Structured Textons for Texture Classification

In this paper, we are proposing a novel texture descriptor, Structured Texton, for image representation. By exploring the local structure existed in the meaningful texture patterns, the structured textons are constructed by the nesting relationship between the different scale local extrema regions, which characterizes not only the local appearance features but also the geometric information. Furthermore, to improve the discriminability of the texture descriptors, higher order texton-words are constructed by means of finding informative patterns from the texton bank, which reduces the ambiguity of the textons through the higher order representation. We conduct two experiments as applications: the texture classification and scene classification. The improvements in the performance demonstrate the effectiveness of the proposed texture descriptors.