Abstract: Detecting facial key point positions on images is a challenging task since facial features differ significantly from one individual to another. Even for a certain individual, there is an occurrence of wide variations due to factors such as size, position, viewing angle, and illumination effects. In this paper, we present a system that trains and compares multiple neural networks and try to optimize their learning rate constantly. This juxtaposes the different levels of accuracy obtained in predicting the facial key points in images even with a wide array of significantly varying facial features. Our method uses a simple three-layer neural network and distinct variations of convolutional neural networks.

Keywords: facial key points, neural networks, hyper-parameter optimization, deep learning, convolution networks.