Age estimation using deep learning

Abstract

The progress of computer vision has seen great strides in these recent years, thanks to the numerous research that was conducted exploring different applications of neural networks on images. One of the applications is age estimation. Proper age estimation can be difficult to achieve since several factors can affect the result, such as face orientations, different illuminations, occluded faces, and even black and white images that can have a signi ficant effect. However, recent advances in deep learning have achieved a state of the art performance on challenging datasets. Sighthound [12] claims to have achieved the lowest mean absolute error so far and they also offer their neural network capabilities as a service to shop owners so that they can gather different statistical data about their customers, such as age and gender. Age estimation is seen as a regression problem due to the number of classi fications that are possible (0-100+) and the extensive research that was conducted using different classification methodologies and it can be viewed as a textured pattern in which the features can be used like local binary patterns, biologically inspired features and Convolutional neural networks which in recent years has seen a surge of popularity due to the outstanding performance in facial recognition. In this project, we will investigate the use of deep learning for age estimation. Instead of creating a convolutional neural network and train it from scratch, we will take advantage of transfer learning, by using a pre-trained model called VGG-FACE (trained on thousands of images for facial recognition) fine-tuning will be done to compute age estimation. A support vector regression (SVR) model is then trained on the features outputted by the CNN to make the final age estimation. A collection of over 60,000 images will be used to train and test the network; however, this quantity of images is not enough to train a deep CNN from scratch, but it is enough for this network due to transfer learning. The photos were taken from the IMBD-WIKI dataset which contains over 500,000 images combined, and the 60,000 previous images are taken from the wiki set.