Training Tesseract 5 for a New Font 

Could you help me with the directory structure. I am a bit confused on how it is made?

I tend to think about tesseract documentation as the Augean Stables

did find any answers?

I'm pretty sure, as long as text2image works correctly. If text2image doesn't work correctly you can either come up with another clever ways (like Python scripts) of automatically generating ground truth data (.gt.txt, .box and .tif files), or worst case, create them manually.

Hey, long time passed, But for others having the same issue, creating an virtual env and then installing requiremnets.txt (of the tesstrain repo) in it may fix the issue, at least for me it worked! also check if all txt files have related box files or not!

I have seen people use the jTessBoxEditor: vietocr.sourceforge.net/training.html

@@AstuteJoe one more question, how would you use the newly trained model in python? Thank you

@@ivanmongebadilla9454 I think just a parameter lang='your_new_model_name' as long as the new model is in the tessdata folder

Is it necessary for the box_file to be for each character or is it okay for it to be one word per bounding box?

Check the comments, a bunch of people train it for this exact intent

Hey, It seems like you were also looking to train tesseract with handwritten text. Did you do it? If so please shade light, I am so lost

@@dhirazz training is not an easy thing as you need huge amt of data and they as well clearly said training is not gonna make any sense ( google ) hence,if u wanna try adjusting the parameters then deep dive into cpp

I am also getting this.. running as sudo helped a bit still getting this "Unable to open '../tmp/fonts.conf' for writing: No such file or directory"

just top add.. I am getting eng_XX.box f eng_XX.tiff and eng_xx.gt.txt

@@jayrigger7508 I have resolved the issue by just changing the --font flag to /usr/share/fonts

fishshell.com/

такая же проблема

Have you been able to figure it out? I'm having the same problem

@@pauliusliaudenskas9269 к сожелению, не смог. Не понимаю, как это сделать

Your folder was probably named "Apex" and not "Apex-ground-truth"

On windows text2image will never find the font name so better install some lnx vertual machine

Hey you collected font but whats the training text data is that of Aurbesh ?

Hi ! I don't know if you resolved your issue since 1 month but don't forget to set a huge MAX_ITERATIONS to the make training. I personally set it to 10000 and it was quite better, it might be really enough for you if you have 100000 sentences. If you want to know the result check the log while the model is training, for example : At iteration 7800/7800/7800, Mean rms=5.642000%, delta=49.022000%, BCER train=97.817000%, BWER train=100.000000%, skip ratio=0.000000%, New best BCER = 97.817000 wrote checkpoint. BCER is the error rate for characters and BWER the error rate for words, you can see that at iteration 7800 it was higher than 95% and after the 9500th iteration I got several improvements.

This comment isn't ai

I mean it's not rendered with ai

Only experimentation will tell, but Tesseract 5 does perform better some times

I think that you should install this font in your system first :)

Honestly, I think your best bet is cloning the GitHub repo, readings the docs and then delving onto code, just reading it, eventually you'll be better at knowing where to look and after trying hard you might be comfortable and understand it. And I'm pretty sure in the docs you can dump and inspect some intermediary steps debug files, finally, be sure to run it on verbose mode, probably -v. Ah, and you can compile it with debugging symbols too, should help if you want to set breakpoints etc

It will never work for windows

Can you send me the whole file? Pastebin or GitHub does it, I believe I know exactly how to fix but I need the whole file to send you the fixed version

Did you solve the problem ? if yes could you share it with me please ?

@@AstuteJoe Did you solve the problem ? if yes could you share it with me please ?

I think you have to type encoding='utf-8' insine open function: with open(training_text_file,'r',encoding='utf-8') as input_file:

Having the same problem. Still trying to understand what it is looking for...

@@TheComputerChip I gave up, looked at another method that uses the Google colab and create my own model there it works pretty well. Don't know the video anymore cause probably between then and now I watched approximately 250 vids. Not kidding I don't have a life

@@Bengeljo hahaha no worries. I actually ended up getting this to work. The error doesn’t seem to affect the output oddly enough. As long as it finds the font everything still runs. Currently waiting as my PC generates the images and then I’ll sleep as it trains. On video #3 since starting the image creation! lol

​@@Bengeljo Can you share the Google Colab workflow you followed to create your model?

MODEL_NAME=foo

I believe it works, but definitely not every step exactly like in the video. But as far as I remember the Tesseract mantainers highly recommend Linux instead

I tried on Windows, not working very well, pls let me know if you are able to solve it

So, are you saying this video is now not useful at all?

