Tesseract OCR Quick Reference

Overview

Tesseract is an open-source Optical Character Recognition (OCR) engine developed by Google. It supports 100+ languages and can extract text from images, PDFs, and scanned documents with high accuracy.

Key Features: - Multi-language support (Latin, Cyrillic, Arabic, Asian scripts) - Multiple output formats (txt, pdf, hocr, tsv, alto) - Page segmentation modes for different layouts - Training support for custom fonts/languages - Command-line and library API (C++, Python, Java)

Version: This guide covers Tesseract 4.x and 5.x (LSTM-based)

Installation

Linux (Fedora/RHEL)

# Install Tesseract engine
sudo dnf install tesseract

# Install language packs
sudo dnf install tesseract-langpack-eng # English
sudo dnf install tesseract-langpack-hun # Hungarian
sudo dnf install tesseract-langpack-deu # German
sudo dnf install tesseract-langpack-spa # Spanish

# List all available language packs
dnf search tesseract-langpack

Ubuntu/Debian

sudo apt update
sudo apt install tesseract-ocr
sudo apt install tesseract-ocr-eng tesseract-ocr-hun

macOS

brew install tesseract
brew install tesseract-lang # All languages

Verify Installation

tesseract --version
tesseract --list-langs # Show installed languages

Basic Usage

Simple Text Extraction

# Extract text from image to stdout
tesseract image.png stdout

# Save to text file
tesseract image.png output    # Creates output.txt
tesseract scan.jpg result -l eng # Specify English language

# Multi-page processing
tesseract page1.png output1
tesseract page2.png output2

Language Selection

# Single language
tesseract document.png output -l hun # Hungarian

# Multiple languages (combined)
tesseract mixed.png output -l eng+hun # English + Hungarian
tesseract arabic.png output -l ara+eng

# Script-based detection
tesseract chinese.png output -l chi_sim # Simplified Chinese
tesseract chinese.png output -l chi_tra # Traditional Chinese

Output Formats

Text Output (Default)

tesseract image.png output     # Creates output.txt
tesseract image.png stdout     # Print to console
tesseract image.png - > file.txt  # Redirect to file

PDF Output (Searchable)

# Create searchable PDF
tesseract scan.jpg output pdf          # Creates output.pdf
tesseract scan.jpg output -l eng pdf      # With language
tesseract image.png output pdf txt       # PDF + TXT both

# Batch convert images to single PDF
tesseract page*.png output pdf # Warning: May not combine pages properly

HOCR (HTML-based OCR format)

# Preserves layout, coordinates, confidence scores
tesseract document.png output hocr       # Creates output.hocr
tesseract scan.jpg output -l eng hocr

TSV (Tab-Separated Values)

# Detailed output with bounding boxes and confidence
tesseract image.png output tsv         # Creates output.tsv

# Columns: level, page_num, block_num, par_num, line_num, word_num, 
#     left, top, width, height, conf, text

ALTO XML (Library Standard)

tesseract document.png output alto       # Creates output.xml

Combined Outputs

# Generate multiple formats at once
tesseract scan.jpg output txt pdf hocr
tesseract image.png output txt tsv alto

Page Segmentation Modes (PSM)

PSM controls how Tesseract detects text layout. Critical for accuracy.

# Syntax
tesseract image.png output --psm <mode>

PSM Modes

Mode	Description	Use Case
0	Orientation and script detection only	Detect document rotation
1	Automatic page segmentation with OSD	General documents
2	Automatic page segmentation (no OSD)	Skewed documents
3	Fully automatic (default)	Most cases
4	Single column of text	Book pages, articles
5	Single vertical block	Asian text
6	Single uniform block	Clean text blocks
7	Single text line	Cropped lines, captions
8	Single word	License plates, labels
9	Single word in circle	Logos, stamps
10	Single character	Character recognition
11	Sparse text (find as much as possible)	Noisy images
12	Sparse text with OSD	Rotated noisy images
13	Raw line (bypass segmentation)	Pre-processed lines

PSM Examples

# Single line of text (e.g., invoice number)
tesseract line.png output --psm 7

# License plate (single word)
tesseract plate.jpg output --psm 8 -l eng

# Handwritten note (sparse text)
tesseract note.jpg output --psm 11

# Book page (single column)
tesseract page.png output --psm 4 -l eng

# Detect rotation angle
tesseract rotated.png output --psm 0

OCR Engine Modes (OEM)

OEM controls the OCR engine used.

