The characters are getting missed, even after fine-tuning.
I never made any progress. I tried many different ways. Some specific
characters are always missing from the OCR result.
On Wednesday, September 13, 2023 at 12:49:20 PM UTC+3 mdalihu...@gmail.com
wrote:
> EasyOCR I think is best for ID cards or something like that image process.
> but document images like books, here Tesseract is better than EasyOCR.
> Even I didn't use EasyOCR. you can try it.
>
> I have added words of dictionaries but the result is the same.
>
> what kind of problem you have faced in fine-tuning in few new characters
> as you said (*but, I failed in every possible way to introduce a few new
> characters into the database.)*
> On Wednesday, 13 September, 2023 at 3:33:48 pm UTC+6 desal...@gmail.com
> wrote:
>
>> Yes, we are new to this. I find the instructions (the manual) very hard
>> to follow. The video you linked above was really helpful to get started.
>> My plan at the beginning was to fine tune the existing .traineddata. But, I
>> failed in every possible way to introduce a few new characters into the
>> database. That is why I started from scratch.
>>
>> Sure, I will follow Lorenzo's suggestion: will run more the iterations,
>> and see if I can improve.
>>
>> Another areas we need to explore is usage of dictionaries actually. May
>> be adding millions of words into the dictionary could help Tesseract. I
>> don't have millions of words; but I am looking into some corpus to get more
>> words into the dictionary.
>>
>> If this all fails, EasyOCR (and probably other similar open-source
>> packages) is probably our next option to try on. Sure, sharing
>> our experiences will be helpful. I will let you know if I made good
>> progresses in any of these options.
>> On Wednesday, September 13, 2023 at 12:19:48 PM UTC+3
>> mdalihu...@gmail.com wrote:
>>
>>> How is your training going for Bengali? It was nearly good but I faced
>>> space problems between two words, some words are spaces but most of them
>>> have no space. I think is problem is in the dataset but I use the default
>>> training dataset from Tesseract which is used in Ben That way I am confused
>>> so I have to explore more. by the way, you can try as Lorenzo Blz said.
>>> Actually training from scratch is harder than fine-tuning. so you can use
>>> different datasets to explore. if you succeed. please let me know how you
>>> have done this whole process. I'm also new in this field.
>>> On Wednesday, 13 September, 2023 at 1:13:43 pm UTC+6 desal...@gmail.com
>>> wrote:
>>>
>>>> How is your training going for Bengali?
>>>> I have been trying to train from scratch. I made about 64,000 lines of
>>>> text (which produced about 255,000 files, in the end) and run the training
>>>> for 150,000 iterations; getting 0.51 training error rate. I was hopping to
>>>> get reasonable accuracy. Unfortunately, when I run the OCR using
>>>> .traineddata, the accuracy is absolutely terrible. Do you think I made
>>>> some mistakes, or that is an expected result?
>>>>
>>>> On Tuesday, September 12, 2023 at 11:15:25 PM UTC+3
>>>> mdalihu...@gmail.com wrote:
>>>>
>>>>> Yes, he doesn't mention all fonts but only one font. That way he
>>>>> didn't use *MODEL_NAME in a separate **script **file script I think.*
>>>>>
>>>>> Actually, here we teach all *tif, gt.txt, and .box files *which are
>>>>> created by *MODEL_NAME I mean **eng, ben, oro flag or language code
>>>>> *because
>>>>> when we first create *tif, gt.txt, and .box files, *every file starts
>>>>> by *MODEL_NAME*. This *MODEL_NAME* we selected on the training
>>>>> script for looping each tif, gt.txt, and .box files which are created by
>>>>> *MODEL_NAME.*
>>>>>
>>>>> On Tuesday, 12 September, 2023 at 9:42:13 pm UTC+6 desal...@gmail.com
>>>>> wrote:
>>>>>
>>>>>> Yes, I am familiar with the video and have set up the folder
>>>>>> structure as you did. Indeed, I have tried a number of fine-tuning with
>>>>>> a
>>>>>> single font following Gracia's video. But, your script is much better
>>>>>> because supports multiple fonts. The whole improvement you made is
>>>>>> brilliant; and very useful. It is all working for me.
