Here is an example of a signature sheet we used to collected 80 signatures from 10 participants.
These signatures were scanned in 200dpi grayscale.
In designing the final neural network we considered many features. For example, how many layers, how many filters, and should the data be binary or not.
Below is two surface plots. The first is showing binarized data while the second is showing non binarized data. The x and y axis represent hidden layer number and filter number, respectively. The z axis shows accuracy across the testing and training data for the top two graphs. And the bottom graph shows performance. The closer the performance value is to zero, the better the performance.