In this project I built a simple Neural Network with Tensorflow aimed at classifying handwritten digits and characters.
The Neural Network was trained on the EMNIST (Extended MNIST) Balanced Dataset and is composed of 3 convolutional layers, 2 max pooling layers and 2 dense layers with dropout regularization. I have reached an accuracy of 88.5%.
Here is the confusion matrix:
In order to test the Neural Network I also created a Shiny app (using this repository as a starting point), which required the import into R of the trained neural network ready to be used for the recognition of alphanumerical characters.
In the web app I implemented a conversion process similar to the one used to get the EMNIST dataset which basically means the removal of the white spaces from the starting image, the application of a Gaussian filter, and the resizing to 28x28.
The Shiny app is available on my website.