A rotationally invariant descriptor based on mixed intensity feature histograms

This paper proposes a novel method to build a rotation invariant local descriptor by mixed intensity feature histogram. Most existing local descriptors based on intensity ordinal information typically encode only one local feature for each sampling point in the image patch. To address this problem, we proposed a method to encode more than one different local features for each pixel in the image patch and construct a 2D mixed intensity feature histogram, from which our proposed MIFH descriptor is then constructed. Since the rotation invariant coordinate system is adopted, the MIFH descriptor does not need to estimate the reference orientation. In order to evaluate the performance and the robustness of the MIFH with other well-known local descriptors (e.g., SIFT, GLOH, DAISY, HRI-CSLTP, LIOP, MROGH), image matching experiments were carried out on standard Oxford dataset, additional image pairs with complex illumination changes and image sequences with different noises. To further investigate the discriminative ability of the MIFH descriptor, a simple object recognition experiment was conducted on three public datasets. The experimental results demonstrate that our descriptor MIFH exhibits better performance and robustness than other evaluated local descriptors.