Age Estimation Based On Facial Image

Abundant private information is represented by human face,including identification,emotion,age,gender,etc.Biometric recognition technology based on human face has become an increasing hot topic in recent research.Facial age estimation shows great potential value in daily life and commercial application,such as access controll,personalized service,image and video retrieval,imformation collection,etc.A typical approach for facial age estimation mainly consists of two key phases: facial age feature extraction and age estimation model learning.New methods are proposed for both two phases based on our deep research.Moreover,multitask model for solving facial age estimation and gender recognition simultaneously is further studied.The main work of this paper is as following:(1)In the age feature extraction phase,convolutional neural network(CNN)is applied for feature learning because of its powerful ability to extract representative image features.The end-to-end architecture can reduce the preprocess procedures.We designed a CNN model with 8 layers,including 6 convolutional layers and 1 fully connected layer.And then,a coarse-to-fine training strategy was adopted to avoid overfitting.(2)In the age estimation model phase,a ranking encoding method for age labels is adopted.The ranking method utilize the relative order of age labels but the absolute value,so the influence of inaccuracy age labeling is reduced.Ordinal classification problem for age labels was transformed into a series of subproblems for comparing the value of ages.The output of CNN was modified to a series of binary classfiers which can be trained jointly.The experiment results show that the ranking model outperforms classification or regression.(3)A multitask CNN model is adopted to tackle age estimation and gender recognition problem simultaneously.Each tsak can share the low level representations in the same CNN,so that the number of models and computation cost can be reduced.In addition,the training of model can benefit from the relevance information between age and gender.The experiment results show that,multitask model can solve both problems well,and promote the performance for each task compared with single task model.