A Structure-Based Method for Building a Database of Extracted Figures from Scientific Documents: A Case Study of Iran Scientific Information Database (GANJ)

Figures in scientific documents are rich source of information. The first step in retrieving information from such figures is to build a valid figure database. To this end, we developed a system for generating figure database from scholarly Persian documents, in large scale. The first step is to parse files and extract figures and their corresponding descriptions. There are two general approaches for extracting figures from documents, one is based on image processing methods and another one is based on processing the file primitives. The focus of this paper is on later one. This approach is shown to be a better choice for the search engines because of its speed and scalability properties. We propose a structure based method that extracts the figures and their descriptions by analyzing the file layout. This information is saved in a database with a specific structure and is indexed for retrieval in the search engine.The proposed algorithm was implemented in Python programming language. As a benchmark we used the basic method in the literature which is based on the processing PDF file. We employed the proposed method in a case study on Iran scientific information database (Ganj). In this regard, 150 scientific documents were randomly chosen from Ganj database and analyzed using two mentioned methods. Based on our experimental results, the proposed method is more efficient than the basic method especially for Persian documents. There many unanswered challenges for Persian documents when using the basic method. The number of noise images resulted from the basic method is high and Persian text extracted is not well organized. Our proposed method overcomes some of these drawbacks and is recommended for generating figure database from scientific Persian documents. The proposed method is able to correctly extract about 40% of the images with their corresponding descriptions which is 10% better than the basic method.