Summary: | m<sub>p</sub>-dissimilarity is a recently proposed data-dependence similarity measure. In the literature, how m<sub>p</sub>-dissimilarity is generally used for matching local image descriptors has been formalized, and three matching strategies have been proposed by incorporating ℓ<sub>p</sub>-norm distance and m<sub>p</sub>-dissimilarity. Each of these three matching strategies is essentially a two-round matching process that utilizes ℓ<sub>p</sub>-norm distance and m<sub>p</sub>-dissimilarity individually. This paper presents two novel similarity measures for matching local image descriptors. The first similarity measure normalizes and weights the similarities that are calculated using ℓ<sub>p</sub>-norm distance and m<sub>p</sub>-dissimilarity, respectively. The second similarity measure involves a novel calculation that takes into account both spatial distance and data distribution between descriptors. The proposed similarity measures are extensively evaluated on a few image registration benchmark data sets. Experimental results will demonstrate that the proposed similarity measures achieve higher matching accuracy and are able to attain better recall results when registering multi-modal images compared with the existing matching strategies that combine ℓ<sub>p</sub>-norm distance and ℓ<sub>p</sub>-dissimilarity.
|