A new efficient method to characterize dynamic textures based on a two-phase texture and dynamism analysis

Dynamic texture (DT) is an extension of texture to the temporal domain. Recently, description and classification of DTs have attracted much attention. In this article, a new method for classifying and synthesizing DTs is proposed. This method is based on a two-phase texture and dynamism analysis. At first, a mathematical model is proposed to model the dynamism of a DT. Then a DT is described in a two-step procedure: describing the texture and the dynamism of a DT. Dual Tree Complex Wavelet Transform (DTCWT) is applied on the textured frames and models of the dynamism of the DT. This makes our algorithm robust enough to illumination and shift variations. The mean and standard deviation of the complex wavelet coefficients, as obtained from textured frames and models of the dynamism of the DT, are concatenated and used to form the DT feature vector. By using Fourier Transform of the feature vector, rotation invariance is also provided. A dictionary of visual words made from these feature vectors is then used to describe each DT. Together, the two phases cover a large variety of DT classification problems, including the cases where classes are different in both appearance and motion and those where appearance is similar for all classes and only motion is discriminant or vice versa. Our approach has many advantages compared with the earlier ones, such as providing high speed and a better performance for two test databases. (C) 2014 Elsevier B.V. All rights reserved.