Dynamic Texture Recognition Based On Multiple Statistical Features With LBP/WLD

Dynamic textures are textures with motion, and they encompass sequences of images that exhibit certain temporal stationary. There are lots of dynamic textures in nature, e.g., smoke, flag in wind, fire and flying birds. Potential applications of dynamic textures include monitoring forest fires and traffic, homeland security applications, animal behavior studying and so on. Due to the unknown spatial and temporal extension, the recognition of dynamic textures becomes a newly and highly challenging problem.The existing methods of dynamic texture recognition mainly fall into two categories:one is based on motion analysis and the other is based on comprehensive analysis. LBMP is a recently proposed effective approach based on comprehensive analysis. It utilizes block matching technique and LBP to abstract dynamic and apprearance feature, and by combining the two features it gives an overall description of dynamic texture. Although it costs simple calculation and can achieve high recognition accuracy, it has disadvantages:in the phase of feature abstraction, the single use of LBP omissives intensity information of textures; in the phase of dynamic feature abstraction, employing LBP as the matching discipline can produce multiple matching points which dinimish the matching accuracy; in the phase of appearance feature abstraction, computing LBP statistical features from several frames with the same time interval is not accurate enough, and it makes the connectivity between appearance feature and dynamic feature low. To overcome the above issues in LBMP, we present a novel approach based on multiple statistical features with LBP and WLD. A newly proposed texture descriptor WLD is combined with LBP in order to get a more full-scale description of dynamic texture; a three-stage matching strategy based on LBP, WLD and Euclidean Distance is used to reduce the deviation caused by multiple matching points; distribution of LBP/WLD based on orientation of dynamic texture’s motion is gathered as appearance feature; concentrating the above two features can represent dynamic texture more generally and precisely. Experiments on dynamic texture database Dyntex demonstrate that the proposed approach has better recognition accuracy than LBMP.