基于人工免疫协同理论和多目标核优化的SAR图像分割框架及关键技术

The satellites with high resolution, multi-polarizations, multi-modes and large bandwidths are greatly proliferated in recent years. They can supply large amounts of observed data in a steady stream and carry out continuous multiscale, omni-directional and multi-spectral observations of the earth. However, the difficulty existing in SAR image processing systems is how to improve their ability of intelligence and precision facing such large image data. As far as we know, image segmentation can discover the general structural information of the image and reveal the essential content and target property. The project is to aim at removing speckle noise in SAR image, discovering the number of land-covers automatically and fast partitioning model of SAR images. The statistical model of speckle noise in SAR images and the performance of the hot studied nonlocal means (NLM) algorithm in this type of noise are analyzed. The project will proposed a union filter framework by combing the maximum likelihood estimation at homogeneous region and fine NLM filter at heterogeneous region of the image. To improve the efficiency of the segmentation model, the project will present SAR image segmentation algorithm by utilizing the theory of artificial immune system (AIS) and co-evolution evolutionary principals. Besides, the searching strategy in AIS with histogram statistics and gene jumping mechanism to discover the clustering number of SAR image also is discussed. Furthermore, the multi-objective SAR image segmentation framework in AIS theory and kernel clustering index is also studied to segment the above despeckling image and discover the pixels distribution model. Finally, the ultimate goal the project is to study and present a feasible fast remote sensing image segmentation framework.

新型的高分辨率、多极化、多模式、大带宽卫星传感器大量涌现，源源不断提供丰富的数据，实现对地球多尺度、全方位、多谱段地连续观测。对现有遥感数据进行智能化和精确化处理是难点问题之一，图像分割可以发现图像的整体结构信息，揭示图像本质内容和目标属性。本课题以SAR图像分割中的斑点噪声抑制、地物类别自动划分、快速分割模型为设计目标，研究SAR图像斑点噪声统计模型，分析最新非局部统计算法在复杂相干斑噪声模型的性能，构建SAR相干斑强度估计算法，提出融合同质区域最大似然估计和异质区域精确化滤波的联合滤波框架。为了实现快速和精确的SAR图像分割，研究最新的协同进化理论模型，构建适合SAR图像分割的人工免疫协同优化模型，设计图像直方图域和基因跳变的地物类别数目自动发现策略。同时，为了使算法能够精确稳定地发现SAR图像像素分布模型，研究多目标核优化SAR图像分割理论，开发出切实可行的快速遥感图像自动分割框架。

课题借鉴SAR图像斑点噪声抑制、多目标优化和协同进化的最新理论和模型，旨在构造出一种计算时间短、分割精度高、通用型SAR图像分割框架，对于SAR图像分割涉及到的斑点噪声强度估计、自适应主成分分析与图像重建、快速非局部均值滤波、人工免疫协同进化理论、多目标核聚类技术等展开深入而系统的研究。.首先，依据SAR斑点噪声统计分布模型推导，在同质区域内，其观测图像的方差均值比统计系数与相干斑的标准差是相等的，为此，我们采用直方图统计分布来估计相干斑的强度，然后来自适应地选择图像分解中主成份数量，完成基于相干斑强度自适应重建的核图划分集成算法。其次，为了提升传统非局部均值算法在SAR图像斑点噪声抑制性能，课题组采用最大似然统计和直方图划分的策略来区分图像同质区域和异质区域，然后仅在异质区域进行快速非局部均值强化滤波。最后，课题组研究了分水岭粗分割、人工免疫多目标核聚类的两阶段分割框架，在图像多尺度分割标记、人工免疫多目标优化、核聚类指标选择均进行了的研究。.课题组在研究期内，发表论文11篇（其中SCI收录5篇，EI收录6篇）；申请软件著作权2项。对照立项时制定的客观目标，已全面保质保量完成。同时，课题负责人获得了2016年度“陕西省青年科技新星”称号，课题组成员获得“陕西省优秀博士论文”。