Research On Age Estimation Based On Multi-task Learning

Human age estimation is an attempt to give the computers the ability to estimate age from a face image. Recently, it has received more and more attentions due to its widely application in face recognition, face aging simulation, security surveillance, man-machine conversation and so on, and it has been one of the most challenging topics in pattern recognition and machine vision. A successful age estimation method typically consists of two key modules:face image feature represent and age estimation mode learning. Our work is based on these two modules.(1) For face image feature representation, we extract the mid-level age feature to characterize the age-related variance. First Dense SIFT descriptors are used to capture the texture of face images. Then the sparse representation is used to coding the original features to a learned dictionary. Through this learned process, we can get the higher and more discriminate Mid-level age features of a face image.(2) For age estimation model learning, we combined Ridge Regression, support vector regression, linear sparse regression and Multi-task learning to propose three effective age estimation methods. First, we use linear sparse regression to select feature, and Ridge Regression and support vector regression are used to learn the age estimation model. Then in order to distinguish the different of individuals, we introduce the multi-task learning into age estimation problem, and we first use this method as a feature selection method, and Ridge Regression is used to learn the age estimation model. Based on those works, we further use linear SVM to classify the face image into several age groups, and then use multi-task learning to learn the age estimation model of these age groups. Multi-task learning not only considers the different of individual, but also shares the common feature of each task, which is more close to the fact of age estimation problem. Multi-task learning provides a new idea for age estimation problem.The paper performs a lot of experiences on the public available FG-NET and MORPH database to verify the efficient of the proposed third methods. The results of the experiments show that our methods can obtain a better or comparable results to the state-of-the-art age estimation methods.