A Comparative Study of Autoencoder Models for Visual Data Reduction in Unmanned Aerial Vehicle (UAV) Operations in Search and Rescue (SAR) Operations

Abstract

This article evaluates the performance of different autoencoder architectures, conventional, variational, and redundancy-penalized, for image compression in autonomous aerial vehicles during search and rescue operations. The main hypothesis anticipated that penalized models would outperform others, balancing computational efficiency and reconstruction fidelity. Experiments were conducted using the SARD 2 dataset, and the analysis considered PSNR, SSIM, MS-SSIM, compression rate, and processing time. Results showed that the conventional autoencoder achieved higher reconstruction quality, while the penalized model excelled in computational efficiency. The variational autoencoder presented intermediate performance. Contrary to expectations, the hypothesis was not fully confirmed, highlighting the need to balance quality and efficiency according to operational constraints. Statistical analysis (Kruskal-Wallis test) indicated significant differences among models. It is important to note the limitation regarding end-to-end latency assessment, as only compression and reconstruction times were measured in the experimental environment. For future work, evaluation in real transmission scenarios and analysis of energy impact and adaptive approaches for different operational conditions are recommended.