Towards an Efficient Framework for Data Extraction from Chart Images

In this paper, we fill the research gap by adopting state-of-the-art computer vision techniques for data extraction stage in a mining system. As shown Fig. 1, contains two subtasks, namely, plot element detection and conversion. For building robust box detector, comprehensively compare different deep learning-based methods find suitable method to detect with high precision. point fully convolutional network feature fusion module is adopted, which can distinguish close points compared traditional methods. The proposed system effectively handle various chart without making heuristic assumptions. conversion, translate detected into semantic value. A measure similarities between legends elements legend matching phase. Furthermore, provide baseline on competition of Harvesting raw tables from Infographics. Some key factors have been found improve performance each stage. Experimental results demonstrate effectiveness