基于小波域HMT模型的动态纹理分割方法研究

发布时间：2018-07-05 17:26

本文选题：LMM-HMT + 动态纹理分割　；参考：《哈尔滨工程大学》2016年硕士论文

【摘要】：纹理分析是图像处理领域的基础性研究课题。动态纹理是纹理的一种类别,它作为纹理分析研究的重要内容,近年来已成为计算机视觉和数字图像处理领域的研究热点,在军事、工业、医学、智能交通、气象遥感、公共安全等领域有着广泛的应用价值。多尺度分析方法也即多分辨率分析,是近年新兴起的一种信号分析技术,其可从不同尺度的角度描述景物,利用不同尺度分析问题、解决问题,更加有效地提取图像或视频特征并进行分析,为信号分析提供了一个全新的视角研究课题。小波变换作为多尺度分析的主要基础方法,其具有良好的时频分析特性,它能通过变换以充分突出问题某些方面的特性。马尔可夫理论作为描述局部统计分布的强有力工具,能有效地与小波域空时变换特性相结合,推动了小波域模型的发展,对准确刻画信号非平稳特性起着至关重要的作用。小波域隐马尔可夫(HMT)模型充分利用小波系数尺度间和尺度内的相关性,在纹理分析等方面表现出了出色的性能。基于以上理论基础,本文主要研究基于小波域隐马尔可夫模型的动态纹理分割算法,主要工作如下:1.提出三维小波域LMM-HMT模型。将HMT模型推广到三维空时域,改进已有的EM算法以进行参数估计,并改进分割过程以适应三维动态纹理的处理。在训练及粗分割阶段,结合纹理图像自身小波系数的分布规律,利用拉普拉斯分布尖峰厚尾的分布特性,对纹理图像的小波系数分布进行匹配,提出LMM-HMT模型及与之相对应的参数估计算法公式。在多尺度融合阶段,提出基于26邻域背景向量的多尺度融合方法。利用动态纹理帧间相互依赖关系,建立一个基于三帧纹理图像的26邻域背景向量模型。在像素级分割阶段,再次使用混合拉普拉斯分布模型,并给出参数估计算法,提高其最终分割效果。2.提出三维小波域HMT-3S模型。通过连接三维小波域七个细节子带的小波系数,建立三维小波域上的HMT-3S模型,并提出对应的模型参数训练方法、似然函数计算及多尺度融合算法。将以上对基于小波域HMT模型的动态纹理分割的改进算法应用到HMT-3S模型中,与HMT模型进行对比,使分割效果得到进一步提升。
[Abstract]:Texture analysis is a basic research topic in the field of image processing. Dynamic texture is a kind of texture. As an important part of texture analysis, dynamic texture has become a hotspot in the field of computer vision and digital image processing in recent years. It has been widely used in military, industrial, medical, intelligent transportation and meteorological remote sensing. Public safety and other fields have extensive application value. Multiresolution analysis, or Multiresolution Analysis, is a new signal analysis technique in recent years. It can describe the scene from different scales and use different scales to solve the problems. More efficient extraction and analysis of image or video features provide a new perspective for signal analysis. Wavelet transform, as the main basic method of multi-scale analysis, has good time-frequency analysis characteristics, and it can fully highlight some aspects of the problem by means of transformation. As a powerful tool to describe the local statistical distribution, Markov theory can effectively combine with the space-time transform characteristics of wavelet domain, promote the development of wavelet domain model, and play an important role in accurately characterizing the non-stationary characteristics of signals. The hidden Markov (HMT) model in wavelet domain makes full use of the correlation between scales and scales of wavelet coefficients and has shown excellent performance in texture analysis. Based on the above theoretical basis, this paper mainly studies the dynamic texture segmentation algorithm based on wavelet domain hidden Markov model, the main work is as follows: 1. A three dimensional wavelet domain LMM-HMT model is presented. The HMT model is extended to 3D space-time domain, the existing EM algorithm is improved to estimate the parameters, and the segmentation process is improved to adapt to the 3D dynamic texture processing. In the stage of training and rough segmentation, the distribution of wavelet coefficients of texture images is matched by using the distribution characteristics of Laplacian distribution of peak and thick tail, combining with the distribution of wavelet coefficients of texture images. The LMM-HMT model and the corresponding formula of parameter estimation algorithm are presented. In the multi-scale fusion stage, a multi-scale fusion method based on 26 neighborhood background vector is proposed. In this paper, a 26 neighborhood background vector model based on three frames texture images is established by using dynamic texture frame interdependence. In the pixel level segmentation stage, the mixed Laplace distribution model is used again, and the parameter estimation algorithm is given to improve the final segmentation effect. A three dimensional wavelet domain HMT-3s model is proposed. The HMT-3S model in 3D wavelet domain is established by connecting the wavelet coefficients of seven subbands of detail in 3D wavelet domain. The corresponding training method of model parameters, likelihood function calculation and multi-scale fusion algorithm are proposed. The improved algorithm of dynamic texture segmentation based on wavelet domain HMT model is applied to HMT-3S model, and compared with HMT model, the segmentation effect is further improved.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】