Dynamic texture description using adapted bipolar-invariant and blurred features

Encoding turbulent properties of dynamic textures (DTs) is a challenging issue of video understanding for various applications in computer vision. It is partly due to the negative impacts of noise, changes of illumination, and scales. In order to deal with those influences, this paper proposes a new approach in which local adapted features of multi-Gaussian-filtered outcomes are exploited for DT representation against the well-known problems. To this end, we firstly take multi-scale 2 D /3 D Gaussian-based filtering kernels into account video analysis in order to correspondingly obtain Gaussian-based filtered outcomes of which the blurred and bipolar-invariant characteristics are complementary. Secondly, due to the sensitivity to noise and near-uniform regions in the encoding of bipolar-invariant features, we propose an essential modification for completed local binary pattern operator to form a more discriminative operator, named Completed AdaptIve Pattern, so that it can be in accordance with the perplexity. Finally, a prominent framework is introduced to efficiently capture DTs’ shape and motion clues in Gaussian-based filtered results. The proposed descriptors are verified on benchmark datasets for DT classification task. Experimental results have validated the interest of our method.