Number Recognizer

A real-time handwritten digit recognition desktop application built with Java and JavaFX, powered by a TensorFlow-trained CNN model.

Demo

Draw any digit (0–9) on the canvas, click Recognize, and get an instant prediction with confidence score.

Tech Stack

• Java / JavaFX — GUI and drawing canvas
• TensorFlow + Python — CNN model training on MNIST dataset
• ONNX Runtime — Model inference in Java
• Maven — Dependency management

How to Run

Prerequisites

• Java 23
• Maven (install via Homebrew: brew install maven)

Steps

1. Clone the repository
2. Navigate to the project folder:

   cd NumberRecognizer

3. Run the application:

   mvn javafx:run

Model & Deployment

A CNN trained on MNIST (99.34% test accuracy), served in Java:

Training (Python) → ONNX export → Inference (Java, ONNX Runtime)

The TensorFlow SavedModel couldn't run in Java on Apple Silicon due to native library incompatibility, so the model is exported to ONNX — decoupling the training stack from the serving stack.

Canvas input is preprocessed before inference: bounding-box detection → centering → 28×28 grayscale, matching MNIST's input format.

Project Structure

• Main.java — Application entry point
• MainController.java — UI logic and mouse drawing
• Predictor.java — Model inference and image preprocessing
• ModelLoader.java — Model loading from resources