An AI that can recognize handwritten characters trained with the mnist special database 19, made using tensorflow/keras
As the dataset can only be downloaded, a data collection and formatting method was required. data_collect.ipynb iterates through the by_merge dataset and extracts all the images/labels using cv2. Afterwards, the data was randomly shuffled and stored in X (images) and y (labels).
My first attempt at making an AI that would read the 64x64 images. Used tensorflow/keras to split and normalize the training and testing data (I used a 85-15 split). Afterwards, a neural network with 3 hidden (dense) layers of 512 nodes each was made, using the softplus activation function for the hidden layers
With 5 epochs, I achieved a loss of 0.4810 (using sparse_categorical_crossentropy) and an accuracy of 0.8452. The model was stored in the model folder.
Using kivy, a GUI was made that would input your own custom image into the model. A canvas which can be written on with a mouse was created, which would pass that canvas to the model folder which would then give the AI's prediction.
Dataset from https://www.nist.gov/srd/nist-special-database-19