🤖 AI Summary
This work addresses the low accuracy and heavy reliance on OCR and computer vision in existing PDF-based chemical structure diagram parsing. We propose the first molecular graph parsing method that directly interprets native PDF vector instructions—including path construction, Bézier curve definitions, and text placement—bypassing rasterization and OCR entirely. Our approach extracts geometric primitives, performs geometric reasoning, enforces chemical topological constraints (e.g., valency, bond angle consistency), and reconstructs structures via an SVG intermediate representation. Evaluated on USPTO and PubMed PDF test sets, our method achieves 92.4% atomic connectivity accuracy—outperforming state-of-the-art CV-based methods by 17.6%. It demonstrates strong robustness to low-resolution scans, hand-drawn styles, and legacy documents. The core contribution is the first semantic interpretation of PDF vector instructions for chemical diagram parsing, eliminating dependence on rendering fidelity and file format specifics.