Réseaux de neurones semi-supervisés pour la segmentation sémantique en télédétection

This work explores different aspects of semantic segmentation of remote sensing data using deep neural networks.The availability of large databases of fully annotated data is the basis for good performances of current neural networks. Although images on remote sensing are abundant, annotations are very rare or difficult to produce. In this context, two questions arise: first, how robust are existing supervised learning strategies with respect to data? Second, is it possible to improve performance of current methods by using non annotated data? Our main contributions are: (i) a strong robustness analysis of existing supervised learning strategies with respect to remote sensing data, (ii) the introduction of a semi-supervised architecture, capable of learning from annotated and non annotated images simultaneously.

Colloque GRETSI 2019