Training Tesseract 4 on Windows

This article covers training Tesseract 4.1.1 for a particular font.

Put a otf or ttf of the font in some folder and make note of the path it is in.

Add the build folder with all the Tesseract executables to your PATH environment variable if it is not there already.

Training scripts

Tesseract source src/training subdirectory has a script tesstrain.sh that already does most of the steps of the training for us. You can copy it to elsewhere if you wish, but you also need these other scripts it depends on: tesstrain_utils.sh, language-specific.sh.

Tesstrain arguments

You can run tesstrain.sh on its own to print usage text [if not, you can check the source and it is in a string right at the top after the licence]. The important arguments are:

1. --fonts_dir FONTS_PATH

A path to a directory that contains the otf or ttf file of the font you want to train for.

2. --fontlist FONTS

The name or list of names of the font(s) you want to train for. That’s not necessarily the name of the file, nor even the ‘Font name’ Windows gives when you open it. A way to find out is run:

text2image --list_available_fonts --fonts_dir=i:/path/to/folder --fontconfig_tmpdir=i:/any/path/doesnt/matter

My font is SofiaPro-Black, and this command gives me the output 0: Sofia Pro Heavy. So ‘Sofia Pro Heavy’ is the font name to use.

3. --lang LANG_CODE

The language you are training for, e.g. eng.

4. --langdata_dir DATADIR

From langdata_lstm repository you need the folder of the language you are training for (e.g. eng), and radical-stroke.txt. You can download this whole thing, but note it is over 1 GB.

5. [OPTIONAL] --tessdata_dir TESSDATADIR

This is only needed if you want to refine an existing model, rather than training from scratch. Download the .traineddata file for the language you are training for, e.g. eng.traineddata. You can get it from tessdata, tessdata_best, tessdata_fast, or other places (you may be able to find people who share the models they’ve trained).

6. [OPTIONAL] training_text TEXTFILE

text2image by default will make the training images based on the training text for the language you are training for which comes with langdata_lstm (e.g. eng.training_text). If you like, you can provide your own text instead, for example if you need only particular characters or only digits. You need quite a lot of text (the default text has almost 200,000 lines), so you’ll probably need to write a simple script to generate some text for you into a text file and provide it with this option.

7. [OPTIONAL] --maxpages N

As mentioned, the training text tends to be very long, so you can limit the number of pages text2image will generate with this option. Pages refers to pages the tif image that is used for training will have. Seems to do 55 lines per page but I am not sure if that is always the case.

8. --linedata_only

We need to provide this flag in order to train for Tesseract 4 LSTM training rather than the legacy box training which was used in Tesseract 3.

9. [OPTIONAL] --save_box_tiff

If you provide this flag, it will save the tif image that is used for training in the output folder, so you can see what it was using.

10. --output_dir OUTPUTDIR

Needless to say, path to the folder it will output stuff to.

Optional: Customise text2image output

You can modify tesstrain_utils.sh to pass different arguments to text2image. I didn’t want it to rotate the images because I will be using my model for screenshots which are never going to be rotated. In line 283 I added to the text2image call the following arguments: --rotate_image=false --degrade_image=false.

Step 1: Initial tesstrain command

You need to preface tesstrain.sh with sh if you are running the script in cmd or PowerShell.

Example command:

sh tesstrain.sh --fonts_dir fonts --fontlist 'Sofia Pro Heavy' --lang eng  --linedata_only --langdata_dir langdata_lstm --tessdata_dir tessdata --save_box_tiff --maxpages 10 --output_dir train

With the following subfolders in the working directory:

• fonts : containing the otf/ttf of the font
• langdata_lstm : containing the aforementioned needed data from langdata_lstm
• tessdata : containing the eng.traineddata model to refine
• train : empty folder for the output

Step 2: Extract lstm from model

The next step is to extract the lstm portion of eng.traineddata (or whatever .traineddata file you are refining). You can skip this step if you are not refining an existing model.

combine_tessdata -e tessdata/eng.traineddata train/eng.lstm

The first path needs to link to the .traineddata file of the model you are refining. The second path can point anywhere; this is where to output the lstm portion of it. The -e stands for ‘extract’.

Step 3: Combine

Again, you can skip this step if you are not refining an existing model.

lstmtraining --continue_from train/eng.lstm --model_output output/Sofia --traineddata tessdata/eng.traineddata --train_listfile train/eng.training_files.txt --debug_interval -1 --max_iterations 500

• the first path (--continue_from) is to the lstm file we just extracted
• the second path (--model_output) can point anywhere; this is where to output your model and checkpoints of it during the training
• the third path (--traineddata) is to the model you are refining
• the fourth path (--train_listfile) is to a txt file that will have been created in Step 1 in the output folder you specified

Other things to note:

• --debug interval -1 is a workaround

Step 4: Finalise

This last step is to stop the training and produce our final .traineddata file, which is the new model we’ve produced.

lstmtraining --stop_training --continue_from output/Sofia_checkpoint --traineddata tessdata/eng.traineddata --model_output output/Sofia.traineddata

• the first path (--continue_from) is to the latest checkpoint of your model. change output/Sofia to whatever you set the second path to in Step 3
• the second path (--traineddata) is to the model you are refining (optional)
• the third path (--model_output) is where to output your finished model

Using your new model

Your new model will be in whatever path you provided in Step 4; output/Sofia.traineddata in my case. You can also see the checkpoints there and continue the training whenever you wish.

Tesseract uses the .traineddata models that are in the path in the TESSDATA_PREFIX environment variable, or in the path you provide it with --tessdata-dir (argument you can give tesseract when you call it). If you want to use your new model for an existing language, for example English, rename it eng.traineddata and place it there. Otherwise you can use it by providing in language whatever your traineddata file is called, for example in my case -l Sofia.

You can see the languages it detects by running tesseract --list-langs.

See also

• Livezingy’s tesstrainsh-win