On-the-spot handwriting number recognition based on a two-layers dense neural net on ImageJ.

• mnist2layersKeras.py
• mnist2layersWeight.txt
• mnist_2layers.ijm

This program imports MNIST dataset, and executes two-layers dense neural net trained using Keras. The parameters of the neural net (weights and biases) will be saved as a text file in '/Applications/ImageJ/plugins/mymacro/'. Please install imageJ and create 'mymacro' folder in the 'plugins' folder before running this program.

An example output of mnist2layersKeras.py.

This program is written in ImageJ macro. Handwriting number will be estimated based on the two-layers neural net with the parameters in mnist2layersWeight.txt.

1. Install imageJ, create 'mymacro' folder in 'plugins' folder.
2. Put 'mnist2layersWeight.txt' and 'mnist_2layers.ijm' in mymacro folder.
3. Start ImageJ.
4. You can see 'mymacro' in Plugin menu. Select 'mnist 2layers' in plugin menu.
5. GUI will guide you. 20-pixel width paintbrush tool is optimal for drawing a number.

• If you activate "show probability", the probability distribution will be shown.
• Press OK after drawing a number, then you get a popup with an estimated number and a bar plot of the probability distribution.
• The image is binned by average into 28x28 pix during estimation.

Run the python program 'mnist2layersKeras.py' in an appropriate environment. Please confirm the package requirements are fulfilled.