Blinkresnet: Cost-Effective Eyelid Localization for Use in a Blink Reflexometer

Abstract

Four keypoints corresponding to upper and lower eyelid positions may be localized in the scan videos of a blink reflexometer, an edge device. Images from prior reflexometer scans are tagged in order to train a residual neural network, dubbed BlinkResNet. In order to provide adequate demographic coverage, comprehensive data collection procedures are implemented. The architecture is thoroughly parameterized, allowing for a multidimensional space of compatible architectures to be swept through using a random search. Experimental results on a dataset of annotated reflexometer frames demonstrate significant improvement over a previous algorithm for eyelid localization. A random search identified variants of BlinkResNet best-suited for reflexometer frame annotation, with BlinkResNet-Default achieving a [Formula: see text] score of 0.966 on the validation dataset. Requirements for deployment on an edge device are met: training ease, generalizable inference performance, and low computational cost. A sweep may be conducted on any combination of hyperparameters and datasets, making it an adaptable tool for the continual development of customized machine learning architectures. The analysis of sweep results provides valuable insights into optimal model organization for specific tasks. Thus, methodical data annotation and sweeps over a parameterized machine learning routine are investigated as measures to achieve rapid, accurate inference results.