This is the official implementation of our paper: "RFL: Simplifying Chemical Structure Recognition with Ring-Free Language". Accepted by AAAI 2025 oral.
Paper arxiv: Paper
- 2025.01.20. Our paper is selected as AAAI 2025 oral, congratulations 👏👏👏.
- The source code including training and inference has relase.
TODO:
- Update paper link in arxiv.
- Update Source Code.
The primary objective of Optical Chemical Structure Recognition is to identify chemical structure images into corresponding markup sequences. In this work, we propose a novel Ring-Free Language (RFL), which utilizes a divide-and-conquer strategy to describe chemical structures in a hierarchical form. RFL allows complex molecular structures to be decomposed into multiple parts. This approach significantly reduces the learning difficulty for recognition models. Leveraging RFL, we propose a universal Molecular Skeleton Decoder (MSD), which comprises a skeleton generation module that progressively predicts the molecular skeleton and individual rings, along with a branch classification module for predicting branch information. Experimental results demonstrate that the proposed RFL and MSD can be applied to various mainstream methods, achieving superior performance compared to state-of-the-art approaches in both printed and handwritten scenarios.
Comparasion of RFL with previous modeling language:
Our Model Architecture:
In Our paper, we use two dataset as follows.
- EDU-CHEMC : A dataset for handwritten chemical structure recognition.
- Mini-CASIA-CSDB : A dataset for printed chemical structure recognition.
Our Ring-Free Language (RFL) utilizes a divide-and-conquer strategy to describe chemical structures in a hierarchical form. For a molecular structure
You can use the following command to generate Ring-Free Language of single samples. We have provided some typical examples for testing in ./RFL/RFL.py:
cd RFL
python RFL.pyBatch generation of multiple process using mutli-processings:
cd RFL
bash RFL_gen.shYou can start training using the following command:
bash train.shNote: The dataset path and related paramaters need to be modified in rain\config.py
bash test_organic.shComparison with state-of-the-art methods on handwritten dataset (EDU-CHEMC) and printed dataset (Mini-CASIA-CSDB).
Ablation study on the EDU-CHEMC dataset, with all systems based on MSD-DenseWAP.
| System | MSD | [conn] | EM | Struct-EM |
|---|---|---|---|---|
| T1 | × | × | 38.70 | 49.45 |
| T2 | × | √ | 44.02 | 55.77 |
| T3 | √ | × | 52.76 | 58.58 |
| T4 | √ | √ | 64.96 | 73.15 |
To prove that RFL and MSD can simplify molecular structure recognition and enhance generalization ability, we design experiments on molecule complexity.
Exact match rate (in %) of DenseWAP and MSD-DenseWAP along test sets with different structural complexity. The left subplot is trained on complexity {1,2}, and the right subplot is trained on complexity {1,2,3}.
Case Study:
If you find our work is useful in your research, please consider citing:
@inproceedings{chang2025rfl,
title={RFL: Simplifying Chemical Structure Recognition with Ring-Free Language},
author={Chang, Qikai and Chen, Mingjun and Pi, Changpeng and Hu, Pengfei and Zhang, Zhenrong and Ma, Jiefeng and Du, Jun and Yin, Baocai and Hu, Jinshui},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={39},
number={2},
pages={2007--2015},
year={2025}
}
If you have any question, please feel free to contact me: qkchang@mail.ustc.edu.cn