Tesseract 4.0发票OCR字符识别错误问题及训练优化咨询

Hey Alexandra, let's tackle your Tesseract 4.0 OCR issues with invoices step by step—these character misrecognition problems are super common when dealing with specific fonts or layout quirks, but we can fix them with targeted training and preprocessing.

1. Training Tesseract 4.0 for Better Invoice OCR #

Tesseract works best when it's trained on your exact invoice fonts and character patterns. Here's a practical workflow to train it for your problematic characters:

• Collect high-quality training samples: Grab 50-100 clear, well-lit images of your invoices, focusing on sections with S, $, 8, and other misrecognized characters. Make sure the resolution is high enough to see fine details in the font.
• Edit character labels with jTessBoxEditor: Use this tool to generate .box files, which map each character in your image to its correct position and label. For every instance where S was mislabeled as $ or 8, manually correct the entry in the .box file—this tells Tesseract exactly what the character should be.
• Generate training data files: Run these Tesseract commands to convert your corrected images and .box files into training-ready .tr files:

  tesseract your_invoice_sample.tif your_training_base nobatch box.train
  

• Build language model components: Combine your custom data with Tesseract's core language files using these commands:

  unicharset_extractor your_training_base.box
  mftraining -F font_properties -U unicharset -O your_custom_lang.unicharset your_training_base.tr
  cntraining your_training_base.tr
  

• Compile and install the trained data: Rename the generated files (inttemp, normproto, pffmtable, shapetable) to add your custom language prefix (e.g., myinvoice.inttemp), then combine them into a .traineddata file with:

  combine_tessdata myinvoice.
  

  Move this .traineddata file into Tesseract's tessdata directory, and test it with:

  tesseract input_invoice.tif output_text -l myinvoice
  

• Iterate and expand: If some S still get misrecognized, add more samples of those tricky cases to your training set and repeat the process—more targeted data = better accuracy.

2. Does the Dot in S.A. Cause S to Be Misrecognized as a Number? #

Yes, it can contribute to the problem! Tesseract looks at both individual character shapes and their surrounding context. If the S in "S.A." is close to the dot, or if the scan quality is low (causing the dot to bleed into the S), Tesseract might confuse it with an 8 (since some fonts make S and 8 look similar when paired with nearby dots).

Here's how to fix this specific case:

• Preprocess images to clean up noise: Use image editing tools (or command-line tools like ImageMagick) to sharpen the image, adjust contrast, and separate characters from nearby dots. For example:

  convert input_invoice.tif -threshold 60% -sharpen 0x1.0 cleaned_invoice.tif
  

• Include "S.A." samples in your training data: Add multiple instances of "S.A." (in your invoice's font) to your training set—this teaches Tesseract to recognize S in this specific contextual pattern, even when it's next to a dot.
• Adjust Tesseract's segmentation mode: Try running Tesseract with segmentation mode 6 (which assumes a single block of uniform text) to force better character separation:

  tesseract input_invoice.tif output_text --psm 6 -l myinvoice
  

内容的提问来源于stack exchange，提问作者Alexandra