An Orientation-Adaptive Extension to Scale-Adaptive Local Binary Patterns

Methods based on Local Binary Patterns have been used successfully in a wide range of texture classification tasks. A restriction shared by all methods based on Local Binary Patterns is the high sensitivity to signal scale. In recent work we presented a general framework for scale-adaptive computation of Local Binary Patterns, improving the accuracy in texture classification scenarios involving varying texture-scales highly. In this work, the scale-adaptive methodology is extended by an orientation-adaptive computation of patterns, leading to a scale- and rotation invariant classification. The results suggest that estimating a global orientation to build orientation-adaptive LBPs is superior to the previously introduced rotation-invariant encodings. The proposed framework allows the use of the highly-discriminative LBPs in less-constrained situations, where both orientation, as well as scale variations, are to be expected.