A New Method for Hierarchical Image Segmentation from Visual Designs

Hierarchical image segmentation recognizes and organizes image elements into a tree structure. The tree structure represents the semantic information of the image. It is one of the most fundamental computer vision problems. This paper focuses on the images that come from visual design such as graphic interfaces, posters, and presentations. Extracting hierarchical structure from such images allows quantitative analysis of visual design choices and reproducing designs from hand drawings or hard copies. We propose a more accurate method that incorporates the common design principles of visual designs. We compare our algorithm with seven existing approaches on the most popular websites screenshots ranked by Alexa. Our method outperforms the state-of-the-art methods in tree edit distance and F-score, and is comparable or better in the bottom-up distance.