We trained a convolutional neural network to perform image segmentation on aerial images of urban areas. The objective of the model is to be able to identify and separate roads from their surrounding environment. The network comprises four convolutional layers of increasing depth followed by a fully connected one. Dropout and image augmentation have been used to improve the performance of the model as well as diminish overfitting.
This project was realized as part of the CS-433 Machine Learning course at EPFL.