Deep Super-Resolution for Autonomous Aerial Robotics

The performance of autonomous aerial robots depends heavily on visual perception, yet limitations in onboard hardware often result in low-resolution imaging and degraded visual quality. This paper introduces AESR++, an enhanced autoencoder-based super-resolution model tailored for aerial robotics. The architecture extends traditional encoder–decoder designs with skip connections, self-attention mechanisms, attention regularization modules, and contrastive latent learning to achieve improved reconstruction robustness and generalization. Evaluated on the Drone Super-Resolution (DSR) dataset, AESR++ demonstrates superior results over recent state-of-the-art methods across multiple upscaling factors (2×, 4×, and 8×) and evaluation metrics (PSNR, SSIM, LPIPS, DISTS, NIQE, BRISQUE). Qualitative analyses further confirm its ability to preserve fine details and natural appearance under diverse aerial scenarios. By enabling high-quality reconstructions from low-resolution inputs, AESR++ provides an effective and scalable solution for remote image processing, with direct implications for tasks such as navigation, mapping, environmental monitoring, and decision-making in aerial robotics.