Super-Resolution for Overhead Imagery Using DenseNets and Adversarial Learning
This work addresses image quality improvement for satellite and aerial imagery, which is incremental as it builds on existing GAN methods with architectural modifications.
The paper tackles super-resolution for overhead imagery by proposing a GAN-based architecture using DenseNets, achieving up to 8x resolution enhancement and reporting results on datasets like SpaceNet and IARPA Multi-View Stereo Challenge.
Recent advances in Generative Adversarial Learning allow for new modalities of image super-resolution by learning low to high resolution mappings. In this paper we present our work using Generative Adversarial Networks (GANs) with applications to overhead and satellite imagery. We have experimented with several state-of-the-art architectures. We propose a GAN-based architecture using densely connected convolutional neural networks (DenseNets) to be able to super-resolve overhead imagery with a factor of up to 8x. We have also investigated resolution limits of these networks. We report results on several publicly available datasets, including SpaceNet data and IARPA Multi-View Stereo Challenge, and compare performance with other state-of-the-art architectures.