Zhenger/MolA

Towards Reliable Optical Chemical Structure Recognition: Better Metrics, Better Data and Better Reconstruction

We introduce a comprehensive approach that improves OCSR evaluation, training data, and model reliability. Key Contributions:

• Better Metrics: We introduce a comprehensive evaluation scheme that directly quantifies manual correction costs, with Perfectly-Matched Annotation Accuracy (PMAA) as the most stringent metric, ensuring exact alignment between model outputs and input images, including atom text, coordinates, and bond types.
• Better Data: We construct the largest and most accurately annotated real-world patent dataset, comprising 2,504,937 chemical structure images. Additionally, we introduce an independent test set, Annotated-USPTO (1,325 images), which covers a diverse range of structural patterns, providing a rigorous benchmark for evaluating model reliability.
• Better Reconstruction: We develop MolA, an end-to-end OCSR model optimized for higher recognition accuracy and reduced manual intervention. Experimental results show that MolA significantly improves PMAA from 3.85% (prior state-of-the-art) to 89.74%, drastically reducing manual corrections and setting a new standard for reliable OCSR.

The code are publicly available for research purposes at https://github.com/Zhenger959/MolA.

If you use this dataset in your research, please cite the following paper:

@article{zheng2025mola,
  title     = {Towards Reliable Optical Chemical Structure Recognition: Better Metrics, Better Data and Better Reconstruction},
  author    = {Jiaxin Zheng and Rong Ma and Cuinan Yu and Xu Wang and Fengfeng Zhou and Jianyang Zeng},
  journal   = {Preprint},
  year      = {2025},
  url       = {https://github.com/Zhenger959/MolA}
}