Weighty LBP: a new selection strategy of LBP codes depending on their information content

This paper presents a novel variation of the use of LBP codes. Similarly to Uniform LBP and Local Salient Patterns (LSP), it aims at both obtaining an effective texture description, and decreasing the length of the feature vectors, i.e., of the chains of LBP histograms. Instead of considering uniform codes, we rather consider the codes providing the highest “representativeness” power with respect to texture features. We identify this subset of codes by a generalized notion of entropy. This allows determining the most informative items in an homogeneous set.