Eye Image Segmentation Method Based on the Modified U-Net CNN Architecture

The paper presents a new eye image segmentation method used to extract the pupil contour based on the modified U-Net CNN architecture. The analysis was performed using two databases which contain IR images with a spatial resolution of 640x480 pixels. The first database was acquired in our laboratory and contains 400 eye images and the second database is a selection of 400 images from the publicly available CASIA Iris Lamp database. The results obtained by applying the segmentation based on the CNN architecture were compared to manually-annotated ground truth data. The results obtained are comparable to the state of the art.

The purpose of the paper is to present the implementation of a robust segmentation algorithm based on the U-Net convolutional neural network that can be used in eye tracking applications such as human computer interface, communication devices for people with disabilities, marketing research or clinical studies.

The proposed method improves uppon existing U-Net CNN architectures in terms of efficiency, by reducing the total number of parameters used from 31 millions to 38k. The advantages of using a number of parameters approximatly 815 times lower than the original U-Net CNN architecture are reduced computing resources consumption and a lower inference time.