Facial Age Estimation by Conditional Probability Neural Network

Abstract

A new label distribution learning algorithm for facial age estimation, namely the Conditional Probability Neural Network (CPNN), is proposed in this paper. CPNN is based on a three-layer neural network which takes both the target variable (e.g., the age) and the conditional feature vector (e.g., the facial features) as its inputs, and the output is the conditional probability of the target variable given the feature vector. As a label distribution learning algorithm, CPNN can effectively utilize the neighboring ages while learning the real age. Compared with the existing label distribution learning algorithm IIS-LLD, it does not presume the underlying model as the maximum entropy model, but learns it from the training data. Thus CPNN is able to match the real problem better. Experimental results on the FG-NET database show that CPNN performs remarkably better than all the other eight compared methods.