Exploring the Utilization of Gradient Information in SIFT Based Local Image Descriptors

The utilization of gradient information is a key issue in building Scale Invariant Feature Transform (SIFT)-like descriptors. In the literature, two types of gradient information, i.e., Gradient Magnitude (GM) and Gradient Occurrence (GO), are used for building descriptors. However, both of these two types of gradient information have limitations in building and matching local image descriptors. In our prior work, a strategy of combining these two types of gradient information was proposed to intersect the keypoint matches which are obtained by using gradient magnitude and gradient occurrence individually. Different from this combination strategy, this paper explores novel strategies of weighting these two types of gradient information to build new descriptors with high discriminative power. These proposed weighting strategies are extensively evaluated against gradient magnitude and gradient occurrence as well as the combination strategy on a few image registration datasets. From the perspective of building new descriptors, experimental results will show that each of the proposed strategies achieve higher matching accuracy as compared to both GM-based and GO-based descriptors. In terms of recall results, one of the proposed strategies outperforms both GM-based and GO-based descriptors.