Weighted Voting-Based Effective Free-Viewing Attention Prediction On Web Image Elements. Issue 2 (20th July 2020)

• Record Type: Journal Article
• Authors: Vidyapu, Sandeep; Vedula, Vijaya Saradhi; Bhattacharya, Samit
• Is Part Of: Interacting with computers. Volume 32:Issue 2(2020)
• Page Start: 170
• Page End: 184
• Publication Date: 2020-07-20
• Subjects: eye-tracking -- visual attention -- image -- webpage -- machine learning -- support vector machine; Human-computer interaction -- Periodicals; 004.019
• DOI: 10.1093/iwcomp/iwaa013
• Languages: English
• ISSNs: 0953-5438

Abstract

Abstract: Quantifying and predicting the user attention on web image elements finds applications in synthesis and rendering of elements on webpages. However, the majority of the existing approaches either overlook the visual characteristics of these elements or do not incorporate the users' visual attention. Especially, obtaining a representative quantified attention (for images) from the attention allocation of multiple users is a challenging task. Toward overcoming the challenge for free-viewing attention, this paper introduces four weighted voting strategies to assign effective visual attention (fixation index (FI)) for web image elements. Subsequently, the prominent image visual features in explaining the assigned attention are identified. Further, the association between image visual features and the assigned attention is modeled as a multi-class prediction problem, which is solved through support vector machine-based classification. The analysis of the proposed approach on real-world webpages reveals the following: (i) image element's position, size and mid-level color histograms are highly informative for the four weighting schemes; (ii) the presented computational approach outperforms the baseline for four weighted voting schemes with an average accuracy of 85% and micro F1-score of 60%; and (iii) uniform weighting (same weight for all FIs) is adequate for estimating the user's initial attention while the proportional weighting (weight the FI in proportion to its likelihood of occurrence) extends to the latter attention prediction.