Description of interest regions based on combined local binary patterns

Abstract

This paper proposes a novel inherently rotation invariant local descriptor which combined intensity information and gradient information of key feature. The CS-LBP shows a better performance than SIFT and do not need large computation. To further enhance its performance and robustness, we calculated the gradient of key feature and computed a combined histogram included intensity and gradient information of sample point. Moreover, our descriptor does not estimate a reference orientation for achieving rotation invariance. We conduct the image matching experiments on Oxford dataset and additional image pairs with large illumination changes to evaluate the performance of our descriptor and existing local descriptors (such as SIFT, GLOH, DAISY, MROGH-S, etc.). The experimental results demonstrate the effectiveness of our descriptors and their superiority to the state-of-the-art descriptors.