Dynamic texture recognition using multiresolution edge-weighted local structure pattern

Abstract Dynamic texture has been found as a powerful cue for modeling natural scenes such as fire, waves and smoke, etc. It combines appearance with motion to characterize the moving scene that exhibits certain spatially repetitive and time-varying visual patterns. This paper proposes a new method of recognizing dynamic texture using the well-known texture descriptor, local binary pattern. The new variant differentiates different structural patterns more efficiently using the additional information from the local patch. This pattern information is further combined with shape information to improve the discriminative power of texture descriptor. The proposed method is extended to multiscale using classifier fusion scheme to capture the spatio-temporal content of a moving scene at multiple scales, thus improves representation capability of the new descriptor. Proposed descriptor is tested on three dynamic texture databases: UCLA, Dyntex and Dyntex++. Results demonstrate that the proposed feature descriptor outperforms various state-of-the-art approaches on all representative databases in terms of classification accuracy.