基于上下文分析的无监督分层迭代算法用于SAR图像分割

发布时间：2019-04-23 13:37

【摘要】：基于聚类的分割算法能够有效地分析目标特征在特征域的分布结构,进而准确判断目标的所属类别,但难以利用图像的空间和边缘信息,而基于区域增长的分割算法能够在空间域利用多种图像信息计算目标之间的相似性,但缺乏对特征结构本身的深层挖掘,容易出现欠分割或过分割的结果.本文结合这两种算法各自的优势,针对合成孔径雷达(Synthetic aperture radar,SAR)图像的特点,提出了一种基于上下文分析的无监督分层迭代算法.该算法使用过分割区域作为操作单元,以提高分割速度,降低SAR图像相干斑噪声的影响.在合并过分割区域时,该算法采用了分层迭代的策略:首先,设计了一种改进的模糊C均值聚类算法,对过分割区域的外观特征进行聚类分析,获得其类别标记,该类别标记包含了特征的分布结构信息.然后,利用多种SAR图像特征对同类区域的空域上下文进行分析,使用区域迭代增长算法对全局范围内的相似区域进行合并,直到不存在满足合并条件的过分割区域对为止,再重新执行聚类算法.这两种子算法分层交替迭代,扬长避短,实现了一种有效的方法来组织和利用多种信息对SAR图像进行分割.对模拟和真实SAR图像的实验表明,本文提出的算法能够在区域一致性和细节保留之间做到很好的平衡,准确地分割出各类目标区域,对相干斑噪声具有很强的鲁棒性.
[Abstract]:The clustering-based segmentation algorithm can effectively analyze the distribution structure of the target feature in the feature domain, and then accurately judge the classification of the target, but it is difficult to utilize the spatial and edge information of the image. The segmentation algorithm based on region growth can calculate the similarity between objects in spatial domain by using a variety of image information, but it lacks deep mining of the feature structure itself, so it is easy to produce the result of under-segmentation or over-segmentation. According to the characteristics of synthetic aperture radar (Synthetic aperture radar,SAR) images, an unsupervised hierarchical iterative algorithm based on context analysis is proposed in this paper, which combines the advantages of the two algorithms. In this algorithm, the over-segmented region is used as the operating unit to improve the segmentation speed and reduce the influence of speckle noise on the SAR image. In the process of merging over-cut regions, the algorithm adopts hierarchical iterative strategy. Firstly, an improved fuzzy C-means clustering algorithm is designed to cluster the appearance features of over-segmented regions and obtain its category markers. The class tag contains information about the distribution structure of the feature. Then, the spatial context of the same region is analyzed by using a variety of SAR image features, and the region iterative growth algorithm is used to merge the similar regions in the global scope until there is no over-cut region pair which satisfies the merging condition. Then re-execute the clustering algorithm. These two sub-algorithms are layered and iterated alternately in order to realize an effective method to organize and segment SAR images by using many kinds of information. Experiments on simulated and real SAR images show that the proposed algorithm can achieve a good balance between region consistency and detail retention, accurately segment all kinds of target regions, and has strong robustness to speckle noise.
【作者单位】：西安电子科技大学智能感知与图像理解教育部重点实验室;
【基金】：国家重点基础研究发展计划(973计划)(2013CB329402) 国家自然科学基金(61072106,61173092,61271302,61272282,61001206,61202176,61271298) 国家教育部博士点基金(20100203120005) 教育部长江学者和创新团队支持计划(IRT1170)资助~~
【分类号】：TN957.52

【共引文献】