An investigation on the use of the RGB color model in the matching process by correlation

Abstract

Image matching is one of the most important processes in Digital Photogrammetry,
since it allows the automation of several stages of the photogrammetric pipeline. In
most of the commercial software nowadays available, the algorithms of image
correlation use the intensity information (gray levels), despising color information,
that could be useful, if used in a suitable way, increasing the robustness of the
current correspondence techniques in Digital Photogrammetry. The aim of this
work is to present a technique that uses the RGB color model in the correlation
process, in which a correlation matrix is generated for each color channel. The trace
of the covariance matrix related to the translations of the reference window is used
to predict which channel can better contribute to the result of the correlation and,
with this, to properly weight the correlation coefficients. The weights to be applied
to each one of the correlations matrixes are computed adaptively, considering the
characteristics of each image. In order to assess this methodology, experiments with
real color aerial images were accomplished and correct correlations were achieved
with the proposed technique but failed with the current techniques using only grey
level images. The results are presented and discussed, showing that the use of color
information increases the robustness of the correlation process.