Tamil OCR Application

A powerful command-line tool for extracting Tamil text from images using advanced OCR technology with AI-powered validation and correction.

🌟 Features

Core Capabilities

• Dual OCR Engines: Choose between ocr_tamil and PaddleOCR for optimal results
• AI-Powered Validation: Qwen2.5 72B model corrects OCR errors and improves accuracy
• Multiple Preprocessing Techniques: 5 different image preprocessing methods
• Batch Processing: Process multiple images in a single command
• Confidence Assessment: Automatic quality evaluation of OCR results
• Beautiful Terminal Output: Color-coded, formatted results with clear visual hierarchy

---

OCR Engines Explained

OCR Engines Explained

1. ocr_tamil (Default) What it is: A specialized OCR library specifically designed for Tamil script recognition. Strengths:

Optimized for Tamil Unicode characters Better handling of complex Tamil ligatures Lightweight and fast processing Pre-trained on Tamil documents

Best for:

Printed Tamil text Clear, high-contrast images Modern Tamil fonts Documents and books

Technical Details:

Uses deep learning models trained on Tamil datasets Includes text detection and recognition modules Handles diacritics and conjunct characters

2. PaddleOCR What it is: An industrial-grade multilingual OCR system developed by Baidu, supporting 80+ languages including Tamil. Strengths:

Robust text detection in complex layouts Better performance on handwritten text Advanced angle correction capabilities Handles skewed or rotated text

Best for:

Complex document layouts Handwritten Tamil text Mixed Tamil-English documents Challenging lighting conditions

Technical Details:

Uses DB (Differentiable Binarization) for text detection CRNN (Convolutional Recurrent Neural Network) for recognition Built-in text angle classification Supports both GPU and CPU inference

📋 Requirements

System Requirements

• Python 3.12.8
• 4GB+ RAM recommended
• CPU-based processing (no GPU required)

Dependencies

Install all required libraries:

pip install -r requirements.txt

🚀 Installation

1. Clone or download the repository

   git clone https://github.com/100-multidisciplinary-projects/Text-Extractor
   cd Text-Extractor

2. Install dependencies

   pip install -r requirements.txt

3. Verify installation

   python app.py image_name.jpg

---

💻 Usage

Basic Usage

Process a single image:

python app.py image.jpg

Process multiple images:

python app.py img1.jpg img2.jpg img3.png

Advanced Options

Specify OCR Engine:

python app.py image.jpg --engine paddleocr

Choose Preprocessing Mode:

python app.py image.jpg --mode aggressive

Enable AI Validation:

python app.py image.jpg --api-key YOUR_OPENROUTER_KEY --ai

Custom Output Directory:

python app.py image.jpg --output ./my_results

Show Detailed Results:

python app.py image.jpg --mode auto --details

---

🔧 Command-Line Arguments

Argument	Short	Description	Default
images	-	Path(s) to input image file(s)	Required
--output	-o	Output directory for results	./ocr_output
--engine	-e	OCR engine: ocr_tamil or paddleocr	ocr_tamil
--mode	-m	Preprocessing mode (see below)	auto
--api-key	-k	OpenRouter API key for AI validation	None
--ai	-	Enable AI validation (requires API key)	Disabled
--details	-d	Show detailed preprocessing results	Disabled
--no-color	-	Disable colored terminal output	Disabled

---

🎨 Preprocessing Techniques

The application offers 5 preprocessing methods optimized for different image types:

1. Original (No Preprocessing)

• Uses the raw image as-is
• Best for: High-quality scanned documents with clear text
• When to use: Modern digital images, screenshots with crisp text

2. Standard (v1)

• Denoising + CLAHE enhancement + Adaptive thresholding
• Best for: Printed text, books, documents
• When to use: Standard quality scans, printed materials
• Techniques:
  • Fast Non-Local Means Denoising
  • CLAHE (Contrast Limited Adaptive Histogram Equalization)
  • Gaussian adaptive threshold

3. Aggressive (v2)

• Heavy denoising + Enhanced CLAHE + Morphological operations
• Best for: Handwritten text, faint text, low-quality images
• When to use: Old documents, faded text, poor lighting
• Techniques:
  • Aggressive denoising (h=30)
  • Mean adaptive threshold
  • Morphological closing to connect broken characters

4. Inverted (v3)

• Inverted threshold for dark backgrounds
• Best for: Dark backgrounds with light text (e.g., blackboards, dark themes)
• When to use: Light text on dark background, inverted images
• Techniques:
  • Binary inverse thresholding
  • Morphological opening to remove noise
  • Elliptical structuring element

