Simultaneous super-resolution and segmentation for remote sensing images

In this paper, we present an algorithm to simultaneously obtain high-resolution images and segmentation maps from low-resolution inputs. Super-resolution and segmentation both are challenging task, but they may have certain relationship. Super-resolution will provide images with more details that may help to improve the segmentation accuracy, while label maps in segmentation dataset may contribute to finer edges during super-resolution process. Therefore, we aim to combine these two tasks and explore the influence for each other. For this end, we proposed a new deep neural network to simultaneously address the super-resolution and segmentation tasks for remote sensing images, which is named S²Net. The S²Net is an integrated network composed of a super-resolution sub-network and a segmentation sub-network, which is trained in an end-to-end manner. Experimental results demonstrate that this combination can enhance the performance on these two tasks.