>>>>>> The only part that I didn't understand is the trick you used in your
>>>>>> tesseract_train.py script. You see, I have been doing exactly to you did
>>>>>> except this script.
>>>>>>
>>>>>> The scripts seems to have the trick of sending/teaching each of the
>>>>>> fonts (iteratively) into the model. The script I have been using (which
>>>>>> I
>>>>>> get from Garcia) doesn't mention font at all.
>>>>>>
>>>>>> *TESSDATA_PREFIX=../tesseract/tessdata make training MODEL_NAME=oro
>>>>>> TESSDATA=../tesseract/tessdata MAX_ITERATIONS=10000*
>>>>>> Does it mean that my model does't train the fonts (even if the fonts
>>>>>> have been included in the splitting process, in the other script)?
>>>>>> On Monday, September 11, 2023 at 10:54:08 AM UTC+3
>>>>>> mdalihu...@gmail.com wrote:
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> *import subprocess# List of font namesfont_names = ['ben']for font
>>>>>>> in font_names: command = f"TESSDATA_PREFIX=../tesseract/tessdata
>>>>>>> make
>>>>>>> training MODEL_NAME={font} START_MODEL=ben
>>>>>>> TESSDATA=../tesseract/tessdata
>>>>>>> MAX_ITERATIONS=10000"*
>>>>>>>
>>>>>>>
>>>>>>> * subprocess.run(command, shell=True) 1 . This command is for
>>>>>>> training data that I have named '*tesseract_training*.py' inside
>>>>>>> tesstrain folder.*
>>>>>>> *2. root directory means your main training folder and inside it as
>>>>>>> like langdata, tessearact, tesstrain folders. if you see this tutorial
>>>>>>> *
>>>>>>> https://www.youtube.com/watch?v=KE4xEzFGSU8 you will understand
>>>>>>> better the folder structure. only I created tesseract_training.py in
>>>>>>> tesstrain folder for training and FontList.py file is the main path as
>>>>>>> *like
>>>>>>> langdata, tessearact, tesstrain, and *split_training_text.py.
>>>>>>> 3. first of all you have to put all fonts in your Linux fonts
>>>>>>> folder. /usr/share/fonts/ then run: sudo apt update then sudo
>>>>>>> fc-cache -fv
>>>>>>>
>>>>>>> after that, you have to add the exact font's name in FontList.py
>>>>>>> file like me.
>>>>>>> I have added two pic my folder structure. first is main structure
>>>>>>> pic and the second is the Colopse tesstrain folder.
>>>>>>>
>>>>>>> I[image: Screenshot 2023-09-11 134947.png][image: Screenshot
>>>>>>> 2023-09-11 135014.png]
>>>>>>> On Monday, 11 September, 2023 at 12:50:03 pm UTC+6
>>>>>>> desal...@gmail.com wrote:
>>>>>>>
>>>>>>>> Thank you so much for putting out these brilliant scripts. They
>>>>>>>> make the process much more efficient.
>>>>>>>>
>>>>>>>> I have one more question on the other script that you use to train.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> *import subprocess# List of font namesfont_names = ['ben']for font
>>>>>>>> in font_names: command = f"TESSDATA_PREFIX=../tesseract/tessdata
>>>>>>>> make
>>>>>>>> training MODEL_NAME={font} START_MODEL=ben
>>>>>>>> TESSDATA=../tesseract/tessdata
>>>>>>>> MAX_ITERATIONS=10000"*
>>>>>>>> * subprocess.run(command, shell=True) *
>>>>>>>>
>>>>>>>> Do you have the name of fonts listed in file in the same/root
>>>>>>>> directory?
>>>>>>>> How do you setup the names of the fonts in the file, if you don't
>>>>>>>> mind sharing it?
