Predicting Saccadic Eye Movements in Free Viewing of Webpages

Attention modeling for webpages has emerged as a new research direction in computer vision. Despite an amount of research effort, most studies have focused on estimating webpage saliency to reveal the static location of human fixations. Without temporal information, existing models cannot interpret the dynamic properties of the actual attention process in free-viewing webpages. To solve this problem, we propose a webpage-based saccadic model in this study to model dynamic visual search behaviors of humans when they view webpages. In the first stage, we utilize the support vector machine to learn the mapping from multilevel saliency features to an initial probability of being fixated. In the second stage, we combine the mechanisms of spatial bias and inhibition of return with the estimation of the initial probability to iteratively predict a sequence of successive fixations for each webpage image. Experimental results on a benchmark eye-tracking data set for webpages have demonstrated that the proposed model outperforms the state-of-the-art saccadic methods.