Eye-Tracking Analysis with Deep Learning Method

Abstract

The eyes are a rich source of information about mental activities as well as providing the perception of the outside world. Because they cannot be consciously controlled, they can reveal unique characteristics such as preferences and intentions. For this reason, eye-tracking technology is widely used in medicine, gaming, and commercial applications. In this study, an estimate of what type of text is read was made using the analysis of eye movements during daily reading activity. In the study, deep learning approaches were preferred due to the insufficient results of machine learning approaches before. Multiplexing was performed using a dataset with 52 features consisting of 20 participants (10 males, 10 females). 627 data were obtained as a result of multiplexing from 20 data. As a result of the creation of visual representations (spectrograms) of the data produced in sufficient numbers and processing with deep learning architectures, a good success rate of 97.88% was achieved with AlexNet. While the best values in news and text types were obtained with AlexNet and Resnet101, better results were produced with ResNet18 and ResNet50 in comedy with high visual content. It was noticed that the success rate in women was higher in documents with visual content.