Glad you liked it! In this tutorial you can see I actually fine-tuned, I started on the eng.traineddata file from Tesseract and trained it further on a new font, this should be enough for most cases.

​@@AstuteJoe Hi Gabriel, when I fine-tune I get a very bad result. I just wanna add some new words and some characters, but the final file that I get is worse than the main traineddata file. I'm trying to fine-tune an RTL language. Thanks a lot.

@@farazsoftinfo That's a very different rabbit hole, that's ML techniques, you might be overfitting (training too much) or underfitting (training too little) your model, have you tried generating all the 193k PDFs to train and leaving it to train for a bit?

@@farazsoftinfo I'm trying to do the same thing and the result is a bunch of "mmmoooomom...". Is yours the same?

​@@gabriel2011gabriel I tried it for Persian, but I couldn't get a good result. The main models are still better than what I got. When I try to add some new words and fonts I get a worse model. Maybe I should check it more to figure out the best settings that work for the RTL languages.

Hello, I am currently stuck. Where did you upload your new font ttf file in the code directory like where should it be place? I have a custom font ttf file that I want to train on?

Hi Just run: set TESSDATA_PREFIX="../tesseract/tessdata" and then run the text2image....

Did you generate the .box files successfully?

​@@AstuteJoe I cleaned the box files but now I get a different error Here is my output: + tesseract data/myFont-ground-truth/eng_2.tif data/myFont-ground-truth/eng_2 --psm 13 lstm.train read_params_file: Can't open lstm.train + tesseract data/myFont-ground-truth/eng_0.tif data/myFont-ground-truth/eng_0 --psm 13 lstm.train read_params_file: Can't open lstm.train + tesseract data/myFont-ground-truth/eng_5.tif data/myFont-ground-truth/eng_5 --psm 13 lstm.train read_params_file: Can't open lstm.train + tesseract data/myFont-ground-truth/eng_7.tif data/myFont-ground-truth/eng_7 --psm 13 lstm.train read_params_file: Can't open lstm.train + tesseract data/myFont-ground-truth/eng_3.tif data/myFont-ground-truth/eng_3 --psm 13 lstm.train read_params_file: Can't open lstm.train + tesseract data/myFont-ground-truth/eng_1.tif data/myFont-ground-truth/eng_1 --psm 13 lstm.train read_params_file: Can't open lstm.train find -L data/myFont-ground-truth -name '*.lstmf' | python3 shuffle.py 0 > "data/myFont/all-lstmf" Error: missing ground truth for training make: *** [data/myFont/list.train] Error 1 Your help will be very appreciated 🙂

@@PsychologicalHeat Did you generate the .txt.gt files? Those are text files with the actual text in them

​@@AstuteJoe Yes, I have all gt.txt, .box, and .tiff files I think the problem is that I want the ocr to read only uppercase letters? I have made a custom training_text file and it only has numbers, '-' and uppercase letters. I played around with it and now this is the output: find -L data/myFont-ground-truth -name '*.gt.txt' | xargs paste -s > "data/myFont/all-gt" unicharset_extractor --output_unicharset "data/myFont/unicharset" --norm_mode 2 "data/myFont/all-gt" Bad box coordinates in boxfile string! 36-XR-34928-PN-54460-TN-50758-XB-02919-JP-10263-DG-99350-MF-07358-PK-31144-MB-35731-ZX-758 Extracting unicharset from plain text file data/myFont/all-gt Other case x of X is not in unicharset Other case r of R is not in unicharset Other case p of P is not in unicharset Other case n of N is not in unicharset Other case t of T is not in unicharset Other case b of B is not in unicharset Other case j of J is not in unicharset Other case d of D is not in unicharset Other case g of G is not in unicharset Other case m of M is not in unicharset Other case f of F is not in unicharset Other case k of K is not in unicharset Other case z of Z is not in unicharset Wrote unicharset file data/myFont/unicharset make: *** No rule to make target `data/myFont-ground-truth/myFont_1.lstmf', needed by `data/myFont/all-lstmf'. Stop.

I imagine you could edit the tex2image utility source code to introduce the variance you need, tesseract is open source

Train it and the use it

It has to be installed on your system, each OS will have a different way of doing it

@@AstuteJoe i'm using ubuntu. Is there any way to specify the directory?

@@asiburrahman3623 askubuntu.com/questions/3697/how-do-i-install-fonts

@@AstuteJoe I have installed the font but still this error shows: Fontconfig warning: "/tmp/fonts.conf", line 4: empty font directory name ignored Fontconfig warning: "/tmp/fonts.conf", line 4: empty font directory name ignored Could not find font named 'Apex'.

