🧠 OCR_AI – Lightweight Neural Network for Handwritten Digit Recognition (No ML Libraries)

OCR_AI is a self-implemented neural network written in Python and powered only by NumPy. It recognizes handwritten digits from the MNIST dataset using a 28x28 input grid. This project is completely independent of machine learning libraries like TensorFlow or PyTorch and is intended for learning, experimentation, or lightweight deployment.

💡 Inspired in part by "500 Lines or Less – Optical Character Recognition" by Marina Samuel. (Read it here)

---

📸 Example Use Case

The model processes 28x28 grayscale images (flattened into 784-dimensional vectors), runs them through a feedforward neural network, and outputs a digit prediction from 0 to 9.

🎬 Demo

Live Testing (Drawing in the Browser)

The user draws a digit on the 28x28 canvas and gets a prediction in real time using the trained model.

Training on MNIST Dataset

The model updates its weights and biases in real time based on new user-supplied training samples.

📁 Project Structure

OCR_AI/
│
├── src/
│ ├── ocr.py # Neural network logic
│ ├── server.py # Web basic server to interact with the model
│ ├── ocr.js # Input preprocessing
│ ├── ocr.html # Page Structure
│ ├── wb_training.py # Script to train the model using mnist │ └── neural_network.json # Stored model weights
│
└── README.md
└── .gitignore

---

🔧 Features

• Fully custom neural network:
  • Input layer: 784 neurons (28x28 pixels)
  • Hidden layer: customizable (default: 25 neurons)
  • Output layer: 10 neurons (digits 0–9)
• Option to train the model on MNIST data or test existing model weights
• Simple training using train_on_instance() for per-sample learning (batch: 1)
• Model weights saved as JSON (neural_network.json)
• Activation functions: ReLU (default), support for sigmoid
• No machine learning libraries used – just NumPy

---

🚀 Run the Server

python src/server.py

🙏 Credits

• MNIST dataset: http://yann.lecun.com/exdb/mnist/

• Base inspiration: 500 Lines or Less – Optical Character Recognition"* by Marina Samuel.

📝 License

This project is licensed under the MIT License.