LBP vs. LBP Variance for Texture Classification

Abstract

Texture classification algorithms are utilised in various image analysis and medical imaging applications. A number of high performing texture algorithms are based on the concept of local binary patterns (LBP) characterising the relationships of pixels to their local neighbourhood. LBP descriptors are simple to calculate, are invariant to intensity changes and can be calculated in a rotation invariant manner as well as at different scales. Incorporating variance information, leading to LBP variance (LBPV) texture descriptors, has been claimed to lead to more versatile and more effective texture features. In this paper, we investigate this in more detail, benchmarking and contrasting the classification performance of several LBP and LBPV descriptors for generic image texture classification as well as two medical tasks. We show that while LBPV-based methods typically lead to improved classification performance this is not always so and that thus the inclusion of variance information is task dependent.