Chinese Digit Classifier

A deep learning project that classifies handwritten Chinese numerals (0-9, 10, 100, 1000, 10000, 1000000) using VGG16 and LeNet architectures, achieving ~97% accuracy on the Chinese MNIST dataset.

---

Demo

Live Demo: chinese.quiet98k.com

---

Features

• Multi-Model Architecture — VGG16 and LeNet models optimized for handwritten character recognition
• Comprehensive EDA — Edge detection, PCA visualization, and intra-class consistency analysis
• Data Augmentation Pipeline — Random rotation (±30°) for robust training
• Hyperparameter Tuning — Grid search over learning rates, batch sizes, dropout, and optimizers

---

Run the Demo

git clone https://github.com/quiet98k/Chinese-Digit-Classifier.git
cd Chinese-Digit-Classifier
docker-compose up --build

Open http://localhost:3788 in your browser.

---

Model & Training

Dataset

• Chinese MNIST — 15,000 handwritten samples from 100 volunteers
• 15 classes: digits 0-9 plus 十 (10), 百 (100), 千 (1000), 万 (10000), 亿 (1000000)
• 64×64 grayscale images

Architectures

Model	Parameters	Description
VGG16	~15M	Deep CNN with batch normalization and dropout
LeNet	~62K	Lightweight model for faster inference

Training Pipeline

• Data Augmentation: Random rotation (±30°), edge detection preprocessing
• Optimization: SGD/Adam with learning rate scheduling, L2 regularization
• Hyperparameter Tuning: Grid search over dropout rates, batch sizes, and epochs

Evaluation

• VGG16 Accuracy: ~97% on test set (best configuration: SGD, lr=0.01, dropout=0.5)
• LeNet Accuracy: ~92% on test set

---

Project Structure

Chinese-Digit-Classifier/
├── minst.ipynb             # Training notebook with EDA & model training
├── models/                 # Pre-trained model weights
│   ├── vgg_16_model.pth
│   ├── lenet_model.pth
│   └── best_ResNet_weights.pth
├── best_model.pth          # Best performing model checkpoint
└── README.md

---

License

This project is licensed under the terms specified in the LICENSE file.