MNIST_image_classification

Character Recognition (MNIST)

Overview

The Character Recognition application is an Edge AI solution that runs on the MSPM0G5187 microcontroller. It classifies single-digit handwritten character images (0-9) in real-time using the classic LeNet-5 convolutional neural network architecture.

Problem and Solution

• Handwritten digit recognition is a fundamental task in document processing and automation
• Manual digit entry is error-prone and time-consuming
• Traditional template matching approaches struggle with handwriting variability
• Edge AI enables robust character recognition directly on resource-constrained microcontrollers

Key Performance Targets

• Real-time digit classification
• High accuracy on handwritten digits
• Low memory footprint suitable for MCU deployment
• Simple UART-based communication with host GUI

System Components

1. Hardware:

   • MSPM0G5187 microcontroller Link
   • UART connection to host PC

2. Software:

   • Code Composer Studio 12.x or later
   • MSPM0 SDK 2.08.00 or later

Dataset Information

The example uses the mnist_image_classification dataset based on the classic MNIST handwritten digit dataset:

• Source: MNIST dataset exported via torchvision
• Classes: 10 digits (0-9)
• Image Format: 28x28 grayscale PNG images
• Download: mnist_classes.zip

For dataset recreation instructions, refer to readme_dataset_creation.md.

Feature Extraction Pipeline

1. Image Input: 28x28 grayscale image
2. Grayscale Conversion: Single channel output
3. Resize: Ensure 28x28 dimensions
4. Tensor Conversion: Convert to PyTorch tensor
5. Normalization: Mean = 0.1307, Scale = 0.3081

Model Architecture Options (Available on Tensorlab CLI Tools)

1. Lenet5 (Default):
   • Classic LeNet-5 CNN architecture
   • ~60,000 parameters
   • Proven architecture for digit recognition

Model Performance

• Accuracy: ~99%

Training and Deployment Process

NOTE: Running the config yaml handles everything including feature extraction, training, quantization, and compilation.

1. Training:

   • Use TI Edge AI Studio (GUI) or tinyml-tensorlab (CLI)
   • Batch size: 64, Learning rate: 0.1
   • Training epochs: 14

2. Quantization:

   • INT8 quantization for reduced model size
   • Maintains accuracy while enabling MCU deployment

3. Compilation:

   • TI Neural Network Compiler converts trained model
   • Generates model artifacts for CPU inference

How to Run

After completing the repository setup, run the following command from the tinyml-modelzoo directory:

Windows:

  .\run_tinyml_modelzoo.bat examples\MNIST_image_classification\config_image_classification_mnist.yaml

Linux:

  ./run_tinyml_modelzoo.sh examples/MNIST_image_classification/config_image_classification_mnist.yaml

References

• MSPM0G5187 Technical Reference Manual https://www.ti.com/product/MSPM0G5187
• https://software-dl.ti.com/mctools/nnc/mcu/users_guide/
• TI Model Training Guide: https://github.com/TexasInstruments/tinyml-tensorlab/tree/main
• http://yann.lecun.com/exdb/mnist/
• https://en.wikipedia.org/wiki/LeNet
• EdgeAI Software Guide: https://dev.ti.com/tirex/explore/node?node=A__AKCnvqDed-Plz2JO5Umb3Q__MSPM0-SDK__a3PaaoK__LATEST
• MSPM0 SDK: https://www.ti.com/tool/MSPM0-SDK