A Light-Weight Human Eye Fixation Solution for Smartphone Applications

Abstract

A wide range of human eye fixation prediction algorithms have been presented in the research with the advent of deep learning. However, to generate better prediction outcomes, these methods are becoming increasingly complicated. In this study, we present a lightweight human eye fixation prediction network that is based on a low-complexity representation learning network and can handle a variety of real-world data. The method includes a simplified multilevel feature extraction network with an emphasis on channel and spatial attention mechanism. We investigate the effectiveness of the present technique in predicting eye fixation maps on a collection of challenging images from the SALICON and MIT1003 datasets. A comprehensive qualitative and quantitative evaluation revealed that the network could learn and capture spatial and semantic information in a scene effectively, resulting in a higher hit rate and fewer false positives in comparison with the competing solutions. The approach is implemented on Samsung Galaxy S23 with SnapDragon-SM8550 mobile platform given its short inference time of 1.4ms and low model complexity.