基于时间-顶点谱图小波变换的动态纹理分类方法

Dynamic Texture Classification Method Based on Spectral Time-Vertex Wavelet Transform

动态纹理在空间和时间上表现出“外观”和“运动”属性,为了有效结合这两种属性进行动态纹理分析,本文提出一种基于时间—顶点谱图小波变换与边缘分布协方差模型的动态纹理分类方法。该方法将动态纹理看成时间—顶点图信号,利用时间—顶点谱图Meyer小波变换对动态纹理进行多尺度分解,再对每个子带应用边缘分布协方差模型,由此得到带内相关性的特征协方差矩阵作为动态纹理特征进行分类。由于时间—顶点图信号的表示可以有效描述动态纹理像素间的空间关系及其沿时间的变化,同时谱图小波变换继承了图表示和小波变换的优势,因此利用时间—顶点谱图小波分解与边缘分布协方差模型,可得到有效的动态纹理特征。在标准动态纹理数据集上的分类实验结果表明,本文方法具有良好的分类性能。

Dynamic textures are sequences of images that change over time.Dynamic texture classification plays an important role in medical testing,industrial production and forest fire control.Dynamic textures exhibit“appearance”and“motion”properties in space and time.Combining these two properties,a dynamic texture classification method based on spectral time-vertex wavelet transform and Marginal distribution covariance model was proposed in this paper.The time-vertex graph signal processing framework was used to represent the dynamic texture as the time-vertex graph signal.Since Meyer wavelet can represent dynamic textures in multiple directions and at multiple scales,so the multi-scale decomposition of dynamic texture was performed by spectral time-vertex Meyer wavelet transform.Then,Marginal distribution covariance model was applied to each sub-band,and the characteristic covariance matrix of intra-band correlation was obtained as dynamic texture feature for classification.Due to the representation of time-vertex graph can effectively describe the spatial relations among dynamic texture pixels and their changes along time,meanwhile,spectral wavelet transform inherits the advantages of graph representation and wavelet transform,so we used spectra time-vertex wavelet decomposition and marginal distribution covariance model to obtain dynamic texture feature effectively.The experiment results on standard dynamic texture data sets show that the proposed method has good classification performance.