# Syntax
tesseract image.png output --oem <mode>

OEM Modes

Mode	Engine	Description
0	Legacy	Tesseract 3.x (older, faster, less accurate)
1	LSTM	Neural network-based (default, most accurate)
2	Legacy + LSTM	Combined engines
3	Default	Automatically choose best

# Use LSTM engine (default in v4+)
tesseract image.png output --oem 1

# Legacy engine (faster, lower accuracy)
tesseract image.png output --oem 0

# Combined engines
tesseract image.png output --oem 2

Practical Use Cases

1. Scan Receipt/Invoice

# Preprocess + OCR + TSV for structured data
convert receipt.jpg -resize 200% -sharpen 0x1 receipt_clean.png
tesseract receipt_clean.png output --psm 6 tsv

# Extract specific fields (e.g., total amount)
tesseract receipt.jpg stdout --psm 6 | grep -i "total"

2. Extract Text from PDF

# Convert PDF to images first
pdftoppm document.pdf page -png

# OCR each page
for img in page-*.png; do
tesseract "$img" "${img%.png}" -l eng pdf
done

# Combine PDFs (using pdfunite or similar)
pdfunite page-*.pdf combined_ocr.pdf

3. Batch Process Scanned Books

#!/bin/bash
# batch_ocr.sh

INPUT_DIR="scans"
OUTPUT_DIR="ocr_output"
LANG="eng"

mkdir -p "$OUTPUT_DIR"

for img in "$INPUT_DIR"/*.jpg; do
base=$(basename "$img" .jpg)
echo "Processing $base..."
tesseract "$img" "$OUTPUT_DIR/$base" -l "$LANG" --psm 4 pdf txt
done

echo "OCR complete. Output in $OUTPUT_DIR/"

4. Extract Table Data

# Use TSV output for tables
tesseract table.png output tsv

# Parse TSV with awk/Python to extract rows/columns
awk -F'\t' '$1==5 {print $12}' output.tsv # Extract word-level text

5. Process Screenshots with Code

# Code screenshots often need specific PSM
tesseract code_snippet.png output --psm 6 -l eng

# Preserve whitespace (important for Python/YAML)
tesseract code.png stdout --psm 6 | cat -A # Show whitespace

6. Handwritten Notes (Low Accuracy)

# Handwriting requires sparse text mode
tesseract handwritten.jpg output --psm 11 -l eng

# Better: Train custom model or use Google Vision API
# Tesseract is NOT optimized for handwriting

7. Multi-Language Documents

# English + Hungarian document
tesseract bilingual.png output -l eng+hun

# Arabic + English
tesseract arabic_doc.png output -l ara+eng --psm 1

8. Extract Text from Video Frames

# Extract frames from video
ffmpeg -i video.mp4 -vf fps=1 frame_%04d.png

# OCR all frames
for frame in frame_*.png; do
tesseract "$frame" "${frame%.png}" -l eng txt
done

9. Verify Text in Images (Testing)

# CI/CD verification: Check if image contains expected text
EXPECTED="Error 404"
RESULT=$(tesseract screenshot.png stdout)

if echo "$RESULT" | grep -q "$EXPECTED"; then
echo "Test passed: Text found"
else
echo "Test failed: Expected text not found"
exit 1
fi

10. Create Searchable PDF Archive

#!/bin/bash
# convert_archive.sh

for doc in archive/*.jpg; do
base=$(basename "$doc" .jpg)
tesseract "$doc" "searchable_pdfs/$base" -l eng pdf
done

# Merge all PDFs
cd searchable_pdfs
pdfunite *.pdf ../complete_archive.pdf

Image Preprocessing

OCR accuracy depends heavily on image quality. Preprocessing is critical.

Using ImageMagick

# Install ImageMagick
sudo dnf install ImageMagick

# Resize (larger = better OCR, up to a point)
convert input.jpg -resize 200% output.png

# Sharpen
convert input.jpg -sharpen 0x1 output.png

# Increase contrast
convert input.jpg -contrast -contrast output.png

# Convert to grayscale
convert input.jpg -colorspace Gray output.png

# Denoise
convert input.jpg -despeckle output.png

# Threshold (binarize)
convert input.jpg -threshold 50% output.png

# Full preprocessing pipeline
convert input.jpg \
-resize 200% \
-colorspace Gray \
-sharpen 0x1 \
-contrast \
-threshold 50% \
preprocessed.png

tesseract preprocessed.png output -l eng

Deskew (Fix Rotation)

# Auto-deskew with ImageMagick
convert input.jpg -deskew 40% deskewed.jpg

# Or use tesseract's built-in OSD
tesseract input.jpg output --psm 0 # Detect rotation
# Then manually rotate:
convert input.jpg -rotate 90 rotated.jpg

