Alinhamento de alta qualidade de imagens aéreas RGB capturadas por UAV para construção de pares de baixa e alta resolução

Abstract

The construction of consistent datasets for aerial image super-resolution depends on the correct spatial correspondence between low- and high-resolution images. In scenarios involving unmanned aerial vehicles (UAVs), variations in perspective, scale, altitude, illumination, and camera positioning may cause misalignments between LR/HR pairs, compromising the training and evaluation of super-resolution models. Therefore, this work proposes an approach for alignment and automatic quality filtering of RGB aerial image pairs. The methodology includes pair formation and verification, selective global registration, block-based local analysis, automatic filtering, and construction of the final dataset. Global registration uses SIFT, Lowe’s ratio test, and RANSAC to estimate candidate homographies, which are applied only when they improve similarity metrics and preserve sufficient valid image area. The local analysis estimates residual displacements using phase correlation. Experiments were conducted on the DSR dataset, in which 1057 pairs were evaluated: 898 were accepted, 134 were assigned for review, and 25 were rejected. The results indicate that the proposed approach evaluates alignment quality, avoids unnecessary transformations, and selects more reliable pairs for super-resolution applications.