Salvetti Francesco ; Mazzia Vittorio ; Khaliq Aleem ; Chiaberge Marcello

Abstract

Convolutional Neural Networks (CNNs) have been consistently provedstate-of-the-art results in image Super-Resolution (SR), representing anexceptional opportunity for the remote sensing field to extract furtherinformation and knowledge from captured data. However, most of the workspublished in the literature have been focusing on the Single-ImageSuper-Resolution problem so far. At present, satellite based remote sensingplatforms offer huge data availability with high temporal resolution and lowspatial resolution. In this context, the presented research proposes a novelresidual attention model (RAMS) that efficiently tackles the multi-imagesuper-resolution task, simultaneously exploiting spatial and temporalcorrelations to combine multiple images. We introduce the mechanism of visualfeature attention with 3D convolutions in order to obtain an aware data fusionand information extraction of the multiple low-resolution images, transcendinglimitations of the local region of convolutional operations. Moreover, havingmultiple inputs with the same scene, our representation learning network makesextensive use of nestled residual connections to let flow redundantlow-frequency signals and focus the computation on more importanthigh-frequency components. Extensive experimentation and evaluations againstother available solutions, either for single or multi-image super-resolution,have demonstrated that the proposed deep learning-based solution can beconsidered state-of-the-art for Multi-Image Super-Resolution for remote sensingapplications.