Improving Machine Understanding of Human Intent in Charts

Charts composed of images and text are a classical and compact method of displaying and comparing various data. Automated data extraction from charts is critical in the machine understanding and recognition of human intent and the meaning inherent in charts. Complex processes of automated chart data extraction have been divided into multiple basic tasks. However, these problems have not been solved well. In this paper, we principally focus on three key sub-tasks, including chart image classification (Task-1), text detection and recognition (Task-2), and text role classification (Task-3). For these tasks, we design and propose a set of effective methods. The experiments on the Adobe Synthetic and PubMedCentral datasets successfully demonstrate the effectiveness of our proposed systems. Notably, our proposed method outperforms competing systems from the ICDAR 2019 and ICPR 2020 CHART-Infographics competitions and achieved state-of-the-art performance. We hope this work will serve as a step toward enhancing the machine understanding of charts and inspiring new avenues for further research in the field of document analysis and recognition.