Abstract
Texture is an important characteristic of images and hence used in a variety of computer vision applications. A group of high performing texture algorithms is based on the concept of local binary patterns (LBP) which describe the relationship of pixels to their local neighbourhoods. A rotation invariant form of this descriptor is typically employed since especially for textured surfaces rotation cannot be controlled. Since conventional LBP discards the magnitude information between the centre pixel and neighbouring pixels, Compound LBP (CM-LBP), a variant of LBP, integrates this information by introducing a 16-bit LBP code. The feature length of CM-LBP is then reduced by splitting this 16-bits into two 8-bit codes. However, this approach does not allow for rotation invariant mappings as in conventional LBP, and CM-LBP hence cannot be applied to images under rotation, thus severly limiting the application of the method. In this paper, we address this problem and present rotation invariant and uniform mappings for CM-LBP. We evaluate our new texture descriptor on Outex and Brodatz benchmark datasets and show it to lead to a significantly improved classification performance compared to CM-LBP.