Face Attribute Learning Based On Multi-task Learning And Metric Learning

Face attribute learning aims to predict some attributes for a given face image.It is very chat-lenging since it has the following two key issues unsolved:First,face attribute learning belongsto multi-task learning,which requires learning relationships among all the tasks.Each attribute can be seen as a classification task.Then,face attribute learning is a multi-task learning problem that consists of several attribute learning problems.It is important for a face attribute learning model to learn relationships among all the tasks(attributes),so as to improve the overall performance.Second,we find that the attributes of the same person are highly similar.It indicates that face identity can help construct local geometric structure in the feature space,which is complementary for the global attribute classification learning.The model can learn more sophisticated attribute relationships and more discriminative face attribute features,once it models the local geometric structure.Based on the analysis above,the contributions of this paper are listed bellow:1.We propose a Partially-Shared Structure and a novel Partially-Shared Multi-task Convolu-tional Neural Network(PS-MCNN)based on the structure.In particular,we construct 4 Task-Specific Networks(TSNet)and 1 Shared Network(SNet)for the face attribute learn-ing problem.TSNets are connected to SNet via Partially-Shared Structures at each layer to construct PS-MCNN.PS-MCNN encourages information exchange among different tasks via an SNet,so that the relationships between different tasks are efficiently modeled.2.We propose the concept "inner-person consistency" of face attributes.And we propose a Lo-cal Constraint Loss(LCLoss)based on the attribute similarity for the same person.LCLoss defines samples with the same identity as the local geometric neighbours.By drawing the local geometric neighbours close to each other,PS-MCNN learns local geometric structure in the feature space,which enables PS-MCNN to learn more sophisticated attribute rela-tionships and more discriminative face attribute features.3.We conduct extensive experiments on face attribute learning datasets.The results show the superior of our method.