Deep Neural Network Algorithms for Upsampling of Surface Images

Physical Sciences Inc. (PSI) proposes to develop the Single Image Super Resolution for Quantitative Analysis (QuantSISR) software suite comprising of state-of-the-art super-resolution (SR) algorithms optimized to reduce errors during subsequent image analysis such as common computer vision tasks (image segmentation, object detection). QuantSISR will be designed to achieve a 50% reduction in edge localization errors while matching pixel-wise accuracy comparable to methods optimized for visual perception quality. The QuantSISR algorithms will improve temporal coverage of data products, such as land cover/land use maps and building footprints.nbsp; These objectives are achieved by generating super-resolved visible and hyperspectral images from low-resolution data sets to increase the temporal revisit frequency of existing high resolution datasets. The algorithms are able to generalize to new sensors and regions without any reference imagery, but can also utilize high-resolution reference imagery, when available, to improve accuracy. nbsp;This feature can be used during Solar System exploration missions to mitigate mismatch between terrestrial training data sets and the newly acquired data. By leveraging multiple observation geometries, high resolution in situ references can be obtained and used to enhance wide area images acquired at lower spatial resolution. QuantSISR algorithms will be capable of 2x-8x up-sampling and support processing of multispectral and hyperspectral data and the high dynamic range (ge; 16 bit) of modern imagers. QuantSISR software suite will incorporate utilities for parsing and assembling common image data types and image pre- and post-processing to enable seamless integration with existing processing infrastructures. QuantSISR will be packaged to operate on a range of user designated computing platforms, from embedded CPU-GPU systems to computer clusters and cloud computing services.