Rubik Gaussian-based patterns for dynamic texture classification

Abstract Illumination, noise, and changes of environments, scales negatively impact on encoding chaotic motions for dynamic texture (DT) representation. This paper proposes a new method to overcome those issues by addressing the following novel concepts. First, different Gaussian-based kernels are taken into account as an effective filtered pre-processing with low computational cost to point out robust and invariant features. Second, a discriminative operator, named Local Rubik-based Pattern (LRP), is introduced to adequately capture both shape and motion cues of DTs by proposing a new concept of complemented components together with an effective encoding method. In addition, it also addresses a novel thresholding to take into account rich spatio-temporal relationships extracted from a new model of neighborhood supporting region. Finally, an efficient framework for DT description is presented by exploiting operator LRP for encoding various instances of Gaussian-based volumes in order to form a robust descriptor against noise, changes of illumination, scale, and environment. Experiments for DT classification on benchmark datasets have authenticated the interest of our proposal.