Automatic Chart Decoding System Based on Deep Learning and Image Processing
Chaofan Liang, F. Xue, Zhangwei Li · 2024
Statistical Charts contain a wealth of information. As an important way to visualize data presentation, statistical charts allow viewers to obtain a complete and intuitive understanding of the content shown in a very short time. At present, the research on automatic extraction and understanding of a large amount of text information has been relatively mature. However, even the latest big artificial intelligence models cannot accurately extract statistical graphs, which are personalized and contain a large amount of information. We propose an automatic bar chart data extraction process by combining deep learning and image processing technology, and construct an intelligent bar chart decoding system. The system is divided into three parts: the classification of statistical chart types, the text detection in the image, the classification of text roles and the image extraction. The original data used to create the chart in the pan-bar graph image is extracted for downstream applications. We evaluate and compare our system on public datasets. The results show that our system has better accuracy.