@@asiburrahman3623 I also have the same problem. I tried to add '--fontconfig_tmpdir={fontconf_dir}'. >> the default is /tmp which doesn't have our font directory in it. fonts.conf is usually located in etc/share/fonts. Now, I could create .box and .tif files. Hope this solution could solve your issue too.

Not even sure what is this file now but here you go, this one is on my home folder: /home/gabri/tesseract_training/apex_legends.otf

This one is on the tesseract project folder:

I have made a good progress today, I am able to train the Apex font, however when I switch to another font Nato Sans, it's able to generate box and tff but it shows error while training "Makefile:219: *** found no data/Noto Sans-ground-truth/*.gt.txt for Sans/all-gt. Stop." . Seem it does not accept font's name with space in middle ..

@@TuanLe-ve7lm That could definitely be it, spaces and Linux (or Windows) don't mix well

That's really odd, I see you changed the START_MODEL so it should work, not super sure now

Do you have ara.traineddata in the tessdata folder?

@@AstuteJoe Yes I have and made sure of it couple of times

@@blndazeez1973 Maybe it's because the Apex model was already created when you were trying it out? And it's already on top of the eng trained data?

@@AstuteJoe I redo the steps with different model name but gives me the same error, that is strange.

lol i love you too

did you manage to train tesseract for bender yet?

I solved this issue 😅

@@Kronzplayz. how you solved it? I got the same issue

@@enriqueortiz5875 solved it: need to run in tesstrain folder: make leptonica tesseract make tesseract-langdata

command : TESSDATA_PREFIX=../tesseract/tessdata make training MODEL_NAME=Apex START_MODEL=eng TESSDATA=../tesseract/tessdata MAX_ITERATIONS=100 combine_tessdata -u ../tesseract/tessdata/eng.traineddata data/eng/Apex

tesseract "data/Apex-ground-truth/eng_44.tif" data/Apex-ground-truth/eng_44 --psm 13 lstm.train + tesseract data/Apex-ground-truth/eng_44.tif data/Apex-ground-truth/eng_44 --psm 13 lstm.train python3 shuffle.py 0 "data/Apex/all-lstmf" + head -n 90 data/Apex/all-lstmf + tail -n 10 data/Apex/all-lstmf combine_lang_model \ --input_unicharset data/Apex/unicharset \ --script_dir data/langdata \ --numbers data/Apex/Apex.numbers \ --puncs data/Apex/Apex.punc \ --words data/Apex/Apex.wordlist \ --output_dir data \ \ --lang Apex Failed to read data from: data/Apex/Apex.wordlist Failed to read data from: data/Apex/Apex.punc Failed to read data from: data/Apex/Apex.numbers Loaded unicharset of size 113 from file data/Apex/unicharset Setting unichar properties Other case É of é is not in unicharset Other case FI of fi is not in unicharset Setting script properties Failed to load script unicharset from:data/langdata/Latin.unicharset Warning: properties incomplete for index 3 = C Warning: properties incomplete for index 4 = H Warning: properties incomplete for index 5 = E Warning: properties incomplete for index 6 = S Warning: properties incomplete for index 7 = - Warning: properties incomplete for index 8 = R Warning: properties incomplete for index 9 = I Warning: properties incomplete for index 10 = K Warning: properties incomplete for index 11 = N Warning: properties incomplete for index 12 = G Warning: properties incomplete for index 13 = B Warning: properties incomplete for index 14 = 8 Warning: properties incomplete for index 15 = 5

@@datarkmveri2228 solved it: need to run in tesstrain folder: make leptonica tesseract make tesseract-langdata

solved it: need to run in tesstrain folder: make leptonica tesseract make tesseract-langdata

thanks

same to me, have you had a solution yet

@@TuanLe-ve7lm No, sadly I gave up on how to train Tesseract 5. I'm going to try to learn how to train Tesseract 4 because there are a lot more videos on youtube.

@@TuanLe-ve7lm I actually fixed the issue by using Linux. But now I get other errors lol

@@faint.2396 Did you fix your problem ?

I think it could be related with text2image itself. You have to provide path to text2image.exe which in general is located in installed tesseract.

Hi, were you able to solve this error? If so, how?

someone can help me?

Are you getting "not recognized" error. I am getting the same error on Windows. The exact command works fine on the Mac. Very strange. Do you find a solution?

fr and I did every step the same and I'm getting errors. Why isn't training Tesseract 5 simple as Tesseract 4? And the thing is there's only ONE video on how to train Tesseract 5 and its this one.