Helpful tesseract fine-tuning scripts from https://www.statworx.com/en/content-hub/blog/fine-tuning-tesseract-ocr-for-german-invoices/

draw_box_file_data.py

#!/usr/bin/python
# -*- coding: utf-8 -*-

"""Script to draw bounding boxes and text from a Tesseract box file.

The script takes an image TIFF_FILE, draws the text and bounding boxes
of the corresponding BOX_FILE and saves the resulting OUT_FILE.

Tesseract box file columns:
<symbol> <left> <bottom> <right> <top> <page>
"""

import pandas as pd
from PIL import Image, ImageDraw, ImageFont
from csv import reader

FONT = '/usr/share/fonts/truetype/dejavu/DejaVuSansMono.ttf'
FONT_SIZE = 26
TIFF_FILE = 'val/val_invoice.tiff'
BOX_FILE = 'val/val_invoice.box'
OUT_FILE = 'val/val_invoice_ocr_deu.tiff'


def draw_tesseract_box(image_path, box_path):
    """Draw a box file coordinates as rectangles on an image. """
    image = Image.open(image_path)
    image_height = image.height
    df_text = get_tesseract_coords(box_path)
    for _, row in df_text.iterrows():
        symbol = row['symbol'].replace('\t', '')
        pillow_coords = convert_coords_to_pillow(row, image_height)
        image = draw_bb_text(image, pillow_coords, symbol)
    image.save(OUT_FILE)


def get_tesseract_coords(box_path):
    """Get the Tesseract text and bb coordinates. """
    csv_reader = reader(open(box_path), delimiter=' ')
    names = ['symbol', 'left', 'bottom', 'right', 'top', 'page']
    box_data = []
    for row in csv_reader:
        # tesseract uses two empty strings as space
        if row[0] == '' and row[1] == '':
            row[0] = ' '
            del row[1]
        box_data.append(row)
    df = pd.DataFrame.from_records(box_data, columns=names)
    df_text = df.groupby([
        'left', 'bottom', 'right', 'top', 'page'])['symbol'].apply(
            lambda x: ''.join(x.values.tolist())
        ).reset_index(drop=False)
    return df_text


def draw_bb_text(image, coords, text):
    """Draw a bb and write text from tesseract box file data. """
    draw = ImageDraw.Draw(image)
    shape = [
        (coords['left'], coords['top']),
        (coords['right'], coords['bottom'])
    ]
    draw.rectangle(xy=shape, outline="green")
    unicode_font = ImageFont.truetype(FONT, FONT_SIZE)
    text_coord = (
        coords['left'],
        coords['top'] + (coords['top'] - coords['bottom'])
    )
    draw.text(text_coord, text, font=unicode_font, fill="red")
    print(text)
    return image


def convert_coords_to_pillow(coords_tess, image_height):
    """Convert coords from Tesseract boxfile coords to pillow coords. """
    coords_json = {
        'left': int(coords_tess['left']),
        'top': image_height - int(coords_tess['top']),
        'right': int(coords_tess['right']),
        'bottom': image_height - int(coords_tess['bottom'])
    }
    return coords_json


def main():
    box_path = BOX_FILE
    image_path = TIFF_FILE
    draw_tesseract_box(image_path, box_path)


if __name__ == '__main__':
    main()

fine_tune.sh

/usr/local/bin/lstmtraining \
 --model_output output/fine_tuned \
 --continue_from lstm_model/deu.lstm \
 --traineddata tesseract/tessdata/best/deu.traineddata \
 --train_listfile train/deu.training_files.txt \
 --eval_listfile eval/deu.training_files.txt \
 --max_iterations 400