This classification project using CNN consists of three phases:
- Processing the data, including data augmentation
- Defining the CNN architecture
- Train and test the model
The data is a subset of the Char74K, which can be downloaded here
The CNN architecture was a simple one, combining Convolutional, Pooling and Fully-Connected layers. After 20 epochs, which took 1min 30s, the final valid accuracy was 99%. After testing the model, every class reached more than 99%, except the number 8 which reached 96% (eight, you are tough!).
An example of a prediction, in parentheses the probability of the prediction:
