The automated recognition of optical chemical structures, with the help of machine learning, could speed up research and development efforts. However, historical sources often have some level of image corruption, which reduces the performance to near zero. To solve this downside, we need a dependable algorithmic program to help chemists to further expand their research. This paper reports the results of research conducted for the Bristol-Myers Squibb-Molecular Translation competition, which was held on Kaggle and which invited participants to convert old chemical images to their underlying chemical structures, annotated as InChI text; we define this work as molecular translation. We proposed a model based on a transformer, which can be utilized in molecular translation. To better capture the details of the chemical structure, the image features we want to extract need to be accurate at the pixel level. TNT is one of the existing transformer models that can meet this requirement. This model was originally used for image classification, and is essentially a transformer-encoder, which cannot be utilized for generation tasks. On the other hand, we believe that TNT cannot integrate the local information of images well, so we improve the core module of TNT—TNT block—and propose a novel module—Deep TNT block—by stacking the module to form an encoder structure, and then use the vanilla transformer-decoder as a decoder, forming a chemical formula generation model based on the encoder–decoder structure. Since molecular translation is an image-captioning task, we named it the Image Captioning Model based on Deep TNT (ICMDT). A comparison with different models shows that our model has benefits in each convergence speed and final description accuracy. We have designed a complete process in the model inference and fusion phase to further enhance the final results.
An Image Captioning Model Based on Bidirectional Depth Residuals and its Application
