Mini-Digits

A comprehensive C++ implementation of Convolutional Neural Networks for MNIST digit classification, featuring both LeNet-5 and VGG16 architectures with a modern Qt-based graphical interface.

Features

Neural Network Core

• LeNet-5 Implementation: Custom C++ implementation achieving >99.8% accuracy on MNIST
• VGG16 Architecture: Complete VGG16 implementation (provided as reference)
• From-Scratch Implementation: Custom CNN layers (convolutional, pooling, fully connected)
• Multiple Activation Functions: ReLU, Leaky ReLU, Sigmoid, Tanh
• Advanced Initialization: Gaussian, Xavier, He initialization methods

User Interface

• Interactive Canvas: Draw digits with customizable brush colors
• Real-time Recognition: Instant digit detection with bounding boxes
• Image Import: Support for PNG, JPG, JPEG, BMP formats
• Undo/Redo System: 16-step history for canvas operations
• Training Monitor: Real-time progress bars and accuracy tracking

Model Management

• Binary Model Format: Custom .cnn format with magic number validation
• Model Persistence: Save and load trained models
• Training Configuration: Adjustable batch size, epochs, learning rates
• Multiple Training Modes: Simple and detailed training with progress monitoring

Quick Start Guide

1. Setup Model

• Choose Model Type: LeNet-5 (recommended) or VGG16
• Create New Model: File → New Model or load existing .cnn file
• Configure training parameters in the Model panel

2. Prepare Data

• Set paths to MNIST dataset files:
• Train Data: train-images.idx3-ubyte
• Train Labels: train-labels.idx1-ubyte
• Test Data: t10k-images.idx3-ubyte
• Test Labels: t10k-labels.idx1-ubyte

3. Train Model

• Click Start Training or use Train menu
• Monitor real-time progress with accuracy and loss
• Stop training anytime with Stop Training

4. Recognize Digits

• Draw on canvas or Import image file
• Click Recognize to detect and label digits
• Export recognized digits for further analysis

Configuration

The application automatically saves settings to config.ini

Development

Key Technical Features

• Custom CNN Implementation: All layers implemented from scratch
• Efficient Linear Algebra: Eigen3 for matrix operations
• Image Processing: OpenCV for digit extraction and preprocessing
• Multi-threading: OpenMP support for parallel computation
• Binary Serialization: Custom format for model persistence

File Formats

• Models: .cnn (custom binary format with magic number 1128)
• Images: PNG, JPG, JPEG, BMP
• MNIST Data: IDX3/IDX1 binary formats

License

MIT License; This project was developed as part of Sichuan University C++ coursework, demonstrating complete CNN implementation from scratch.