Code to create and train a neural network
Neural Networks are becoming fashionable.
There are multiple applications for images studies.
The goal here is to create, train, and use a neural network to predict landuse in a Finnish historical map.
Four students in the SIGMA master degree of the ENSAT INP School and the Jean Jaures Toulouse University, willing to learn about neural networks and the Deep Learning.
To use those scripts you will previously need to install MuseoToolbox, tensorflow, pydot, graphviz, opencv, and numpy libraries.
The code "extract_label_nn.py" is used to the images you want to work with, it uses the polygones stored in the gpkg file and the compress image of a Finnish historical maps(example_entire_map_compress.tif). By default, we create images of 30x30 pixels and all the label associated (all the data is furnished in the git). We then create and train a model (neural_networks_FinlandMaps.py) to process the images. At the end the script gives you the overall accuracy of the model predictions and a png files of the architecture of the model.