Digit Recognition

This project is devoted to a classic ML problem - digit recognition.
Two models were trained on the MNIST dataset with the help of the Keras framework and deployed in a simple android app to test their live preformance. The difference between the models lies in the training data. The first model used the data from MNIST in its initial form - centered and normalized by size. The second model employed a different approach: the data was augmented - rotated and zoomed in order to simulate the inconsistency of user input.

Demo

Note: the first model, user input is accurate

Note: first vs second model in case of sloppy user input