Questions regarding fine tuning of Tesseract 4.00alpha LSTM

[Wang Zhimin]

Hi all, 

Thank you in advance.

I have questions regarding the accuracy improvement with fine tuning of the LSTM model. 

BACKGROUND: 

I want to use tesseract to recognise DNA/RNA sequences from PDF/TIFF. However, the accuracy is not great as the images have different font types and sizes.

Method:

I understand that I probably have two options:

1. With the source images, I run the tesseract to generate the boxes, manually correcting them using jTessBoxEditor to edit them and retrain a new eng_dna.traindata file.
2. With the current eng best LSTM train data file, fine tune the network with a bunch of sequences texts.

Questions and concerns:

• Can I mix different font type in the training data images?
• Do I need to rely on any existing train data file? Since I want to recognise some normal words and numbers in the DNA/RNA sequence images too.
• I understand LSTM is line based recognition. Will it accept the mix font training images with boxes.
• Which one is the right one for my problem? Really have no clue and experience when it comes to training your own model.

Thank you all.