Document Classifier

A machine learning system for classifying different types of documents using computer vision and pattern recognition techniques. The system can distinguish between comics, books, manuscripts, typewritten documents, and tickets.

Features

• Multiple Classifier Configurations: Four different classifier setups with varying preprocessing techniques
• Document Scanning: Automatic document rectification and perspective correction
• Flexible Preprocessing: Support for grayscale conversion and image transformations
• Dimensionality Reduction: Optional Linear Discriminant Analysis (LDA) for feature reduction
• Model Persistence: Save and load trained models for reuse
• Comprehensive Evaluation: Detailed performance metrics and classification reports

Document Classes

The system classifies documents into five categories:

• Comics: Comic books and graphic novels
• Libros (Books): Regular printed books
• Manuscrito (Manuscript): Handwritten documents
• Mecanografiado (Typewritten): Typewritten documents
• Tickets: Receipts, tickets, and similar small documents

Classifier Configurations

Classifier	LDA	Document Scanner	Grayscale
C1	❌	❌	❌
C2	✅	❌	❌
C3	❌	✅	✅
C4	✅	✅	✅

Requirements

pip install opencv-python numpy scikit-learn joblib argparse

Dependencies

• opencv-python: Image processing and computer vision
• numpy: Numerical computations
• scikit-learn: Machine learning algorithms
• joblib: Model serialization
• argparse: Command-line argument parsing

Installation

1. Clone or download the repository
2. Install the required dependencies
3. Ensure you have the scanner.py module (DocumentScanner class)
4. Create the following directory structure:

project/
├── data/
│   ├── Learning/
│   │   ├── comics/
│   │   ├── libros/
│   │   ├── manuscrito/
│   │   ├── mecanografiado/
│   │   └── tickets/
│   └── Test/
│       ├── comics/
│       ├── libros/
│       ├── manuscrito/
│       ├── mecanografiado/
│       └── tickets/
├── models/ (created automatically)
└── main.py

Usage

Training Mode

Train all classifier configurations:

python main.py --train

This will:

• Load training images from ./data/Learning/
• Train four different classifier configurations
• Save models to ./models/ directory

Evaluation Mode

Evaluate all trained classifiers:

python main.py --test

This will:

• Load test images from ./data/Test/
• Evaluate each classifier configuration
• Display accuracy scores and classification reports

Single Image Prediction

Classify a single image:

python main.py path/to/image.jpg

Uses the best performing classifier (C2 by default) to predict the document class.

Configuration

Key Parameters

• IMG_SIZE = (400, 300): Standard image dimensions for processing
• BEST_CLASSIFIER = 1: Index of the best performing classifier for predictions
• TRAIN_DIR = "./data/Learning": Training data directory
• TEST_DIR = "./data/Test": Test data directory

Classifier Configurations

Modify CLASSIFIER_CONFIGS to experiment with different preprocessing combinations:

CLASSIFIER_CONFIGS = {
    'C1': {'use_lda': False, 'use_scanner': False, 'to_gray': False},
    'C2': {'use_lda': True, 'use_scanner': False, 'to_gray': False},
    'C3': {'use_lda': False, 'use_scanner': True, 'to_gray': True},
    'C4': {'use_lda': True, 'use_scanner': True, 'to_gray': True}
}

Data Organization

Organize your training and test data in the following structure:

data/
├── Learning/
│   ├── comics/
│   │   ├── comic1.jpg
│   │   ├── comic2.png
│   │   └── ...
│   ├── libros/
│   │   ├── book1.jpg
│   │   └── ...
│   └── ... (other classes)
└── Test/
    ├── comics/
    │   ├── test_comic1.jpg
    │   └── ...
    └── ... (other classes)

Model Architecture

Base Classifier

• Algorithm: Support Vector Machine (SVM) with linear kernel
• Features: Raw pixel values or LDA-transformed features
• Preprocessing: StandardScaler for feature normalization

Optional Components

• Document Scanner: Automatic document rectification using perspective transformation
• Linear Discriminant Analysis: Dimensionality reduction to 4 components
• Grayscale Conversion: Color to grayscale transformation for certain configurations

Output Examples

Training Output

[TRAINING COMPLETED] Classifier C1
[TRAINING COMPLETED] Classifier C2
[TRAINING COMPLETED] Classifier C3
[TRAINING COMPLETED] Classifier C4

Evaluation Output

[EVALUATING] Classifier C1
Accuracy: 0.8542
              precision    recall  f1-score   support
...

=== RESULTS SUMMARY ===
C1: 0.8542
C2: 0.8790
C3: 0.8333
C4: 0.8611

Prediction Output

Prediction: libros

File Structure

• main.py: Main script with all functionality
• scanner.py: Document scanner module (required dependency)
• models/: Directory containing trained models
  • model_C*.joblib: Trained SVM models
  • scaler_C*.joblib: Feature scalers
  • model_C*_lda.joblib: LDA transformers (for applicable classifiers)

Performance Tips

1. Image Quality: Use high-quality, well-lit images for better results
2. Consistent Sizing: The system resizes images to 400x300, maintain aspect ratios when possible
3. Data Balance: Ensure balanced representation of all document classes in training data
4. Scanner Usage: Use document scanner for images with perspective distortion or poor alignment

Troubleshooting

Common Issues

1. "Could not load image" error: Check file path and image format compatibility
2. "Could not rectify image" error: Document scanner failed - image may be too distorted
3. Low accuracy: Ensure sufficient training data and proper class balance
4. Missing models: Run training mode before evaluation or prediction

Supported Image Formats

• JPEG (.jpg, .jpeg)
• PNG (.png)
• BMP (.bmp)
• TIFF (.tiff)

License

This project is provided as-is for educational and research purposes.

Contributing

When contributing to this project:

1. Maintain code style and documentation standards
2. Test new features with multiple classifier configurations
3. Update this README for any new functionality or configuration options