离散ADMM方法下像素与对象基元协同优化的遥感影像无监督语义分割

语义分割是遥感影像分析中的重要技术之一。现有的方法（如基于深度卷积神经网络的方法等）虽然在语义分割中取得了显著进展，但往往需要大量训练数据。基于图模型的马尔可夫随机场模型（Markov random field model，MRF）提出了一种不依赖训练数据的无监督语义分割思路，可以有效地刻画地物空间关系，并对地物空间分布的统计规律进行建模。但现有的MRF模型方法通常建立在基于像素或对象的单一粒度基元上，难以充分利用影像信息，语义分割效果不佳。针对上述问题，引入交替方向乘子法 （alternative direction method of multiplier，ADMM）并将其离散化，提出了一种像素与对象基元协同的MRF模型无监督语义分割方法（MRF-ADMM）。首先构建像素基元和对象基元两个概率图，其中像素基元概率图用于刻画影像的细节信息，保持语义分割的边界；对象基元概率图用于描述较大范围的空间关系，以应对遥感影像地物内部的高异质性，使分割结果中地物内部具有良好的区域完整性。在模型求解过程中，针对像素和对象基元的特点，提出了一种离散化的ADMM方法，并将其用于两种基元类别标记的传递与更新，实现像素基元细节信息和对象基元区域信息的协同优化。高分二号和航拍影像等不同数据库不同类型遥感影像的语义分割实验结果表明，相较于现有的MRF模型，提出的MRF-ADMM方法能有效地协同不同粒度基元的优点，优化语义分割结果。

Unsupervised semantic segmentation of remote sensing image based on collaborative optimization of multigranularity primitives under discrete ADMM method

Semantic segmentation is one of the important techniques in remote sensing image analysis. Existing methods, such as methods based on deep convolutional neural networks, etc., have made significant progress in semantic segmentation, but these methods often require a large amount of training data. The Markov random field model(MRF) based on the graph model proposed an idea of unsupervised semantic segmentation that did not rely on training data, which could effectively describe the spatial relationship of objects, and analyze the spatial distribution of objects. However, the existing MRF model methods were usually based on a single granularity primitive based on pixels or objects, and it is difficult to make full use of image information, resulting in poor semantic segmentation. Aiming AT the above problems, this paper introduced the alternative direction method of multipliers(ADMM) and discretized it, then proposed an unsupervised semantic segmentation method(MRF-ADMM) based on the pixel and object primitives collaborative MRF model. Firstly, it constructed two probability maps of pixel primitive and object primitive, in which the pixel primitive probability map was used to describe the detailed information of the image and maintain the boundary of semantic segmentation; used the object primitive probability map is to describe a large range of spatial relationships and deal with the high heterogeneity inside the remote sensing images. In the process of model solving, according to the characteristics of pixels and object primitives, it proposed a discretized ADMM method, and used it to transfer and update of the two primitive category labels. Compared with the existing MRF models, the experimental results of semantic segmentation of different types of remote sensing images in different databases such as Gaofen-2 and aerial images can effectively synergize the advantages of different granularity primitives and optimize semantics results.