A deep learning framework for scientific chart data extraction and reconstruction

Scientific figures encapsulate key empirical evidence and are critical to data-driven discovery and scientific integrity assessment. Despite substantial advances in scientific natural language processing for scientific document analysis, automatic quantitative extraction of quantitative data from visual plots remains a challenge due to graphical heterogeneity, noise, and non-standardized formats. In this paper, we present ChartRecover, a deep learning-based framework for end-to-end extraction and reconstruction of chart data from scientific figures. ChartRecover adopts object detection to robustly identify graphical elements across diverse plot types, and introduces a coordinate transformation strategy that aligns visual pixel coordinates with real-world numerical values through axis tick-mark alignment and adaptive conversion. We benchmark ChartRecover across a wide range of figure styles and perturbation conditions, demonstrating strong generalization and high-fidelity recovery performance. The extracted structured data enables downstream applications such as structure-property relationship mining and automated scientific verification. Our work advances machine understanding of scientific figures and provides a scalable tool to enhance research transparency, reproducibility, and data accessibility.