面向SAR图像变化检测的稀疏判别学习模型与方法研究

The synthetic aperture radar (SAR) image is an active detection system with long distance, all-weather and timeless imaging capacity, which has been widely used in a variety of civil and military areas. As one typical task of SAR image interpretation and understanding, the change detection has gains widespread application in various civilian and military fields. However, due to the coherence characteristics of imaging system, the quality of SAR image is degraded serious by speckle. Therefore, it is a difficult task to capture accurate and discriminant features in the change detection task. To handle this problem, this project focuses on learning discriminant sparse model and method for SAR images based on the research results in the theory of sparse representation and visual cognition. Specifically, the key research points of this project are listed as follows. First, a novel similarity denoising model of SAR image is provided by analyzing the statistical characteristics of noising. Second, a discriminant feature extracting approach is proposed by applying the sparse coding. Apart from this, the non-negative matrix factorization plays an important role in analyzing the form of data in SAR images. Finally, based on the image saliency similarity, the priori information of the changed parts and the unchanged parts can be obtained from the SAR images. Combining with the sparse representation theory and the classifier algorithm, the accurate and discriminant feature could be learned to realize very precise SAR images change detection. The study of this project aims to enrich the theory systems of feature extraction and learning for SAR image. With the learnt discriminant features, difficulties of change detection could be overcome effectively, which are caused by SAR imaging system. With the implementation of this project, it expects to provide a solid theoretical basis and feasible methods for discriminant feature extracting of SAR images in practical application.

受成像系统相干特性的影响，SAR图像中存在大量的相干斑噪声，这使得在变化检测中很难获得区域的判别性特征表示。针对该问题，结合稀疏表示理论和视觉认知机理的研究成果，将重点研究SAR图像变化检测的判别稀疏模型和方法。本项目拟开展以下三个方面的研究工作：（1）综合分析噪声的统计特性，建立SAR图像的相似性降噪模型；（2）结合图像的数据特性，基于非负稀疏编码方法，对SAR图像中的判别性特征进行提取；（3）利用视觉认知机理挖掘SAR图像的先验信息，结合稀疏表示理论和分类器算法，研究提取SAR图像判别性特征的学习方法，提高SAR图像变化检测的精度。该项目的研究旨在丰富SAR图像特征提取及学习方法的理论体系和技术手段，克服成像特性对SAR图像变化检测所带来的困难，实现对SAR图像中地物变化的准确检测。通过本项目的实施，预期为SAR图像变化检测的实际应用提供坚实的理论依据和可行性的方法。

在SAR图像的处理领域中，变化检测是一个十分重要的研究课题。本项目结合计算机视觉和机器学习领域的理论及技术手段，对SAR图像的变化检测任务进行系统深入的研究。通过研究噪声块相似性度量模型，建立了一种抑制SAR图像中相干斑噪声的模型，有效抑制相干斑噪声对SAR图像处理造成干扰的问题。通过研究SAR图像判别特征的学习方法，本项目提出了一种基于稀疏表示和胶囊网络的SAR图像变化检测方法。实验验证了稀疏判别特征有效降低图像中斑点噪声对变化检测准确率的影响。为了提高检测的准确性构建了一个轻量级的胶囊网络模型，该模型能够挖掘特征之间的空间信息，训练好的网络可以实现对变化和未变化区域的精确检测。受视觉注意机制的启发，设计了一种显著性区域检测策略，为了挖掘检测图像的先验信息，提出了一种基于显著性检测和注意力胶囊网络的SAR图像变化检测方法。通过空间注意力机制提取感兴趣的特征，胶囊网络实现精确地分类。项目研究从SAR图像中提取更有效的特征来训练网络模型，提高网络模型的分类能力，提出了一种基于U-Net和胶囊网络的SAR图像变化检测方法。选取训练样本时，只在显著性区域内选择，这样可以平衡变化类样本和未变化类样本的数量，大大提高网络训练的效率。结合非负矩阵分解理论，本项目提出了一种基于稀疏非负矩阵分解和极限学习机的SAR图像变化检测方法。充分考虑差异图的邻域空间信息，构造了基于稀疏非负矩阵分解和分层模糊C均值的预分类模型，选择极限学习机作为分类模型学习两时相SAR图像中的非线性关系，通过实验验证了所提方法实现了对SAR图像中变化信息的准确检测。