Summarization of Graph Using Question Answer Approach

Abstract

From past 400 years, charts, graphs, and other visual illustration of data have become a most important medium for communicating quantitative information (Giles et al. in Proceedings of the 6th ACM/IEEE-CS Joint Conference on Digital Libraries. ACM, pp. 339–340, 2006 [1]). It contains important information and can provide essential knowledge that is not duplicated in other data formats. One important type of graph summarization is to produce small, and interactive question answer type summary based on user-based asked questions. Graphical representation, diagrams, charts, and graphs are main types of visual data in scientific domains. To detect, extract, and recognize text present in graphical are the important steps of the process. Generally textual data in charts/graphs includes axes titles, tick labels, legends, captions, notes. In general, OCR engines used, such as Tesseract OCR, ABBYY Fine Reader (Fei-Fei et al. in Proceedings of the 24th Annual ACM Symposium on User Interface Software and Technology. ACM, pp. 393–402, 2011 [2]), MODI (Microsoft Office Document Imaging) (Fomina and Vassilieva in Pattern Recogn Image Anal 23:139–144, 2013 [3]), are used to find and identify text in scanned document images. We used Tesseract OCR engine (Fei-Fei et al. in Proceedings of the 24th Annual ACM Symposium on User Interface Software and Technology. ACM, pp. 393–402, 2011 [2]) in our experiments. Particularly, the accuracy of recognition by Tesseract OCR is up to 97% when the scanned document images are used. Tesseract OCR also has some disadvantages like, it become failure in the detection of small text regions composed of one or a few words; such text regions are typical for chart images. In our proposed work, we try to resolve the failure outputs. After being cropped and deskewed, the detected text regions are passed to an OCR engine. When the proposed algorithm for text detection and localization is used as a preprocessing step, the experimental results show a significant increase in the recognition rate. For the work presented in this paper, the technical charts such as bar chart, pie chart are taken as input.