Training model for English Data

[Bhavik-Ardeshna]

Hi @csxmli2016 

Thank you for creating such a good project!

I want to train this model for English data. I also have some custom LR and HR-paired image data. After reading your paper, I came to know that you are using synthetic data generated by GAN. 

• Can I use my custom data to train the model?

• To train with English data, I have to use only English corpus, right? And, the data generated using text corpus and background image will get stored or not?

• Regarding the model inference and size: To deploy the model of inferencing, do I have to use all the checkpoint files present in the MACRONet release section, or do I only need to use OCR and model checkpoints?

There are some good OCRs, or text recognition, MORAN for English data. I think I can use this model for OCR.

Thank you!

[csxmli2016]

Hi @csxmli2016 

Thank you for creating such a good project!

I want to train this model for English data. I also have some custom LR and HR-paired image data. After reading your paper, I came to know that you are using synthetic data generated by GAN. 

• Can I use my custom data to train the model?
• To train with English data, I have to use only English corpus, right? And, the data generated using text corpus and background image will get stored or not?
• Regarding the model inference and size: To deploy the model of inferencing, do I have to use all the checkpoint files present in the MACRONet release section, or do I only need to use OCR and model checkpoints?

There are some good OCRs, or text recognition, MORAN for English data. I think I can use this model for OCR.

Thank you!

• Can I use my custom data to train the model?

If your training data contains LR input, HR ground-truth, the location of each character, and the recognition of each character, you can use them to train the model.

• To train with English data, I have to use only English corpus, right? And, the data generated using text corpus and background image will get stored or not?

Yes, you can replace our text corpus with your English corpus. We synthesize the training data online, and will not store them.

• Regarding the model inference and size: To deploy the model of inferencing, do I have to use all the checkpoint files present in the MACRONet release section, or do I only need to use OCR and model checkpoints?

Yes. All the models are necessary for reference.

Yes, you can use MORAN for English data. But it is better to fine-tune the detection of character location. In my experience, the Prior generation model and SR model are very stable. The performance limitations are character recognition and location detection.