A new efficient feature-combination-based method is proposed using Semi-Random Starting Parameter Dynamic Bayesian Networks (SRSP-D) and by designing an appropriate DBN to solve modeling and classification of Dynamic Textures (DTs). Our approach is based on a BN scheme, especially DBN. Feature combination is accomplished by maximizing the joint probability of different feature vector types during learning process of the DBN. To increase the efficiency of D, SRSP-D are proposed. Experimental results on the DT datasets show that the average classification time of a new DT sample is about 31 and 11 seconds and the average and best correct classification rate is 99.5% and 100% for the UCLA and 98% and 100% for the DynTex datasets, respectively. The approach is invariant to the scale, shift, illumination and rotation variations. In addition, a novel heuristic approach based on a two-phase texture and dynamism analysis is proposed to solve modeling, classification, and synthesis of DTs. In this approach, we capture the texture form textured frames and the dynamism by proposing a mathematical model for dynamism of a DT. The average and best correct classification rate is 97% and 98% for the UCLA and 95.33% and 100% for the DynTex datasets, respectively. This approach is also invariant to the shift, illumination and rotation variations. Key Words Dynamic texture, Modeling, Classification, Dual tree complex wavelet transform, Semi-random starting parameter dynamic Bayesian network, Dictionary of visual words
