A Novel Image-Based Human Age Estimation Algorithm using Local Texture Curvelet Transform Features and Improved Active Shape Model

Abstract

Automatic age estimation of human facial images with its broad applications in human-computer interaction (HCI) still has a certain complexity that makes it a challenging issue. This paper introduces a novel and efficient method to estimate the age utilizing the combination of texture and shape features which describe human’s age. To extract the proposed features related to image texture, the new generation multi-scale Curvelet transform is applied to the local facial images so the texture features which are the most stable components of the image features are extracted. After that, by adaptively thresholding on Curvelet transform’s entropy and computing each curvelet’s sub-band entropy, the most relevant features of local texture images are chosen. The image shape features are extracted using a new active shape model which has three improvements compared with other active shape models. Then they are merged with the local texture features to create the age feature vector. The genetic algorithm (GA) with a special fitness function is utilized to choose the best aging features so the most effective features from texture and shape is chosen. Finally the neural network classifier is used to classify the age characteristics in different ages. Experimental results shows mean absolute error (MAE) of 3.47 years for the proposed algorithm on FG-NET aging database images which is the lowest reported error expressed so far that proves the accuracy of the proposed algorithm.