search

Kohulan / DECIMER-Image_Transformer

DECIMER Image Transformer is a deep-learning-based tool designed for automated recognition of chemical structure images. Leveraging transformer architectures, the model converts chemical images into SMILES strings, enabling the digitization of chemical data from scanned documents, literature, and patents.
MIT License
216 stars 52 forks source link

Model doesn't perform well #47

Closed AlmaEvans2910 closed 1 year ago

AlmaEvans2910 commented 1 year ago

The model performs badly on hand drawn images or pictures taken by phone.

AlmaEvans2910 commented 1 year ago

It wasn't able to recognize structures like these, as well. 22w image

OBrink commented 1 year ago

Thank you for your problem report! We are continuously working on improving our models and diversifying our training data. Could you also send us examples of hand-drawn images where DECIMER Image Transformer failed? We try to imitate the features in images that are taken by phone in our artificial training data generation pipeline and are currently working on a model that has been trained on more of these types of images. Keep an eye on the repository in the coming months. We usually update the repository far before publications about new features are published.

atef199 commented 1 year ago

The model cannot recognize pretty simple structures. I tried all of these and it gives wrong results

ethanol structural-formula-propane-chemical-formula-skeletal-formula-chemical-compound-png-favpng-at97Lc9JR8GCLVnKdUFphiUpg ethane-flat3134352308631709013 q_3_21_july201166500841672020