Automatic Data Extraction from 2D and 3D Pie Chart Images

Abstract

Due to the perceptual advantages, pie charts are frequently used in digital documents to represent meaningful numerical data. Automatic extraction of underlying slice data of pie charts is necessary for further processing of chart data. In this paper, a novel technique have been presented for identification of pie charts in document images followed by the extraction of chart data. To identify pie charts in documents, a Region-based Convolutional Neural Network (RCNN) model has been trained with 2D and 3D pie chart images. Then, different slices of a pie chart are analyzed using image gradients as one of the primary feature and compute the values of different slices. The algorithm successfully identifies different 3D structural information of a 3D pie chart which are used only for a 3D representation of such charts and are excluded from processing. To demonstrate the superiority, the algorithm has been tested on a number of 2D and 3D pie chart images.