Cross-age Face Recognition

Recently,great advancement has been shown in face recognition due to the development of the computer vision and deep learning strategy.However,face recognition accuracy is still limited by large intra-class variations such as lighting,pose,expression and age.Face recognition with age variation has not achieved adequate attention compared with other sources of variations.Actually,cross-age face recognition is required in many practical applications.The sharp and texture of a face is subject to remarkable changes caused by the aging process.How to deal with the changes of face appearance across age is the key to tackle the cross-age face recognition.In this thesis,the goal is to construct a robust cross-age face recognition system and the focus is to extract age invariant face feature.The person specific feature is relatively stable across age,which is one of the requirements that face recognition must meet as the identity authentication.It is inevitable the extracted face feature still contains the age variations if using the general face recognition methods.In this paper,the age estimation task is innovatively introduced to guide the feature selection for face recognition by reducing the chance of choosing age sensitive features.Both the traditional subspace learning and the deep learning perspectives are explored.The main work of this paper can be described as follows:1)Age invariant face feature space learningIn this thesis,the identity and age dictionaries are introduced to encode the hand-crafted face features onto two separated subspace.The age estimation task guides the age feature subspace to catch the age sensitive feature,which help separate the age variations from the identity specific features.The face recognition task guides the identity feature subspace to catch the stable identity specific features.Moreover,the label matrix constraints are introduced to ensure the discriminative ability of the learned subspace.Thus,the cross-age face recognition can perform on the identity feature subspace.The experiments,together with feature visualization demonstrate the effectiveness of the proposed method.2)Cross-age face recognition based on deep convolution networkIn this thesis,deep convolution network are introduced to learn the age invariant face feature.The model is based on a pre-trained model which may overcome the shortcomings of the limited number of the cross-age face databases.Feature visualization and neuron energy distribution are used to guide the design of the deep convolution network architectures.The age estimation share the low level texture feature with the face recognition task and guide the face recognition to choose the age insensitive feature.Center loss constraints are introduced to improve the discriminative ability of the identity feature.The performance is evaluated by the experiments.3)Propose the cross-age face recognition framework based on multi-task learningFrom the perspective of traditional subspace learning and deep learning,a multi-task cross-age face recognition framework is proposed.The age estimation is introduced to extract age sensitive features,which guide the face recognition task to focus on age insensitive features.Cosine distance and the nearest neighbor classifier are used for face recognition.Extensive experiments on two well-known face aging datasets: MORPH and FGNET demonstrate the effectiveness of the proposed framework.Compared with the single task,the age estimation task can help improve the generalization ability of the mode and enhance the rate of the cross-age face recognition.