Dynamic Textures Retrieval Based On Surfacelet Transform

Dynamic textures are sequences of images that exhibit spatial repeatability, temporal continuity and comply with certain statistical characteristics. Many phenomena like flowing water, wind-shaken trees and rising smokes can be classified as examples of dynamic textures. The analysis of dynamic textures is an extremely important problem in the field of digital image processing and computer vision. The description and retrieval of dynamic textures is a key research topic.Recently, multi-resolution and multi-direction analysis on multidimensional signal is a hot topic in academe. Aimed at the issue of dynamic textures retrieval, this paper combined the research of dynamic textures analysis with the method of 3-d wavelet transform, and proposed a dynamic texture retrieval method based on surfacelet transform. The main work is as following:Firstly, this paper introduced some traditional methods for describing dynamic textures, mainly includes optic flow, computing geometric properties in the spatiotemporal domain, local spatiotemporal filtering, global spatiotemporal transforming and estimating model parameters. Considered the method based on optic flow is currently the most popular approach and is the basis of most current methods, this paper mainly introduced the optic flow method.Secondly, this paper introduced a new Multi-resolution and multi-direction wavelets named Surfacelet. In order to overcome the disadvantages of optic flow, this paper used surfacelet transform in dynamic textures description and retrieval. This thesis introduced the theory, the sub-band coefficients distribution of surfacelet transform, and proposed an algorithm that extracts features based on surfacelet transform. This paper constructed a prototype system, which combined with the mean and the variance of energy distribution of the transform coefficients for each sub-band at each decomposition level for feature description and the Euclidean distance for similarity measure.This paper tested its performance through a simulation. Compared with the 3-D DWT method, Horn&Schunck method and Lucas&Kanade method, the dynamic texture retrieval experimental results show that the method based on surfacelet transformation provided a better performance.