Um estudo de design science research (DSR) utilizado em uma abordagem de inteligência artificial na redução de dados em bordas computacionais

Loading...
Thumbnail Image

Journal Title

Journal ISSN

Volume Title

Publisher

Universidade Federal de Juiz de Fora (UFJF)

Abstract

The growing volume of digital data, the pillar of Artificial Intelligence (AI), originating from the computational edge, especially in applications with Unmanned Aerial Vehicles (UAVs) that capture large amounts of images in the field, poses a critical challenge: the need to optimize processing and latency on devices with severe hardware and power constraints. Aiming to analyze and compare AI-based image compression methods to mitigate latency at the edge, this work employed the Design Science Research (DSR) methodology in two iterative cycles. The solution developed consisted of implementing Autoencoder (AE) models of different complexities. Conventional, Variational Autoencoder (VAE), and Redundancy-Penalized Autoencoder, trained in a high-performance cloud environment using UAV images from the SARD-2 dataset. In Cycle 1, the models were compared and optimized for reconstruction quality and computational efficiency, revealing that a simple and optimized structure is more effective than overly complex architectures. The Optimized Conventional AE outperformed the more complex variants, achieving the best balance between quality (PSNR = 20.71 dB; MS-SSIM = 0.9359) and processing time. In Cycle 2, the winning model was converted and run in a simulated edge environment using the OpenVINO platform, with restricted hardware simulation and FP32 precision. Experimental validation demonstrated an average latency of 21.2 ms and adequate temporal stability for near real-time applications, confirming the viability of the lightweight model in embedded devices. The results consolidate the thesis that efficiency at the edge depends on the structural adequacy of the model to the hardware, and not only on algorithmic sophistication. This work contributes as a replicable methodological baseline for intelligent UAV image compression, reconciling performance, computational sustainability, and Open Science principles.

Description

Citation

Endorsement

Review

Supplemented By

Referenced By

Creative Commons license

Except where otherwise noted, this item's license is described as Acesso Aberto