Remove Borders/Noise

# Trim white borders
convert input.jpg -trim output.jpg

# Remove salt-and-pepper noise
convert input.jpg -median 2 output.jpg

Configuration Files

Custom Configuration

# Use config file
tesseract image.png output -c tessedit_char_whitelist=0123456789

# Create custom config file (e.g., digits_only)
echo "tessedit_char_whitelist 0123456789" > digits.conf
tesseract image.png output digits

Common Configurations

# Digits only (useful for phone numbers, IDs)
tesseract image.png output -c tessedit_char_whitelist=0123456789

# Alphanumeric only
tesseract image.png output -c tessedit_char_whitelist=ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789

# Preserve whitespace
tesseract image.png output -c preserve_interword_spaces=1

# Min confidence threshold (reject low-confidence chars)
tesseract image.png output -c tessedit_char_min_confidence=80

Advanced Examples

Extract Metadata from TSV

# Get word-level confidence scores
tesseract image.png output tsv

# Parse TSV to find low-confidence words
awk -F'\t' '$11 < 80 && $12 != "" {print $12, $11}' output.tsv

# Average confidence score
awk -F'\t' '$11 > 0 {sum+=$11; count++} END {print sum/count}' output.tsv

Parallel Processing

# Process 100 images in parallel (GNU parallel)
find images/ -name "*.jpg" | \
parallel -j 8 tesseract {} {.} -l eng pdf

# Using xargs
ls images/*.png | \
xargs -n 1 -P 4 -I {} tesseract {} {}.txt -l eng

Integration with Python

#!/usr/bin/env python3
import pytesseract
from PIL import Image

# Simple OCR
text = pytesseract.image_to_string(Image.open('image.png'), lang='eng')
print(text)

# Get detailed data
data = pytesseract.image_to_data(Image.open('image.png'), output_type='dict')
print(f"Confidence: {data['conf']}")

# Custom config
custom_config = r'--psm 6 --oem 1'
text = pytesseract.image_to_string(Image.open('image.png'), config=custom_config)

OCR + Spellcheck

# Extract text and fix typos with aspell
tesseract image.png stdout -l eng | aspell -a

# Or hunspell
tesseract image.png stdout -l eng | hunspell -l

Troubleshooting

Problem: Low Accuracy

Solutions: 1. Preprocess image (resize, sharpen, denoise) 2. Try different PSM modes (--psm 4, --psm 6, --psm 11) 3. Verify correct language (-l eng, -l hun) 4. Check image DPI (300+ DPI recommended) 5. Binarize image (convert to black/white)

# Debug: Save preprocessed image and test
convert input.jpg -resize 300% -sharpen 0x1 -threshold 50% debug.png
tesseract debug.png output --psm 6 -l eng

Problem: Missing Text

Causes: - Wrong PSM mode (use --psm 11 for sparse text) - Language not installed (tesseract --list-langs) - Text too small (resize image to 200-300%)

# Check installed languages
tesseract --list-langs

# Install missing language
sudo dnf install tesseract-langpack-<lang>

Problem: Garbled Output

Causes: - Wrong character encoding - Mixed languages (use -l eng+hun) - Special characters not in language model

# Try combined languages
tesseract image.png output -l eng+hun+deu

# Check encoding
file output.txt
iconv -f ISO-8859-1 -t UTF-8 output.txt -o fixed.txt

Problem: Slow Processing

Solutions: 1. Use legacy engine (--oem 0) 2. Reduce image resolution 3. Use faster PSM mode (--psm 6 vs --psm 3) 4. Parallel processing (GNU parallel)

# Faster processing (lower accuracy)
tesseract image.png output --oem 0 --psm 6

Problem: PDF Input Not Working

Solution: Convert PDF to images first

# Install poppler-utils
sudo dnf install poppler-utils

# Convert PDF to images
pdftoppm input.pdf page -png -r 300 # 300 DPI

# OCR each page
for img in page-*.png; do
tesseract "$img" "${img%.png}" -l eng pdf
done

Performance Tips

Optimal Image Settings

DPI: 300 DPI (minimum 150, maximum 600)
Format: PNG or TIFF (lossless)
Color: Grayscale (color not needed for OCR)
Size: Resize to 2000-3000px width for normal text

Speed vs Accuracy Trade-offs

# Fastest (lowest accuracy)
tesseract img.png out --oem 0 --psm 6

# Balanced (good for production)
tesseract img.png out --oem 1 --psm 3 # Default

# Most accurate (slowest)
tesseract img.png out --oem 2 --psm 1 -c tessedit_char_min_confidence=90

Batch Processing Best Practices

# Use tmpfs for temp files (faster I/O)
mkdir -p /tmp/ocr_temp
for img in *.jpg; do
cp "$img" /tmp/ocr_temp/
tesseract /tmp/ocr_temp/"$img" output/"${img%.jpg}" -l eng
rm /tmp/ocr_temp/"$img"
done

