Age Estimation with Multilevel Residual Networks in Unconstrained Conditions
Zhang Ke, Gao Ce, Guo Liru, Yuan Jinsha, Zhao Zhenbing, and Li Baogang
(Department of Electronic and Communication Engineering, North China Electric Power University, Baoding 071000)
Age is the inherent attribute of human, and plays a fundamental role in human social interaction. Therefore, age estimation is one of the important problems in the field of artificial intelligence. In order to solve the problems of age estimation in unconstrained conditions, a new CNN based method for age estimation in unconstrained conditions leveraging multilevel residual networks is proposed. Firstly, the multilevel residual neural network is constructed for the high-resolution image datasets. Secondly, the multilevel residual network is pre-trained by ImageNet dataset to obtain the basic features of the image. Thirdly, the multilevel residual network model is fine-tuned by combining the stochastic depth algorithm on the unconstrained face dataset. Finally, the comparative experiments for age estimation are performed on the unconstrained Adience dataset. Excellent experimental results show that this method can significantly improve the accuracy of age estimation in unconstrained conditions, and effectively alleviate the over-fitting problem caused by small-scale datasets while improving the learning ability of network.
multilevel residual networks; age estimation; unconstrained conditions; stochastic depth algorithm; ImageNet and Adience datasets