Vision-based Web Data Records Extraction

This paper studies the problem of extracting data records on the response pages returned from web databases or search engines. Existing solutions to this problem are based primarily on analyzing the HTML DOM trees and tags of the response pages. While these solutions can achieve good results, they are too heavily dependent on the specifics of HTML and they may have to be changed should the response pages are written in a totally different markup language. In this paper, we propose a novel and language independent technique to solve the data extraction problem. Our proposed solution performs the extraction using only the visual information of the response pages when they are rendered on web browsers. We analyze several types of visual features in this paper. We also propose a new measure revision to evaluate the extraction performance. This measure reflects perfect extraction ratio among all response pages. Our experimental results indicate that this visionbased approach can achieve very high extraction accuracy.