Box clustering segmentation: A new method for vision-based web page preprocessing. Issue 3 (May 2017)

• Record Type: Journal Article
• Authors: Zeleny, Jan; Burget, Radek; Zendulka, Jaroslav
• Is Part Of: Information processing & management. Volume 53:Issue 3(2017:May)
• Page Start: 735
• Page End: 750
• Publication Date: 2017-05
• Subjects: Clustering -- Segmentation -- Vision-based page segmentation -- VIPS; Information storage and retrieval systems -- Periodicals; Information science -- Periodicals; Systèmes d'information -- Périodiques; Sciences de l'information -- Périodiques; Information science; Information storage and retrieval systems; Periodicals; 658.4038
• DOI: 10.1016/j.ipm.2017.02.002
• Languages: English
• ISSNs: 0306-4573

Abstract

Highlights: New, purely vision-based, segmentation technique is formally described. Only a few simple visual cues are used to assess similarity of the rectangles. Its performance better by an order of magnitude when compared with competition. Rectangle clustering is a viable way to perform web page segmentation. Abstract: This paper presents a novel approach to web page segmentation, which is one of substantial preprocessing steps when mining data from web documents. Most of the current segmentation methods are based on algorithms that work on a tree representation of web pages (DOM tree or a hierarchical rendering model) and produce another tree structure as an output. In contrast, our method uses a rendering engine to get an image of the web page, takes the smallest rendered elements of that image, performs clustering using a custom algorithm and produces a flat set of segments of a given granularity. For the clustering metrics, we use purely visual properties only: the distance of elements and their visual similarity. We experimentally evaluate the properties of our algorithm by processing 2400 web pages. On this set of web pages, we prove that our algorithm is almost 90% faster than the reference algorithm. We also show that our algorithm accuracy is between 47% and 133% of the reference algorithm accuracy with indirect correlation of our algorithm's accuracy to the depth of inspected page structure. In our experiments, we also demonstrate the advantages of producing a flat segmentation structure instead of an hierarchy.