Geometric Filtering of Matches Between Points in Bands of Hyperspectral Cubes
Nenhuma Miniatura disponível
Data
2019-03-01
Autores
Berveglieri, Adilson [UNESP]
Tommaselli, Antonio Maria Garcia [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Ieee-inst Electrical Electronics Engineers Inc
Resumo
Hyperspectral images can present low contrast, noisy pixels, and illumination variation among bands, which complicates the extraction of interest points and reduces the number of reliable image matches affecting subsequent tasks as band registration and bundle adjustment. Once matched points have been determined, a technique to select correct matches in sets with outliers is required, as well as to fix mismatches. In this letter, we apply a filtering technique that uses a majority voting algorithm combined with a 2-D Helmert geometric transformation to identify consistent matches. The correct matches also allow the estimation of parameters of a geometric transformation, which enables point transfer between images. Thus, mismatches can be fixed to their correct positions. Experiments were performed with the proposed technique using hyperspectral images that were collected with a lightweight camera using the time-sequential principle, while onboard an unmanned aerial vehicle. Scale-invariant feature transform was used for both keypoint extraction and image matching. Reliable matches were extracted from the sets with outliers, and incorrect matches were fixed. The results of the technique were compared with an algorithm based on random sample consensus. In the comparison, the proposed technique was efficient in extracting a larger number of correct matches. In addition, 85% of the incorrect matches were recovered, which significantly increased the density of matched pairs.
Descrição
Palavras-chave
Correlation, filtering, image analysis, image matching, stereo image processing
Como citar
Ieee Geoscience And Remote Sensing Letters. Piscataway: Ieee-inst Electrical Electronics Engineers Inc, v. 16, n. 3, p. 492-496, 2019.