Complex Wavelet Based Dynamic Texture Classification

Texture is a natural property of the surface of the natural object. According to differentrepresentation methods, it can be divided into image texture, dynamic texture andthree-dimensional texture. Dynamic texture is a spatially repetitive, time-varying visualpattern that forms an image sequence with some spatio-temporal stationary properties.Dynamic texture classification plays an important role in texture analysis, and it could bewidely applied in many fields, such as military affairs, industries, medical treatments,intelligent transportation, meteorology, public safty and so on. Thus, dynamic textureclassification has been an interesting and challenging research field in recent years.Although the wavelet transform has been widely used in texture classification, thetraditional real wavelet transform has great limitations with shift-variance, poor spatialorientations and the lack of phase information. However, the complex wavelet transformcould overcome the shortcomings of real wavelet transform. Thus, we are mainly focus on thedynamic texture classification algorithms in complex wavelet transform field as follows:1. Dynamic texture classification based on dual-tree complex wavelet transform. Thedistribution of phase components is approximate uniform in all subbands of complex wavelettransforms. If the phase information of the complex wavelet subband is used directly, thegood performance could not be display sufficiently in texture classification. And this paperproposes a new dynamic texture feature based on the magnitude information and the phaseinformations in dual-tree complex wavelet transform domain.We introduce the probabilitydensity function of the magnitude of complex wavelet coefficients for dynamic textureclassification where both the real and imaginary parts of each subband are characterized bythe generalized Gaussian distribution model. Then, we demonstrate the Von Mises distributionfits well with the relative phase informations of complex wavelet coefficients from dynamictexture. The real univariate generalized Gaussian distribution is always employed to modelthe real wavelet coefficients. Thus we present a fully-complex distribution denoted ascomplex generalized Gaussian distribution, and the model parameters serve as the texturefeature for the dynamic texture classification. To evaluate the classification performance ofthe proposed feature, we utilize two classifiers, k-nearest neighbor (kNN) classifier and support vector machines (SVMs), in our experiments. The experimental results demonstratethe superior performance of the new texture features compared with the current existingapproaches.2. Dynamic texture classification based on spatio-temporal complex wavelet transform.Firstly, a new spatio-temporal complex wavelet transform is proposed in the paper. Itmaintains the excellent performance of the spatio-temporal real wavelet transform, andapproximate shift invariance and good directional selectivity properties make it an excellentcandidate for representing the dynamic texture features. Secondly, we develop thespatio-temporal complex wavelet transform to extract the motion information of dynamictexture, and then take the Weibull distribution parameters of each subband as the texturefeature for dynamic texture classification. Finally, we break each three-dimensional detailsubband of the spatio-temporal complex wavelet into some subblocks and modle the medianmagnitudes of them with the Gumbel distribution. The distribution parameters are estimatedto form the dynamic texture feature. The experimental results demonstrate the superiorperformance of the new texture features compared with the method base on the mean andvariance.