ChartReader: Automatic Parsing of Bar-Plots
Chinmayee Rane, Seshasayee Mahadevan Subramanya, Devi Sandeep Endluri, Jian Feng Wu, Clyde Lee Giles · 2021
Scientific figures such as bar graphs are a critical part of scientific research and a predominant method used to represent trends and relationships in data. However, manually interpreting and extracting information from graphs is often tedious. Since data consumption has exponentially evolved over the past few decades, there is a need for automated data inference from these bar graphs. ChartReader presents a fully automated end-to-end framework that extracts data from bar graphs in scientific research papers focusing on process engineering and environmental science journals. ChartReader uses a deep learning-based classifier to determine the chart type of a given chart image. We then develop novel heuristic methods for analyzing scientific figures (text detection, pixel grouping, object detection) and address prime challenges like axis detection, legend parsing, and label detection. Our framework achieves 98% and 68% accuracy in parsing x-axis and y-axis ticks, respectively. It achieves 83% accuracy in parsing legends and 42% accuracy in parsing data values in the testing corpus. We compare the proposed method with state-of-the-art methods and address its limitations.