Mini-Vision-V3: EMNIST Balanced Handwritten Character Classifier

Welcome to Mini-Vision-V3, the third model in the Mini-Vision series. Following the MNIST digit recognition task in V2, this model expands capabilities to 47 classes of handwritten characters (Digits & Uppercase & Lowercase letters) using the EMNIST Balanced dataset. It features a deeper yet highly efficient 3-layer CNN architecture, achieving over 90% accuracy with less than half a million parameters.

Model Description

Mini-Vision-V3 is a custom 3-layer CNN architecture tailored for 28x28 grayscale images. While maintaining a lightweight footprint with only 0.40M parameters (half the size of V2), it handles the significantly increased complexity of 47 character classes. This project demonstrates how depth and Batch Normalization can improve performance on more complex classification tasks without increasing model size.

• Dataset: EMNIST Balanced (28x28 grayscale images, 47 classes)
• Framework: PyTorch
• Total Parameters: 0.40M

Model Architecture

The network utilizes a deeper structure compared to V2, featuring three convolutional blocks. This allows for better feature extraction in the more complex 47-class task.

Layer	Input Channels	Output Channels	Kernel Size	Stride	Padding	Activation	Other
Conv Block 1	1	32	3	1	1	ReLU	MaxPool(2), BatchNorm
Conv Block 2	32	64	3	1	1	ReLU	MaxPool(2), BatchNorm
Conv Block 3	64	128	3	1	1	ReLU	MaxPool(2), BatchNorm
Flatten	-	-	-	-	-	-	Output: 1152
Linear 1	1152	256	-	-	-	ReLU	Dropout(0.3)
Linear 2	256	47	-	-	-	-	-

Training Strategy

The training strategy was adjusted for the larger dataset and increased class complexity, utilizing a higher initial learning rate and a StepLR scheduler for convergence.

• Optimizer: SGD (Momentum=0.8)
• Initial Learning Rate: 0.05
• Scheduler: StepLR (Step size=5, Gamma=0.5)
• Loss Function: CrossEntropyLoss
• Batch Size: 256
• Epochs: 50 (Best model at Epoch 40)
• Data Preprocessing:
  • EMNIST specific alignment: Rotate -90 degrees and Flip Horizontal (to match standard image orientation).
  • Random Crop (28x28 with padding=2)
  • Random Rotation (10 degrees)

Performance

The model achieved solid results on the EMNIST Balanced test set (18800 samples), selected based on the best performing epoch (Epoch 40):

Metric	Value
Test Accuracy	90.06%
Test Loss	0.28
Train Loss	0.28
Parameters	0.40M

Training Visualization (TensorBoard)

Below are the training and testing curves visualized via TensorBoard.

1. Training Loss

(Recorded every epoch)

2. Test Loss & Accuracy

(Recorded every epoch)

Quick Start

Dependencies

• Python 3.x
• PyTorch
• Torchvision
• Gradio (for demo)
• Pillow

Inference / Web Demo

Run the Gradio demo to draw characters and see predictions in real-time:

python demo.py

Note: The demo supports inverted drawing (white ink on black background) to match the EMNIST format.

File Structure

.
├── model.py                    # Model architecture definition (MiniVisionV3)
├── train.py                    # Training script
├── demo.py                     # Gradio Web Interface
├── Mini-Vision-V3.pth          # Trained model weights (Epoch 40)
├── Mini-Vision-V3.safetensors  # Safetensors format model weights
├── config.json
├── README.md
└── assets
      ├── train_loss.png   # Visualized train loss graph
      └── test_loss.png    # Visualized test loss graph

License

This project is licensed under the MIT License.