Rotation Invariant Local Binary Pattern Descriptor for Remote Sensing Image Matching

Abstract

Distinctive and efficient description of image features is an essential task for image registration inphotogrammetry and remote sensing.The majority of existing descriptors estimate a dominant orientation parameter for rotation invariant image matching.The dominant orientation assignment is an error-prone process, and it decreases the capability of the descriptors.In this paper, a novel feature descriptor based on the local binary pattern operator named RILBP (Rotation Invariant Local Binary Pattern) is proposed that is inherently rotation invariant.To compute the RILBP descriptor, the pixels in the given image region are divided into several sub-regions based on distance and intensity order constraints.Then, a local binary pattern histogram is generated for each sub-region based on a rotation invariant coordinate system.To increase the descriptor robustness against geometric distortions, a special weighting process based on a combined ring and Gaussian functions is applied.The proposed RILBP descriptor was successfully applied for matching of various remote sensing images as: SPOT 5, ETM+, Sentinel 2, IKONOS, IRS P6 and ZY3 sensors, and the results demonstrate its capability compared to common feature descriptors such as CS-LBP, SIFT, LSS, and MROGH.Compared to the standard CS-LBP descriptor, the RILBP descriptor indicates an average performance improvement of about 25%, 10% and 30%, in terms of Recall, Precision and number of correct matches, respectively.