如何基于已生成的box文件训练Tesseract 3.04（适配pytesseract）？

Got it, let's walk through the exact next steps to turn your PNG/box file pairs into a custom Tesseract model that works seamlessly with pytesseract. I’ll break this down into actionable, easy-to-follow steps tailored for v3.04 (since it has different training workflows than newer Tesseract versions):

1. Organize Your Training Files First #

Start by setting up a dedicated training folder (e.g., custom_ocr_train) to keep everything tidy:

• Drop all your PNG images and matching .box files here—make sure each image and box file share the exact same filename (like invoice01.png and invoice01.box).
• Create a font_properties file in this folder. This tells Tesseract about the font characteristics of your training data. The format is simple:

  my_custom_font 0 0 0 0 0
  

  Replace my_custom_font with a name for your font (you’ll use this name later). The five 0s represent italic, bold, fixed, serif, fraktur—set them to 1 only if your font has those attributes, otherwise leave as 0.

2. Generate Training Text (.tr) Files #

For each PNG/box pair, run this command to generate a .tr file (Tesseract’s core training input format):

tesseract sample_image.png sample_image nobatch box.train

If you have dozens of files, automate this with a shell script:

for img in *.png; do
    base=$(basename "$img" .png)
    tesseract "$img" "$base" nobatch box.train
done

3. Create the Unicharset #

The unicharset is a list of all unique characters in your training data. Generate it with:

unicharset_extractor *.box

This will create a file named unicharset in your training folder.

4. Generate Intermediate Training Files #

Next, we need to build four critical files that Tesseract uses to learn character shapes and patterns. Run these commands in order, replacing my_custom_font with the name you used in font_properties:

# Cluster character shapes
shapeclustering -F font_properties -U unicharset *.tr

# Build font-specific unicharset and character metrics
mftraining -F font_properties -U unicharset -O my_custom_font.unicharset *.tr

# Generate count-based training data
cntraining *.tr

After running these, you’ll have four files: inttemp, pffmtable, shapetable, and normproto.

5. Rename Intermediate Files #

Rename each of these four files to include your font name—this is required for the final step:

mv inttemp my_custom_font.inttemp
mv pffmtable my_custom_font.pffmtable
mv shapetable my_custom_font.shapetable
mv normproto my_custom_font.normproto

6. Build the Final Traineddata File #

Combine all the renamed files into a single .traineddata file—this is the model pytesseract will use:

combine_tessdata my_custom_font.

Don’t forget the trailing dot! This tells combine_tessdata to look for all files prefixed with my_custom_font.. You’ll see a confirmation message if the file is created successfully (e.g., my_custom_font.traineddata).

7. Use Your Custom Model in Pytesseract #

To use your new model, just specify the lang parameter with your font name when calling pytesseract:

import pytesseract
from PIL import Image

test_img = Image.open("your_test_image.png")
extracted_text = pytesseract.image_to_string(test_img, lang="my_custom_font")
print(extracted_text)

If you want to use your custom model alongside Tesseract’s default English model, separate the language codes with a plus sign: lang="eng+my_custom_font".

Quick Pro Tips #

• Ensure all Tesseract training tools (unicharset_extractor, shapeclustering, etc.) are in your system’s PATH—these usually come with a separate "training" package (e.g., tesseract-ocr-training on Debian/Ubuntu).
• For better accuracy, use 100+ training samples covering all characters you need to recognize.
• Double-check your box files for alignment errors—even one mispositioned character can hurt training results. Tools like jTessBoxEditor can help visualize and edit box files easily.

内容的提问来源于stack exchange，提问作者user98651