This project is based on the INRIA Aerial Image Labeling Dataset. The aim is to build an algorithm based on neural network performing semantic segmentation on buildings.

The open source library buzzard developed by Airware has been used to manage georeferenced data. The results are saved and store at the geoJSON format.

The solution implemented is inspired by the TernausNet based on an autoencoder that reuses the features learned during the encoding part for the decoding part. The used architecture was :

All the convolutional layers except the last one (sigmoid) are using ReLu activation function. The size of the filters is (3,3) and there are 2 strides for the deconvolution.

The network is trained on n_samples croped images of size (image_size,image_size) resampled with a given downsampling_factor.

The treatment for the whole image is proceed by tiling the whole image into several tiles of size (image_size,image_size) and then untiling them to reconstruct the entire binary image. The results are polygons saved at the geoJSON format (same name of the original image)

Due to limited computational performances, we only chose a dataset of 180x5 images with downsampling_factor=3. The batch number is fixed at 40 and the number of epochs at 50. Here is the history of the loss function (binary cross entropy):

The results are shown for the first train file of the dataset:

The original image:

The reference binarized image:

The predicted binarized image:

To evaluate the accuracy of the model, with use the intersection over union metric. The results for the different towns displayed in the following table: