The Research On Key Technologies Of Facial Image Based Age Estimation

The research on human face images is the focus of pattern recognition and machine learning. Face detection, face recognition and expression recognition et al are all the important research areas on face human images. As there are more and more age based applications, the research of facial image based age estimation become a new research focus. Facial image based age estimation problem refers to using the computer tech-nology to model the pattern of facial aging, so that the machine can estimate the human age based on facial image. It has two key technologies. They are the age feature rep-resentation and the age estimation model learning. We introduce two feature learning methods and the soft two level age estimation model. The methods are described below in detail.Based on facial aging pattern, we raise the feature learning method based on weight distribution and the feature fusing method based on cell granularity. Firstly, different face patches age differently, and have different contributions to age feature representation. We raise feature learning method based on weight distribution. Use the sliding window to get the face image patch, learn the age contributions, get the face age weight distribution. We use pixel granularity, cell granularity and cell threshold strategies to combine the weight distribution and the facial feature to get the age fea-ture based on weight distribution. Secondly, face aging not only include the change of shape, but also include the change of texture. We raise the feature fusing method based on cell granularity. Divide the face image into image cells, and then connect many different facial features on the cell granularity, at last get the fused feature. The experiments prove that the feature learning method based on weight distribution and feature fusing method based on cell granularity both can represent age information more effectively.We raise the soft two level age estimation model for the age estimation method research. The facial aging process can be roughly divided into two phrases, so we use the coarse to fine strategy. At first classify the age, and then construct the age estimation models for different age phrases. But the facial aging is a slow process. It exists the problem of age boundary fuzzy. So in the paper we use the strategy of setting the overlap region at the age border. Expand the application scope of the age estimation models, so that we can fix the error of the age group classification. The experiments prove that the soft two level age estimation model can achieve a better age estimation result.