A Systematic Review of Deep Learning Approaches to Visual Saliency Prediction on Webpage Images

A webpage provides important visual information for branding and business goals. Studying visual attention and saliency on webpages can lead to understanding where human eyes look at the salient region. With the growth of machine learning and deep learning, models from previous studies have achieved incredible improvements in saliency prediction, particularly for natural image datasets. However, webpage image datasets are different from natural images. Unfortunately, there is a lack of extensive research on predicting visual saliency for webpage images. To address the scientific gap in saliency prediction for webpage images, we conducted a systematic literature review (SLR) based on Kitchenham’s method. This approach enables a thorough examination of existing research in this area. Our review includes 13 relevant articles published between 2014 and 2023. This study also reviews several deep learning models or methods developed in previous papers, compares the most common saliency evaluation metrics, identifies the different saliency prediction models used by researchers for benchmarking, and provides a comprehensive discussion of current limitations and future developments in this field of study. Our analysis reveals several key challenges in the current research. One significant issue is the limited fixation data available in existing webpage datasets. Additionally, there is a difficulty in adapting saliency models trained on natural images to webpage contexts due to notable distributional shifts between these domains.