InkNet-Mini: Personalized Handwritten Digit Recognition from Minimal Data

InkNet-Mini is a lightweight Convolutional Neural Network (CNN) experiment designed to learn and recognize a single individual's handwriting style using only 100 base images (10 per digit).

While standard digit recognition models rely on massive datasets like MNIST (60,000+ images), this project demonstrates how deep learning can be adapted for highly personalized, low-data scenarios. By leveraging aggressive data augmentation and a constrained network architecture, InkNet-Mini learns the unique quirks of your handwriting without needing thousands of examples.

✨ Features

• Minimal Data Footprint: Requires only 10 original images per digit (100 total).
• Automated Data Augmentation: Generates additional unique variations per image (rotation, shifting, zooming, noise) to prevent overfitting.
• Custom CNN Architecture: Optimized for small datasets with strategic dropout layers.
• Multi-Digit Inference: Can process full phone numbers or postal codes, not just isolated digits.
• Modular Pipeline: Clean separation between data augmentation, model training, and inference.

📂 Project Structure

InkNet-Mini/
├── augment_data.py          # Script to generate augmented training images
├── train_custom.py          # Script to train the CNN on the enhanced dataset
├── predict.py               # Inference script for single/multi-digit images
├── requirements.txt         # Python dependencies
├── custom_data/             # Your dataset directory
│   ├── 0/                   # Contains 00.png to 09.png (and generated 10.png-209.png)
│   ├── 1/
│   └── ...
└── my_handwriting_model.keras  # The trained model (generated after training)

🛠️ Prerequisites

• OS: Ubuntu 24.04
• Python: 3.13 or higher
• Hardware: CPU is sufficient (GPU optional but faster)

🚀 Installation

1. Clone the repository:

   git clone https://github.com/IvanDeus/InkNet-Mini.git && cd InkNet-Mini

2. Create and activate a virtual environment:

   python3 -m venv venv
   source venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

📖 Usage Guide

Step 1: Prepare Your Base Dataset

Create the custom_data directory structure and add exactly 10 images for each digit (0-9).

• Name them 00.png through 09.png inside their respective folders.
• Tip: Write the digits slightly differently each time (vary the slant, size, and pen pressure) to give the model more natural variance to learn from.

mkdir -p custom_data/{0,1,2,3,4,5,6,7,8,9}
# Add your 100 images (10 per folder) here.

Step 2: Augment the Data

Run the augmentation script to expand your 100 images into 3,100 images. This prevents the neural network from simply memorizing your original 100 pictures.

python augment_data.py

You will now see 10.png through 309.png in each folder. Open a few to verify they look like slightly altered versions of your handwriting. Original images will be reformatted.

Step 3: Train the Model

Train the CNN on your newly expanded dataset. The script will automatically split the data into 80% training and 20% testing.

python train_custom.py

Watch the console output. You want to see both accuracy (training) and val_accuracy (testing) climb together. This will generate my_handwriting_model.keras.

Step 4: Predict / Inference

Use the prediction script to test the model on brand new images of your handwriting.

Recognize a single digit:

python predict.py path/to/single_digit.png

Expected output:

Recognized: 4
Confidence: 99.58%

Recognize a sequence (like a phone number or postal code):

python predict.py path/to/phone_number.png --multi

Expected output:

Recognized: 093205417
Confidences: ['99.99%', '100.00%', '97.88%', '91.39%', '100.00%', '99.97%', '100.00%', '99.95%', '99.99%']

🧠 How It Works

1. The Overfitting Problem: If you train a deep learning model on only 10 images per class, it will achieve 100% training accuracy but fail completely on new data. It memorizes the exact pixels rather than learning the concept of the digit.
2. The Augmentation Solution: augment_data.py applies random affine transformations (rotations up to 15°, shifts, zooms, blur, and noise). This forces the model to learn the structural features of your handwriting (loops, lines, intersections) rather than pixel-perfect memorization.
3. The Architecture: The CNN uses Dropout(0.5) in its final dense layer. During training, this randomly turns off 50% of the neurons, further preventing the network from relying on any single memorized feature.

💡 Tips for Best Results

• Increase Accuracy: If you decide you want higher accuracy for a practical application, do not increase augmented images, but start increasing your original images.
• Contrast is King: Ensure your original images have high contrast (black ink on white paper). The scripts automatically invert colors if needed, but clean source images yield the best results.
• Centering: Try to draw the digits roughly in the center of the image frame.
• Isolate Digits for Multi-Prediction: When using the --multi flag, ensure there is clear, empty space between each digit in the source image so the contour detection can separate them accurately.

---

2026 [ ivan deus ]