Eye pupil localization algorithm using convolutional neural networks

Eye pupil localization is one of the indispensable technologies in various computer vision applications such as virtual reality and augmented reality. In general, the algorithm consists of finding the approximate eye region and finding the pupil position by extracting the semantic feature from each eye region. However, the performance of the eye pupil location is affected not only by illumination and image resolution but also by glasses wear. Therefore, this paper proposes an eye pupil localization algorithm which is robust against the above disturbance conditions and provides high accuracy. First, a face is detected from an input image and it is determined whether to wear glasses using the detected face. If glasses are present, the glasses are removed to find the correct eye region. Then, facial landmarks are extracted, and eye regions are detected based on facial landmarks. Next, the pupil region is segmented using fully convolutional networks. Finally, the position of the segmented pupil is calculated. Experimental results show that the proposed algorithm outperforms the state-of-the-art algorithms for public databases such as BioID and GI4E by up to 3.44% 0.5%, respectively.