基于尺度间上下文关系模型的动态纹理分割

发布时间：2018-09-17 12:50

【摘要】：动态纹理是由空间重复且随时间变化的一系列的图像序列组成,并且在空时域上具有某种自相似性。动态纹理分析在很多领域表现出潜在的应用前景,动态纹理分割作为动态纹理研究的关键问题之一,得到越来越多的关注,促使动态纹理的研究成为了一个热点问题。动态纹理分割就是将自然纹理的图像序列分割成互不相叠的若干区域,并且不同的区域有不同的纹理,同一区域内纹理表现出均匀一致性。尺度间上下文关系可以充分利用不同尺度间标号之间的关系刻画动态纹理表现出的“运动”和“外观”特征。因此,本文提出基于尺度间上下文关系模型的方法进行动态纹理分割。本文的主要工作如下:1.提出了基于小波域马尔可夫链的上下文关系的动态纹理分割算法。对动态纹理进行小波变换后同一帧图像尺度内各子带间以及相邻尺度子带间存在较强的依赖关系,这种关系可以用于提高动态纹理的刻画能力。本算法的标记场模型采用尺度间因果马尔可夫随机场模型和尺度内的非因果马尔可夫随机场模型,特征场建模采用高斯马尔可夫随机场模型进行建模,通过邻域交互参数矩阵考虑了每一小波系数向量与同一尺度上相邻小波系数向量之间的作用关系。实验结果表明该算法可以很好的实现动态纹理分割。2.提出了基于马尔可夫随机场能量的上下文关系的动态纹理分割算法。根据空时邻域系统和多尺度随机场模型确定标记场的邻域系统和能量函数。利用高斯分布描述观测场,形成基于多尺度随机场模型的马尔可夫随机场的动态纹理分割方法,运用最大后验准则分割动态纹理。最后,将仿真结果与现有模型算法的仿真结果进行对比,得到本方法具有更好的分割结果。
[Abstract]:Dynamic texture is composed of a series of image sequences which are repeated in space and change with time, and have some self-similarity in space-time domain. Dynamic texture analysis has a potential application prospect in many fields. As one of the key issues in dynamic texture research, dynamic texture segmentation has attracted more and more attention, which makes the research of dynamic texture become a hot issue. Dynamic texture segmentation is to segment natural texture image sequences into several regions that are not overlapped with each other, and different regions have different textures, and the texture in the same region shows uniform consistency. The contextual relationship between scales can make full use of the relationships between different scales to characterize the "motion" and "appearance" of dynamic textures. Therefore, this paper proposes a dynamic texture segmentation method based on the context relation model between scales. The main work of this paper is as follows: 1. A dynamic texture segmentation algorithm based on context relation of Markov chain in wavelet domain is proposed. After the dynamic texture is transformed by wavelet transform, there are strong dependencies between subbands and adjacent sub-bands in the same frame image scale, which can be used to improve the performance of dynamic texture characterization. The mark-up field model of this algorithm adopts the inter-scale causal Markov random field model and the non-causal Markov random field model in the scale, and the Gao Si Markov random field model is used to model the characteristic field. The interaction relationship between each wavelet coefficient vector and adjacent wavelet coefficient vector on the same scale is considered by neighborhood interaction parameter matrix. Experimental results show that the algorithm can achieve dynamic texture segmentation. 2. 2. A dynamic texture segmentation algorithm based on the context relation of Markov random field energy is proposed. Based on the space-time neighborhood system and the multi-scale random field model, the neighborhood system and energy function of the label field are determined. The dynamic texture segmentation method of Markov random field based on multi-scale random field model is formed by using Gao Si distribution to describe the observation field, and the dynamic texture is segmented by maximum a posteriori criterion. Finally, the simulation results are compared with those of the existing model algorithms, and better segmentation results are obtained.
【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】