Dynamic Texture Classification with Relative Phase Information in the Complex Wavelet Domain

Abstract

In recent years, dynamic texture classification has caused widespread concern in the image sequence analysis field. We propose a new method of combining relative phase information of dynamic texture in the complex wavelet domain with probability distribution models for dynamic classification in this paper. Instead of using only real or magnitude information of dynamic texture, relative phase information is an effective complementary measure for dynamic texture classification. Firstly, the finite mixtures of Von Mises distributions (MoVMD) and corresponding parameter estimation method based on expectation-maximization (EM) algorithm are introduced. Subsequently, the dynamic texture features based on MoVMD model for dynamic texture classification are proposed. Besides, the relative phase information of dynamic texture is modeled with MoVMDs after decomposing dynamic texture with the dual-tree complex wavelet transform (DT-CWT). Finally, the variational approximation between different dynamic textures is measured using the Kuller-Leibler divergence (KLD) variational approximation. The effectiveness of the proposed method is verified by experimental evaluation of two popular benchmark texture databases (UCLA and DynTex++).