>>>>>>>> On Monday, September 11, 2023 at 4:27:27 AM UTC+3
>>>>>>>> mdalihu...@gmail.com wrote:
>>>>>>>>
>>>>>>>>> You can use the new script below. it's better than the previous
>>>>>>>>> two scripts. You can create *tif, gt.txt, and .box files *by
>>>>>>>>> multiple fonts and also use breakpoint if vs code close or anything
>>>>>>>>> during
>>>>>>>>> creating *tif, gt.txt, and .box files *then you can checkpoint to
>>>>>>>>> navigate where you close vs code.
>>>>>>>>>
>>>>>>>>> command for *tif, gt.txt, and .box files *
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> import os
>>>>>>>>> import random
>>>>>>>>> import pathlib
>>>>>>>>> import subprocess
>>>>>>>>> import argparse
>>>>>>>>> from FontList import FontList
>>>>>>>>>
>>>>>>>>> def create_training_data(training_text_file, font_list,
>>>>>>>>> output_directory, start_line=None, end_line=None):
>>>>>>>>> lines = []
>>>>>>>>> with open(training_text_file, 'r') as input_file:
>>>>>>>>> lines = input_file.readlines()
>>>>>>>>>
>>>>>>>>> if not os.path.exists(output_directory):
>>>>>>>>> os.mkdir(output_directory)
>>>>>>>>>
>>>>>>>>> if start_line is None:
>>>>>>>>> start_line = 0
>>>>>>>>>
>>>>>>>>> if end_line is None:
>>>>>>>>> end_line = len(lines) - 1
>>>>>>>>>
>>>>>>>>> for font_name in font_list.fonts:
>>>>>>>>> for line_index in range(start_line, end_line + 1):
>>>>>>>>> line = lines[line_index].strip()
>>>>>>>>>
>>>>>>>>> training_text_file_name = pathlib.Path(
>>>>>>>>> training_text_file).stem
>>>>>>>>>
>>>>>>>>> line_serial = f"{line_index:d}"
>>>>>>>>>
>>>>>>>>> line_gt_text = os.path.join(output_directory, f'{
>>>>>>>>> training_text_file_name}_{line_serial}_{font_name.replace(" ", "_"
>>>>>>>>> )}.gt.txt')
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> with open(line_gt_text, 'w') as output_file:
>>>>>>>>> output_file.writelines([line])
>>>>>>>>>
>>>>>>>>> file_base_name = f'{training_text_file_name}_{
>>>>>>>>> line_serial}_{font_name.replace(" ", "_")}'
>>>>>>>>> subprocess.run([
>>>>>>>>> 'text2image',
>>>>>>>>> f'--font={font_name}',
>>>>>>>>> f'--text={line_gt_text}',
>>>>>>>>> f'--outputbase={output_directory}/{file_base_name}
>>>>>>>>> ',
>>>>>>>>> '--max_pages=1',
>>>>>>>>> '--strip_unrenderable_words',
>>>>>>>>> '--leading=36',
>>>>>>>>> '--xsize=3600',
>>>>>>>>> '--ysize=330',
>>>>>>>>> '--char_spacing=1.0',
>>>>>>>>> '--exposure=0',
>>>>>>>>> '--unicharset_file=langdata/eng.unicharset',
>>>>>>>>> ])
>>>>>>>>>
>>>>>>>>> if __name__ == "__main__":
>>>>>>>>> parser = argparse.ArgumentParser()
>>>>>>>>> parser.add_argument('--start', type=int, help='Starting line
>>>>>>>>> count (inclusive)')
>>>>>>>>> parser.add_argument('--end', type=int, help='Ending line
>>>>>>>>> count (inclusive)')
>>>>>>>>> args = parser.parse_args()
>>>>>>>>>
>>>>>>>>> training_text_file = 'langdata/eng.training_text'
>>>>>>>>> output_directory = 'tesstrain/data/eng-ground-truth'
>>>>>>>>>
>>>>>>>>> font_list = FontList()
>>>>>>>>>
>>>>>>>>> create_training_data(training_text_file, font_list,
>>>>>>>>> output_directory, args.start, args.end)
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Then create a file called "FontList" in the root directory and
>>>>>>>>> paste it.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> class FontList:
>>>>>>>>> def __init__(self):
>>>>>>>>> self.fonts = [
>>>>>>>>> "Gerlick"
>>>>>>>>> "Sagar Medium",
>>>>>>>>> "Ekushey Lohit Normal",
>>>>>>>>> "Charukola Round Head Regular, weight=433",
>>>>>>>>> "Charukola Round Head Bold, weight=443",
>>>>>>>>> "Ador Orjoma Unicode",
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> ]
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> then import in the above code,
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> *for breakpoint command:*
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> sudo python3 split_training_text.py --start 0 --end 11
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> change checkpoint according to you --start 0 --end 11.
