A Study On Automatic Facial Age Estimation Based On Multi-Label Learning

Age estimation is a novel and significant research direction in the area of biometricidentification, and there is plenty of age concerned information contained in facialimages. However, because of the difficulties in getting enough training samples as wellas the special characteristics of aging process, the estimation results based on facialimages still need to be improved. As a result, motivated by the slow and continuousaging process, multi-label based model, which is fitter for aging characteristics, isintroduced to represent the facial images in age estimation.In most of the previous work, age estimation is generally formulated as asingle-label based problem. However, since aging is a gradual process and people arealways in transition period between ages, labeling a facial example with an exact age is adifficult problem. Meanwhile, sufficient training data is lack for many ages. In this paper,to improve the accuracy of age estimation, a novel approach by applying Multi-LabelLearning to the age features is proposed. In that approach, each facial image is treated asan example associated with the origin label as well as its neighboring ages, which makesthe data more reliable and sufficient. The motivation comes from the observation that,with age changes slowly and smoothly, people would look quite like themselves beforeand after several years. And in this paper, in order to capture comprehensive informationin facial images and further better the previous approach, multi-instance model is alsointroduced in the representation of facial images.Experiments show that the proposed multi-label approach outperforms thetraditional age estimation approaches. Though the multi-instance improvement doesn’tachieve better results, analysis as well as the future work is shown in this paper.