Dynamic texture recognition using local tetra pattern—three orthogonal planes (LTrP-TOP)

Abstract

Dynamic texture (DT) is the annexation of an altering series of images where temporal regularity is present. Characterization, classification and identification of DTs have been investigated by many researchers in the past 2 decades. Especially, several image texture descriptors based on local binary pattern (LBP) have been enhanced for DT classification in the spatiotemporal domain. Local tetra pattern (LTrP) is an extension of LBP in which the calculation of the feature code depends on the referenced central pixel as well as its neighbor’s directions. In this work, we have proposed an LTrP-based novel spatiotemporal descriptor for characterization and subsequent identification of textures in videos. We term this descriptor as local tetra patterns on three orthogonal planes (LTrP-TOP). In our proposed work, the video frames are initially divided into three orthogonal planes on the basis of the XY, YT and XT directions. The direction of the center and neighborhood pixels are then detected in vertical and horizontal orientations using first-order derivatives. This sequence of operations computes the LTrP for each center pixel. Subsequently, we create the histogram for a particular plane by considering the LTrP codes for all the pixels in that plane. Finally, the histograms from the three orthogonal planes are concatenated for forming the final texture descriptor of the video. Comprehensive experimental assessments on benchmark DT databases (Dyntex++ and UCLA) demonstrate that the LTrP-TOP descriptor exhibits better performance than most of the other state-of-the-art DT descriptors (in terms of classification accuracy %).