5. Otsu (v4)

• Automatic threshold calculation using Otsu's method
• Best for: Images with clear contrast and bimodal histograms
• When to use: High-contrast images, clear separation between text and background
• Techniques:
  • Gaussian blur for noise reduction
  • Otsu's automatic thresholding
  • No morphological operations (preserves original structure)

Auto Mode

When using --mode auto, the application:

1. Runs all 5 preprocessing methods
2. Saves each preprocessed image
3. Uses the "Original" method for final output
4. Displays all results if --details is enabled

---

🤖 AI Validation with Qwen2.5 72B

What It Does

• Corrects OCR recognition errors
• Fixes spelling mistakes in Tamil text
• Ensures proper grammar and sentence structure
• Preserves the original meaning

How to Enable

1. Get an API Key:

   • Sign up at OpenRouter.ai
   • Get your API key from the dashboard

2. Use AI Validation:

   python app.py image.jpg --api-key YOUR_KEY --ai

3. Output:

   • Shows both OCR-extracted text and AI-corrected text
   • Saves both versions to the result file

---

📊 Confidence Assessment

The application automatically evaluates OCR quality:

• Excellent (80%+): High-quality recognition, minimal errors expected
• Good (65-79%): Reliable recognition, minor errors possible
• Fair (50-64%): Acceptable recognition, review recommended
• Poor (30-49%): Low quality, significant errors likely
• Very Poor (<30%): Very low quality, manual review required

Assessment factors:

• Text length (longer text = more reliable)
• Character quality (ratio of valid characters)
• Image clarity (Laplacian variance)

---

📁 Output Structure

For each processed image, the application creates:

ocr_output/
└── image_name/
    ├── image_name_original.png          # Original preprocessing
    ├── image_name_standard.png          # Standard preprocessing
    ├── image_name_aggressive.png        # Aggressive preprocessing
    ├── image_name_inverted.png          # Inverted preprocessing
    ├── image_name_otsu.png              # Otsu preprocessing
    └── image_name_result.txt            # Final text results

Result File Contents

• Image metadata (filename, size)
• OCR engine used
• Preprocessing method applied
• Confidence score
• Character count
• OCR extracted text (formatted)
• AI-corrected text (if AI validation was enabled)

---

📖 Examples

Example 1: Quick OCR

python app.py document.jpg

Output: Basic OCR with standard preprocessing, results in ./ocr_output/

Example 2: High-Quality Processing

python app.py old_manuscript.jpg --mode aggressive --engine paddleocr

Best for: Old documents, faded text

Example 3: AI-Enhanced Processing

python app.py handwritten.jpg --api-key sk-or-v1-xxxxx --ai --mode auto --details

Best for: Maximum accuracy with error correction

Example 4: Batch Processing

python app.py *.jpg --output ./batch_results --mode standard

Best for: Processing multiple images at once

Example 5: Dark Background Text

python app.py blackboard.jpg --mode inverted

Best for: Light text on dark backgrounds

---

🎯 Tips for Best Results

1. Image Quality:

   • Use high-resolution images (300 DPI or higher for scans)
   • Ensure good lighting and minimal shadows
   • Avoid blurry or out-of-focus images

2. Choosing OCR Engine:

   • ocr_tamil: Faster, good for standard text
   • paddleocr: More accurate for complex layouts

3. Preprocessing Selection:

   • Not sure? Use --mode auto to try all methods
   • Printed text: Use standard or original
   • Handwritten/faded: Use aggressive
   • Dark backgrounds: Use inverted
   • High contrast: Use otsu

4. AI Validation:

   • Significantly improves accuracy
   • Essential for handwritten text
   • Requires internet connection
   • Uses OpenRouter API (costs apply)

---

📞 Support

For issues, questions, or feature requests:

• Check the troubleshooting section
• Review command-line help: python app.py --help
• Refer to library documentation:
  • PaddleOCR Docs
  • OpenCV Docs
  • OpenRouter API

---

🎓 Technical Details

Image Preprocessing Pipeline

1. Resizing: Images > 2000px are scaled down
2. Denoising: Fast Non-Local Means algorithm
3. Enhancement: CLAHE for contrast improvement
4. Binarization: Adaptive or Otsu thresholding
5. Morphology: Opening/closing operations (method-dependent)

OCR Processing Flow

Input Image → Preprocessing → OCR Engine → Text Extraction
                                              ↓
                                    Sentence Formatting
                                              ↓
                                    Confidence Assessment
                                              ↓
                                    AI Validation (optional)
                                              ↓
                                    Save Results & Display

---

Happy OCR Processing! 🎉