STM32F429 Online handwritten character recognition

This project demonstrates Neural Networks inference in MCUs with constrained resources (memory and CPU frequency).

Binaries

See Releases. To upload an ELF file, use STM32 CubeProgrammer.

Task

Online handwritten character classification.

Dataset

UJIpenchars2

Model

Gated Recurrent Unit.

• input shape: (30, 2) - 30 points of X and Y from a touch screen;
• GRU hidden unit shape: (100,);
• output shape: (26,) - 26 English letters.

from keras import layers, models

gru = models.Sequential()
gru.add(layers.GRU(units=100, activation='tanh', recurrent_activation='hard_sigmoid',
                   return_sequences=False, implementation=1,
                   input_shape=(30, 2)))
gru.add(layers.Dense(units=26, activation='softmax'))

Model training script: gru.py

Pretrained model dir: models

X-CUBE-AI

X-CUBE-AI

Usage: GRU inference

Hardware

STM32F429 Discovery board.

• Flash: 2048 Kb
• RAM: 256 Kb
• CPU: 180 MHz

YouTube demo