SwinOCSR: end-to-end optical chemical structure recognition using a Swin Transformer
Zhanpeng Xu, Jianhua Li, Zhaopeng Yang, Shiliang Li, Honglin Li
Abstract
Optical chemical structure recognition from scientific publications is essential for rediscovering a chemical structure. It is an extremely challenging problem, and current rule-based and deep-learning methods cannot achieve satisfactory recognition rates. Herein, we propose SwinOCSR, an end-to-end model based on a Swin Transformer. This model uses the Swin Transformer as the backbone to extract image features and introduces Transformer models to convert chemical information from publications into DeepSMILES. A novel chemical structure dataset was constructed to train and verify our method. Our proposed Swin Transformer-based model was extensively tested against the backbone of existing publicly available deep learning methods. The experimental results show that our model significantly outperforms the compared methods, demonstrating the model's effectiveness. Moreover, we used a focal loss to address the token imbalance problem in the text representation of the chemical structure diagram, and our model achieved an accuracy of 98.58%.