A new efficient feature-combination-based method for dynamic texture modeling and classification using semi-random starting parameter dynamic Bayesian networks

Abstract

Dynamic texture (DT) is an extension of texture to the temporal domain. Recently, modeling and classification of DTs have attracted much attention. In many pattern recognition and computer vision problems, such as our case, applying only one descriptor to extract one type of feature vector is not sufficient to obtain all of the relevant information from the input data. Hence, it is necessary to apply two or more descriptors to extract two or more different feature vector types with different dimensions and domains. In this paper, for the purpose of DT classification, we propose a novel approach to efficiently combine all different types of feature vectors describing the DT in their original form, dimensionality, and domain. On the other hand, each DT has two types of information: texture and dynamism. In addition to classification, the two above-mentioned aspects of a DT are efficiently simulated in order to model DTs, using the novel proposed approach. Therefore, a new method for simultaneous modeling and classification of DTs is proposed. Our approach is based on a Bayesian Network (BN) scheme, especially Dynamic Bayesian Network (DBN). To increase the efficiency of DBNs, we propose Semi-Random Starting Parameter Dynamic Bayesian Networks (SRSP-DBNs). The proposed approach is sufficiently fast and outperforms the state-of-the-art DT classification methods in terms of accuracy. Furthermore, it is invariant to all types of changes that may occur in the DT, including shift, illumination, rotation, and scale variations.