Toward Understanding the Impact of Input Data for Multi-Image Super-Resolution

Super-resolution reconstruction is a common term for a variety of techniques aimed at enhancing spatial resolution either from a single image or from multiple images presenting the same scene. While single-image super-resolution has been intensively explored with many advancements proposed attributed to the use of deep learning, multi-image reconstruction remains a much less explored field. The first solutions based on convolutional neural networks were proposed recently for super-resolving multiple Proba-V satellite images, but they have not been validated for enhancing natural images so far. Also, their sensitiveness to the characteristics of the input data, including their mutual similarity and image acquisition conditions, has not been explored in depth. In this paper, we address this research gap to better understand how to select and prepare the input data for reconstruction. We expect that the reported conclusions will help in elaborating more efficient super-resolution frameworks that could be deployed in practical applications.