/Variational-Autoencoder-For-Satellite-Imagery

This is my implementation of a special Variational Autoencoder under TF 2.0, which make it possible to squeeze N images to generate one single representation of all the dataset with colors segmentation of the difference objects

Primary LanguageJupyter Notebook

Segmenter Variational Autoencoder

This VAE has the ability to generate from within the distribution of satellites images a new image.

Hyperparameters

Hyperparameters Values
Epochs 2
batch size 32
learning rate 1e-4
dropout 0.5
mse_weight 1
KLDiv_weight 1e-3
Hidden activation SeLU
Latent activation Sigmoid

The dataset

Data represents satellites images of cities with multiples objects. Capture d’écran 2020-12-03 à 2 22 20 AM

Generative representation of the dataset

The resulted image shows interesting result, where objects are segmented with a specific color which make us able to distinguish an item from another. Capture d’écran 2020-12-03 à 2 31 58 AM