Pupil Detection System Implementation for Low-Resolution Eye Images Based on a Fully-Connected Neural Network Classifier

This paper presents an implementation of a pupil center detection system that uses a classifier based on neural networks. Being intended for the processing of low-resolution eye images, the implementation is done with a fully connected neural network. The classes of the classifier are defined according to the possible positions of the pupil center. The neural network was trained, tested and validated using 39,000 eye images from 18 databases, images obtained from different human subjects and under variable and non-uniform lighting conditions. The solution is distinguished by the fact that it directly provides the coordinates of the center of the pupil and not just the area where it is located. The results show both high accuracy and high processing speed, the proposed solution being suitable for low-budget implementations of real-time gaze detection systems.