Comparison of Three Different CNN Architectures for Age Classification

Abstract

As one of the powerful tools of machine learning, Convolutional Neural Network (CNN) architectures are used to solve complex problems like image recognition, video analysis and natural language processing. In this paper, three different CNN architectures for age classification using face images are compared. The Morph dataset containing over 55k images is used in experiments and success of a 6-layer CNN and 2 variants of ResNet with different depths are compared. The images in the dataset are divided into 6 different age classes. While 80% of the images are used in training of the networks, the rest of the 20% is used for testing. The performance of the networks are compared according to two different criteria namely, the ability to make the estimation pointing the exact age classes of test images and the ability to make the estimation pointing the exact age classes or at most neighboring classes of the images. According to the performance results obtained, with 6-layer network, it is possible to estimate the exact or neighboring classes of the images with less than 5% error. It is shown that for a 6 class age classification problem 6-layer network is more successful than the deeper ResNet counterparts since 6-layer network is less susceptible to overfitting for this problem.