A New Hybrid Pupil Detection Algorithm for Real Time Applications

This paper presents a new hybrid pupil detection algorithm based on both AI and classical image processing techniques. The problem addressed by the AI technique of neural networks is the detection of the binarization threshold. Thus, using the percentile function calculated in a sufficiently large number of points, a convolutional neural network with a reduced number of layers is used to determine the appropriate threshold necessary to obtain the binarized eye image. Furthermore, classic image processing solutions (morphological operations, Laplacian based edge detection, convexity correction, ellipse fitting) are used to detect the center of the pupil. Due to a good immunity to variable and non-uniform lighting conditions, the proposed algorithm shows a high detection rate when tested on three representative databases.