CurveHunter: A Multi-Agent Vision-Language System for Automated Adsorption Isotherm Curve Extraction

Extracting numerical data from scientific curve plots has become a critical task in enabling large-scale data analysis and knowledge discovery from chemistry literature, where accurate curve digitization is commonly required for meta-analysis and computational modeling. Since manual extraction is both costly and time-consuming, recent research has focused on developing automatic curve extraction systems. However, current approaches usually suffer from issues of either geometric precision or semantic understanding. To address these limitations, we introduce CurveHunter, a novel multi-agent vision-language system that synergistically combines classical computer vision for geometric precision with the semantic reasoning of modern VLMs. First, to solve the critical challenge of identifying plots within cluttered document pages, the system employs our innovative Contour-Based Cropping Recovery (CBCR) module for robust and accurate figure localization. Second, to decipher visually complex graphs, our Multi-Modal Curve Element (MMCE) framework accurately digitizes and matches intricate patterns, effectively resolving ambiguities from overlapping or stylistically similar curves. Extensive experiments on a challenging dataset of nitrogen adsorption isotherms from 346 peer-reviewed articles demonstrate the effectiveness of CurveHunter. The results confirm its superior performance, achieving high curve completeness above 0.79 and strong geometric fidelity (R² > 0.49).