>>>>>>>>>
>>>>>>>>> *and training checkpoint as you know already.*
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Monday, 11 September, 2023 at 1:22:34 am UTC+6
>>>>>>>>> desal...@gmail.com wrote:
>>>>>>>>>
>>>>>>>>>> Hi mhalidu,
>>>>>>>>>> the script you posted here seems much more extensive than you
>>>>>>>>>> posted before:
>>>>>>>>>> https://groups.google.com/d/msgid/tesseract-ocr/0e2880d9-64c0-4659-b497-902a5747caf4n%40googlegroups.com
>>>>>>>>>> .
>>>>>>>>>>
>>>>>>>>>> I have been using your earlier script. It is magical. How is this
>>>>>>>>>> one different from the earlier one?
>>>>>>>>>>
>>>>>>>>>> Thank you for posting these scripts, by the way. It has saved my
>>>>>>>>>> countless hours; by running multiple fonts in one sweep. I was not
>>>>>>>>>> able to
>>>>>>>>>> find any instruction on how to train for multiple fonts. The
>>>>>>>>>> official
>>>>>>>>>> manual is also unclear. YOUr script helped me to get started.
>>>>>>>>>> On Wednesday, August 9, 2023 at 11:00:49 PM UTC+3
>>>>>>>>>> mdalihu...@gmail.com wrote:
>>>>>>>>>>
>>>>>>>>>>> ok, I will try as you said.
>>>>>>>>>>> one more thing, what's the role of the trained_text lines will
>>>>>>>>>>> be? I have seen Bengali text are long words of lines. so I wanna
>>>>>>>>>>> know how
>>>>>>>>>>> many words or characters will be the better choice for the train?
>>>>>>>>>>> and '--xsize=3600','--ysize=350', will be according to words of
>>>>>>>>>>> lines?
>>>>>>>>>>>
>>>>>>>>>>> On Thursday, 10 August, 2023 at 1:10:14 am UTC+6 shree wrote:
>>>>>>>>>>>
>>>>>>>>>>>> Include the default fonts also in your fine-tuning list of
>>>>>>>>>>>> fonts and see if that helps.
>>>>>>>>>>>>
>>>>>>>>>>>> On Wed, Aug 9, 2023, 2:27 PM Ali hussain <mdalihu...@gmail.com>
>>>>>>>>>>>> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>> I have trained some new fonts by fine-tune methods for the
>>>>>>>>>>>>> Bengali language in Tesseract 5 and I have used all official
>>>>>>>>>>>>> trained_text
>>>>>>>>>>>>> and tessdata_best and other things also. everything is good but
>>>>>>>>>>>>> the
>>>>>>>>>>>>> problem is the default font which was trained before that does
>>>>>>>>>>>>> not convert
>>>>>>>>>>>>> text like prev but my new fonts work well. I don't understand why
>>>>>>>>>>>>> it's
>>>>>>>>>>>>> happening. I share code based to understand what going on.
