Multiscale and Nonlinearility Convolutional Regression for Locating the Eye's Pupil Center

• Published in: Proceedings of the ... International Joint Conference on Computer Science and Software Engineering (Online) pp. 47 - 52
• Main Authors: Lertsiravarameth, Phitchapha, Taeprasartsit, Pinyo
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.11.2020
• Subjects: convolutional neural network; data augmentation; deep learning; Image color analysis; Iris; Lighting; Measurement; Monitoring; pupil localization; Pupils; Statistics
• ISSN: 2642-6579
• DOI: 10.1109/JCSSE49651.2020.9268375

Pupil detection has been an active research topic in recent years due to its usability in many areas including interaction design and medical diagnosis. Many existing methods are accurate in pupil localization; however, many of them were designed and tested under specialized lighting conditions. Furthermore, there are challenges to cope with images whose color of the iris and the pupil are both dark. In this work, we proposed deep regression models to locate the eye's pupil center in target groups whose iris and pupil are both dark in color and images were acquired in a typical lighting condition. The proposed neural network model concurrently utilizes features from multiple convolutional layers for final regression. In other words, features from different layers which contain different degrees of nonlinearities are concatenated so that a neural network model can explicitly employ these features together for predicting the location of the eye's pupil. Ablation analysis indicated that these model characteristics were essential for robustness. All experimented models were trained and validated against 2,500 eye images from the MPIIGaze dataset. Additional 2,515 test images were obtained from the MMU2 and BioID datasets to evaluate robustness of the method. Our experiments showed that robust and accurate localization of the pupil center could be achieved over datasets whose image were acquired under non-specialized lighting conditions. Our best model had average pixel error of 1.11 pixels, and had error within two pixels for 95.5 percent of test samples.