基于共性视觉特征与反馈机制的SAR图像目标检测方法研究

How to use visual attention mechanism to detect objects quickly and efficiently from a mass of SAR data has become one of the key problems in SAR image processing, which is very important to alleviate the contradiction between high-speed acquisition and low-speed interpretation of SAR images. Most existing methods only take single-image visual feature into account. Hence, they can not effectively suppress the interference caused by the salient clutter in the background, neither can they rule out SAR images without objects of interest. To solve the above problems, this project simultaneously analyses multiple SAR images with similar target characteristics. Firstly, we combine superpixel segmentation with statistical analysis of visual characteristics, to construct a single-image visual saliency feature analysis model with strong anti-speckle noise capability, and the initial target areas are thus extracted effectively. Then, the common visual saliency features of these regions are fully explored and integrated with the target similarity based division strategy, to construct a feedback mechanism on the basis of feature similarity measurement. Such mechanism can sufficiently suppress the interference caused by the salient clutter, which facilitates not only fast and efficient target detection in the SAR images, but also effective elimination of SAR images that do not contain common targets. Lastly, the project will establish a set of geometric characteristic based evaluation indices, to achieve accurate and valid assessment of target detection in the SAR images. The above research will offer new research ideas for accurate and efficient target detection in SAR images. The related results will provide important theoretical and technical supports for the automatic SAR image interpretation technology.

如何利用视觉注意机制快速、高效地从海量SAR数据中检测目标已成为SAR图像处理领域亟待解决的关键问题之一，对缓解SAR图像高速获取与低速解译间的矛盾具有重要意义。针对现有方法仅考虑单幅图像视觉特征，难以抑制背景中显著性较高的杂波干扰，无法排除无感兴趣目标的SAR图像等问题，本项目拟从具有相似目标特征的多幅SAR图像入手，首先结合超像素分割与视觉特征统计分析，构建具有抗相干斑噪声能力的单图视觉显著特征分析模型，有效提取初始目标区域；然后充分发掘这些区域的共性视觉特征，结合目标相似度划分构建基于特征相似性度量的反馈机制，充分抑制显著性较高的杂波干扰，实现SAR图像目标的快速、高效检测并将无共性目标图像有效排除；最后建立基于几何特征的评价指标，实现对SAR图像目标检测方法的有效评估。上述研究将为准确高效地检测SAR图像目标开辟新的研究思路，相关成果将为SAR图像自动解译提供重要的理论与技术支持。

本项目针对传统SAR图像目标检测方法仅考虑单幅图像的视觉特征，难以抑制背景中显著性较高的杂波干扰，无法有效排除那些无感兴趣区域或目标的SAR图像等问题，从多幅SAR图像的共性视觉特征入手，结合反馈机制构建了新的多幅SAR图像共性感兴趣目标提取策略。本项目的主要研究工作可以概括为如下四个方面：1、针对单幅SAR图像的视觉显著性分析问题，项目组先后提出基于多维特征向量聚类与主动轮廓模型的SAR图像视觉显著性分析与油库目标精确检测方法，基于一致增强扩散模型的SAR图像显著性分析与感兴趣目标检测方法，基于统计显著性分析与测地线模型的单幅SAR图像油库目标检测算法，基于幅度谱聚类与背景先验显著性分析模型的SAR图像居民区检测方法，上述算法基本实现了单幅SAR图像中感兴趣区域或目标的有效检测与提取；2、针对具有相似地物特征的一组SAR图像，项目组构建了基于共性视觉特征与灰度共生直方图的SAR图像联合显著性分析与感兴趣区域检测方法，从而有效弥补了由真实颜色信息缺失所引起的SAR图像目标检测精度下降问题；3、项目组结合目标相似度划分策略与反馈机制，进一步增强待检测目标的共性视觉显著特征，充分抑制背景中显著性较高的杂波干扰，最终完成了多幅SAR图像共性感兴趣目标的准确、高效检测，同时也实现了不含共性目标SAR图像的有效排除；4、项目组在上述工作基础上，进一步建立了基于几何特征的评价指标，从而实现了对SAR图像显著性分析与目标检测方法的有效评估。项目组的主要研究工作为SAR图像目标的准确高效检测开辟了新的研究思路，所取得的理论成果不仅对缓解SAR图像高速获取与低速解译间的矛盾具有重要意义，而且也为SAR图像的自动目标提取与识别提供了重要的理论依据与技术支持。所提出的主要算法与申请的发明专利在舰船目标提取、港口油库区建设以及城乡土地利用监测等诸多SAR图像应用领域都具有重要的实际应用价值。