Low-power neural networks for semantic segmentation of satellite images
Résumé
Semantic segmentation methods have made impressiveprogress with deep learning. However, while achievinghigher and higher accuracy, state-of-the-art neural net-works overlook the complexity of architectures, which typ-ically feature dozens of millions of trainable parameters.Consequently, these networks requires high computationalressources and are mostly not suited to perform on edgedevices with tight resource constraints, such as phones,drones, or satellites. In this work, we propose two highly-compact neural network architectures for semantic segmen-tation of images, which are up to 100 000 times less complexthan state-of-the-art architectures while approaching theiraccuracy. To decrease the complexity of existing networks,our main ideas consist in exploiting lightweight encodersand decoders with depth-wise separable convolutions anddecreasing memory usage with the removal of skip connec-tions between encoder and decoder. Our architectures aredesigned to be implemented on a basic FPGA such as theone featured on the Intel Altera Cyclone V family of SoCs.We demonstrate the potential of our solutions in the case of binary segmentation of remote sensing, in particular for extracting clouds and trees from satellite images.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...