Illumination-Robust remote sensing image matching based on oriented self-similarity

Establishing feature correspondences between different images is a key step in many remote sensing applications. In this paper, a novel descriptor based on an extended self-similarity measure, named HOSS (histogram of oriented self-similarity), is proposed for illumination-robust remote sensing image matching. The novel idea of HOSS is a computation of the self-similarity values in multiple directions using an oriented rectangular patch to increase the descriptor distinctiveness. For HOSS construction based on self-similarity values in various directions, a novel index map called RIMC (rotation index of the maximal correlation) incorporated with an adaptive log-polar spatial structure is proposed. Unlike conventional 2D self-similarity descriptor and its extension, HOSS is a 3D histogram of oriented self-similarity values, which is very robust to significant illumination variations. An illumination-robust rotation assignment approach is also developed based on self-similarity gradients to obtain rotation invariance. Experimental analysis on three categories of synthetic and real remote sensing images from various sensors, demonstrate the superior capability of the HOSS over state-of-the-art descriptors, including DOBSS, AB-SIFT, PIIFD, and DAISY, in terms of the recall, precision, and positional accuracy. The average performance of the HOSS in all applied images is about 36%, 51%, and 0.9 pixels in terms of recall, precision, and positional accuracy, respectively. The demo code of the HOSS approach can be downloaded from https://www.researchgate.net/publication/328733655_HOSS_Descriptor.