高斯混合模型(GMMs)是统计学习理论的基本模型,在可视媒体领域应用广泛。近些年来,随着可视媒体信息的增长和分析技术的深入,GMMs在(纹理)图像分割、视频分析、图像配准、聚类等领域有了进一步的发展。从GMMs的基本模型出发,从理论和应用的角度讨论和分析了GMMs的求解算法,包括EM算法、变化形式等,论述了GMMs的模型选择问题:在线学习和模型约简。在视觉应用领域,介绍了GMMs在图像分段、视频分析、图像配准、图像降噪等领域的扩展模型与方法,详细地阐述了一些最新的典型模型的原理与过程,如用于图像分段的空间约束GMMs、图像配准中的关联点漂移算法。最后,讨论了一些潜在的发展方向与存在的困难问题。

Gaussian Mixture Models(GMMs) is the basic model of statistical machine learning and widely applied to visual media fields. In recently years, with the rapid growth of visual media information and deep development of analytical techniques GMMs have obtained further developments in such fields as (texture) image segmentation, video analysis, image registration and clustering. This paper begins from the basic models of GMMs, discusses and analyzes from both theoretical and application aspects the solving methods of GMMs including EM algorithms and its variants, and expounds the two problems of model selection: online learning and model reduction. In visual applications, this paper introduces GMM-based models and methods in image segmentation, video analysis, image registration and image de-noising, expatiates the principles and processes of some newest and classical models, such as space-variant GMMs for image segmentation, coherent point draft algorithm for image registration. At last, this paper gives some possible latent directions and difficult problems.