Integration Examples

Shell Script: Automated Invoice Processing

#!/bin/bash
# process_invoices.sh

INVOICE_DIR="invoices"
OUTPUT_DIR="extracted_data"
LANG="eng"

mkdir -p "$OUTPUT_DIR"

for invoice in "$INVOICE_DIR"/*.pdf; do
base=$(basename "$invoice" .pdf)

# Convert PDF to image
pdftoppm -png -r 300 "$invoice" "$OUTPUT_DIR/$base"

# Preprocess
convert "$OUTPUT_DIR/${base}-1.png" \
-resize 200% \
-colorspace Gray \
-sharpen 0x1 \
"$OUTPUT_DIR/${base}_clean.png"

# OCR with structured output
tesseract "$OUTPUT_DIR/${base}_clean.png" \
"$OUTPUT_DIR/$base" \
-l "$LANG" \
--psm 6 \
tsv txt

# Extract specific fields (example: invoice number)
grep -i "invoice" "$OUTPUT_DIR/$base.txt" | head -1

# Cleanup temp files
rm "$OUTPUT_DIR/${base}-1.png" "$OUTPUT_DIR/${base}_clean.png"
done

echo "Processing complete."

Python: Extract Text with Confidence Scores

#!/usr/bin/env python3
import pytesseract
from PIL import Image
import pandas as pd

def ocr_with_confidence(image_path):
"""Extract text with confidence scores"""
img = Image.open(image_path)

# Get detailed data
data = pytesseract.image_to_data(img, output_type='dict')

# Filter low-confidence words
df = pd.DataFrame(data)
df = df[df['conf'] > 60] # Keep only >60% confidence

# Group by line
lines = df.groupby(['block_num', 'par_num', 'line_num'])['text'].apply(' '.join)

return '\n'.join(lines)

# Usage
text = ocr_with_confidence('document.png')
print(text)

Bash: Monitor Folder for New Scans

#!/bin/bash
# watch_scans.sh - Auto-OCR new files with inotifywait

WATCH_DIR="$HOME/scans"
OUTPUT_DIR="$HOME/ocr_results"

mkdir -p "$OUTPUT_DIR"

inotifywait -m -e close_write --format '%f' "$WATCH_DIR" | \
while read filename; do
if [[ "$filename" =~ \.(jpg|png|tif)$ ]]; then
echo "Processing $filename..."
tesseract "$WATCH_DIR/$filename" \
"$OUTPUT_DIR/${filename%.*}" \
-l eng --psm 3 pdf txt
echo "Done: ${filename%.*}.pdf"
fi
done

Training Custom Models

For specialized fonts, historical documents, or custom languages.

Quick Training Overview

# 1. Generate training images with text
# 2. Create box files (bounding boxes)
# 3. Train Tesseract model

# Generate box file
tesseract training_image.png training_image lstm.train

# (Full training process is complex - see official docs)

Note: Custom training is advanced. Use pre-trained models when possible.

Alternative Tools

Tool	Use Case	Comparison
Google Cloud Vision API	High accuracy, handwriting	Paid, cloud-based, more accurate
Amazon Textract	Forms, tables, AWS integration	Paid, excellent for structured docs
EasyOCR	80+ languages, Python-native	Easier API, GPU support
PaddleOCR	Asian languages, lightweight	Fast, good for Chinese/Japanese
ABBYY FineReader	Commercial, high accuracy	Expensive, best-in-class accuracy

Useful Resources

Official Documentation: https://tesseract-ocr.github.io/
GitHub: https://github.com/tesseract-ocr/tesseract
Language Data: https://github.com/tesseract-ocr/tessdata
Training Tutorial: https://tesseract-ocr.github.io/tessdoc/Training-Tesseract.html

Quick Command Reference

# Basic OCR
tesseract image.png output

# Specify language
tesseract image.png output -l hun

# Multiple languages
tesseract image.png output -l eng+hun

# Output formats
tesseract image.png output pdf
tesseract image.png output hocr
tesseract image.png output tsv

# Page segmentation mode
tesseract image.png output --psm 6

# OCR engine mode
tesseract image.png output --oem 1

# Combined options
tesseract image.png output -l eng --psm 6 --oem 1 pdf txt

# Config options
tesseract image.png output -c tessedit_char_whitelist=0123456789

# List languages
tesseract --list-langs

# Version info
tesseract --version

License

Tesseract is released under the Apache License 2.0.

Last Updated: 2026-04-13 Author: Miklos Greczi Repository: docs-repository