ChartMaster: An End-to-End Method to Recognize, Redraw, and Redesign Chart

Abstract

Chart is one kind of ubiquitous information, which is widely utilized and easy for people to understand. Due to there are so many different kinds and different styles of charts, it is not an easy task for a computer to recognize a chart, as well as to redraw the chart or redesign it. This study proposes a three-stage method to chart recognition: analyze the classification of charts, analyze the structure of charts, and analyze the content of charts. When classifying charts, we choose ResNet-50. When recognizing the structure and content of charts, we use different deep frameworks to extract key points based on different types of charts. Besides, we also introduce two datasets, UCCD and UCID, to train deep models to classify and recognize charts. Finally, we utilize some traditional geometric methods to obtain an original table of a chart, so we can redraw it.