grained identity-related description,e.g.,shape of facial features,facial texture,etc.Compare with the traditional face recognition,facial attribute analysis is the more detailed bi-ological feature analysis,which has the extensive potential applications,e.g.,national defense security,anthropometry,medical treatment,social entertainment,etc.Existing facial attribute methods usually focus on the obvious facial attributes,e.g.,hair color,shape of facial features,skin color,gender,etc.However,there are many challenges for the fine-grained facial attribute estimation,e.g.,single/double eyelids,shape of lips,density of eyebrows,etc.Besides,empir-ically,facial aging is a main factor to change the facial attributes.In short,one person's facial attributes are changing as the age growth,e.g.,facial shape and skin texture.Therefore,given a face photo,how to aesthetically re-render an aging face at any future age(called age synthesis)is one of main topics in facial attribute analysis.Age synthesis not only can be applied into the social entertainment,but also can collaboratively solve the cross-age face verification and age estimation.To sum up,this thesis studies both facial attribute estimation and age synthesis,and then the main research achievements are summarized as follows:1 This thesis studies the facial attribute analysis problem,especially the fine-grained faical attribute estimation.First,fined-grained facial attributes are defined and then a facial attribute dataset is collected.Second,a new deep convolutional neural network is proposed to estimate the fine-grained facial attributes.And then,the proposed deep network is applied into a face reading system.2 A kinship-guided age synthesis method is proposed to guide the age direction in age syn-thesis,and then render the authentic aging face.In particular,the kinship consistency between the child and parent can be taken as a prior to direct the aging result.The proposed method can render the aging face with identity preservation and natural authenticity at any future age for an input face.3 An age synthesis based on coupled aging dictionary learning(CDL-AS)is proposed by utilizing the aging dictionaries in different age groups to reconstruct the aging face.First,a face in the human aging process can divided into the aging-variant pattern and the aging-invariant pattern.Second,human age process can be into several age groups,where each age group is designed an aging dictionary to represent its aging characteristics.These aging dictionaries are trained on the collected intra-person face pairs.Given an input face,its aging face can be reconstructed by the liner combination of the different dictionary bases with the corresponding weights.Finally,experimental results show that the proposed method achieves the best aging results.4 Based on CDL-AS,the coupled aging dictionary learning model is extended to a more efficient bilevel aging dictionary learning(BDL)model.Specifically,BDL model takes the younger faces in the face pairs as the inputs of the auxiliary objective function,and takes the older faces in the face pairs as the inputs of the main objective function.And then,an effective optimization procedure is presented to solve the optimization problem of this objective function.In experiments,the proposed age synthesis based on bilevel aging dictionary learning(BDL-AS)has not only solved the personalized age synthesis,but also achieves the best aging results and the best performance of cross-age face verification.