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> *codes for creating tif, gt.txt, .box files:*
>>>>>>>>>>>>> import os
>>>>>>>>>>>>> import random
>>>>>>>>>>>>> import pathlib
>>>>>>>>>>>>> import subprocess
>>>>>>>>>>>>> import argparse
>>>>>>>>>>>>> from FontList import FontList
>>>>>>>>>>>>>
>>>>>>>>>>>>> def read_line_count():
>>>>>>>>>>>>> if os.path.exists('line_count.txt'):
>>>>>>>>>>>>> with open('line_count.txt', 'r') as file:
>>>>>>>>>>>>> return int(file.read())
>>>>>>>>>>>>> return 0
>>>>>>>>>>>>>
>>>>>>>>>>>>> def write_line_count(line_count):
>>>>>>>>>>>>> with open('line_count.txt', 'w') as file:
>>>>>>>>>>>>> file.write(str(line_count))
>>>>>>>>>>>>>
>>>>>>>>>>>>> def create_training_data(training_text_file, font_list,
>>>>>>>>>>>>> output_directory, start_line=None, end_line=None):
>>>>>>>>>>>>> lines = []
>>>>>>>>>>>>> with open(training_text_file, 'r') as input_file:
>>>>>>>>>>>>> for line in input_file.readlines():
>>>>>>>>>>>>> lines.append(line.strip())
>>>>>>>>>>>>>
>>>>>>>>>>>>> if not os.path.exists(output_directory):
>>>>>>>>>>>>> os.mkdir(output_directory)
>>>>>>>>>>>>>
>>>>>>>>>>>>> random.shuffle(lines)
>>>>>>>>>>>>>
>>>>>>>>>>>>> if start_line is None:
>>>>>>>>>>>>> line_count = read_line_count() # Set the starting
>>>>>>>>>>>>> line_count from the file
>>>>>>>>>>>>> else:
>>>>>>>>>>>>> line_count = start_line
>>>>>>>>>>>>>
>>>>>>>>>>>>> if end_line is None:
>>>>>>>>>>>>> end_line_count = len(lines) - 1 # Set the ending
>>>>>>>>>>>>> line_count
>>>>>>>>>>>>> else:
>>>>>>>>>>>>> end_line_count = min(end_line, len(lines) - 1)
>>>>>>>>>>>>>
>>>>>>>>>>>>> for font in font_list.fonts: # Iterate through all the
>>>>>>>>>>>>> fonts in the font_list
>>>>>>>>>>>>> font_serial = 1
>>>>>>>>>>>>> for line in lines:
>>>>>>>>>>>>> training_text_file_name = pathlib.Path(
>>>>>>>>>>>>> training_text_file).stem
>>>>>>>>>>>>>
>>>>>>>>>>>>> # Generate a unique serial number for each line
>>>>>>>>>>>>> line_serial = f"{line_count:d}"
>>>>>>>>>>>>>
>>>>>>>>>>>>> # GT (Ground Truth) text filename
>>>>>>>>>>>>> line_gt_text = os.path.join(output_directory, f'{
>>>>>>>>>>>>> training_text_file_name}_{line_serial}.gt.txt')
>>>>>>>>>>>>> with open(line_gt_text, 'w') as output_file:
>>>>>>>>>>>>> output_file.writelines([line])
>>>>>>>>>>>>>
>>>>>>>>>>>>> # Image filename
>>>>>>>>>>>>> file_base_name = f'ben_{line_serial}' # Unique
>>>>>>>>>>>>> filename for each font
>>>>>>>>>>>>> subprocess.run([
>>>>>>>>>>>>> 'text2image',
>>>>>>>>>>>>> f'--font={font}',
>>>>>>>>>>>>> f'--text={line_gt_text}',
>>>>>>>>>>>>> f'--outputbase={output_directory}/{
>>>>>>>>>>>>> file_base_name}',
>>>>>>>>>>>>> '--max_pages=1',
>>>>>>>>>>>>> '--strip_unrenderable_words',
>>>>>>>>>>>>> '--leading=36',
>>>>>>>>>>>>> '--xsize=3600',
>>>>>>>>>>>>> '--ysize=350',
