Arduino Uno Based Handwriting Recognition System Using a Feed-Forward Neural Network

Human-machine interaction (HCI) plays important roles for the development of assistive devices. Smooth communication between humans and machines (or devices) can improve quality of life. This study focuses on the application of machine learning within an embedded system for handwriting recognition. A 10×16 pixel LCD was used to display handwriting from 2D inputs (a joystick). 18,000 free and public domain fonts were used to generate a dataset of digits. A model was developed as a simple densely connected feed-forward neural network with a single hidden layer of 16 nodes defined and trained with Keras. The results showed some significant overfitting with diminishing returns around 30 epochs with a batch size of 64. Increasing the batch size to 256 and training for only 30 epochs yielded the same validation results in much less time. After training, the model was stored in flash memory on an Arduino Uno. Inference was shown to be perceptually instant on-device, while retaining compatibility with the online simulation environment.