Autoencoder-based Super-Resolution Approach for Aerial Robot Navigation

Images are increasingly used in the navigation of autonomous aerial vehicles. However, processing images locally in aerial robots can be very computationally and energy-intensive. In this sense, it is increasingly common to use remote data processing for applications involving mobile robots. This paper addresses the problem of super-resolution as a strategy for remote processing of aerial images, considering the proposed algorithm as a technique that reconstructs high-resolution images from low-resolution images captured by the aerial vehicle. To this end, we propose an approach called AESR, based on Autoencoder architecture for generating high-definition images, integrating skip-connections and self-attention modules to improve the learning process. The results indicate the high performance of our approach, validating the learning process based on autoencoder. Moreover, the approach achieved high accuracy even in challenging images representing different environments. The experiments demonstrate the robustness and accuracy of our approach, overcoming state-of-the-art techniques, through the use of the quality metrics PSNR, SSIM, LPIPS, DISTS, NIQE, and BRISQUE.