>>>>>>>>>>>>> '--char_spacing=1.0',
>>>>>>>>>>>>> '--exposure=0',
>>>>>>>>>>>>> '--unicharset_file=langdata/ben.unicharset',
>>>>>>>>>>>>> ])
>>>>>>>>>>>>>
>>>>>>>>>>>>> line_count += 1
>>>>>>>>>>>>> font_serial += 1
>>>>>>>>>>>>>
>>>>>>>>>>>>> # Reset font_serial for the next font iteration
>>>>>>>>>>>>> font_serial = 1
>>>>>>>>>>>>>
>>>>>>>>>>>>> write_line_count(line_count) # Update the line_count in
>>>>>>>>>>>>> the file
>>>>>>>>>>>>>
>>>>>>>>>>>>> if __name__ == "__main__":
>>>>>>>>>>>>> parser = argparse.ArgumentParser()
>>>>>>>>>>>>> parser.add_argument('--start', type=int, help='Starting
>>>>>>>>>>>>> line count (inclusive)')
>>>>>>>>>>>>> parser.add_argument('--end', type=int, help='Ending line
>>>>>>>>>>>>> count (inclusive)')
>>>>>>>>>>>>> args = parser.parse_args()
>>>>>>>>>>>>>
>>>>>>>>>>>>> training_text_file = 'langdata/ben.training_text'
>>>>>>>>>>>>> output_directory = 'tesstrain/data/ben-ground-truth'
>>>>>>>>>>>>>
>>>>>>>>>>>>> # Create an instance of the FontList class
>>>>>>>>>>>>> font_list = FontList()
>>>>>>>>>>>>>
>>>>>>>>>>>>> create_training_data(training_text_file, font_list,
>>>>>>>>>>>>> output_directory, args.start, args.end)
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> *and for training code:*
>>>>>>>>>>>>>
>>>>>>>>>>>>> import subprocess
>>>>>>>>>>>>>
>>>>>>>>>>>>> # List of font names
>>>>>>>>>>>>> font_names = ['ben']
>>>>>>>>>>>>>
>>>>>>>>>>>>> for font in font_names:
>>>>>>>>>>>>> command = f"TESSDATA_PREFIX=../tesseract/tessdata make
>>>>>>>>>>>>> training MODEL_NAME={font} START_MODEL=ben
>>>>>>>>>>>>> TESSDATA=../tesseract/tessdata MAX_ITERATIONS=10000
>>>>>>>>>>>>> LANG_TYPE=Indic"
>>>>>>>>>>>>> subprocess.run(command, shell=True)
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> any suggestion to identify to extract the problem.
>>>>>>>>>>>>> thanks, everyone
>>>>>>>>>>>>>
>>>>>>>>>>>>> --
>>>>>>>>>>>>> You received this message because you are subscribed to the
>>>>>>>>>>>>> Google Groups "tesseract-ocr" group.
>>>>>>>>>>>>> To unsubscribe from this group and stop receiving emails from
>>>>>>>>>>>>> it, send an email to tesseract-oc...@googlegroups.com.
>>>>>>>>>>>>> To view this discussion on the web visit
>>>>>>>>>>>>> https://groups.google.com/d/msgid/tesseract-ocr/406cd733-b265-4118-a7ca-de75871cac39n%40googlegroups.com
>>>>>>>>>>>>>
>>>>>>>>>>>>> <https://groups.google.com/d/msgid/tesseract-ocr/406cd733-b265-4118-a7ca-de75871cac39n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>>>>>>>>> .
>>>>>>>>>>>>>
>>>>>>>>>>>>
--
You received this message because you are subscribed to the Google Groups
"tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to tesseract-ocr+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/tesseract-ocr/042222e0-335f-41b4-9539-efdc34d79eb8n%40googlegroups.com.
