基于特征量统计与联合稀疏迁移学习的极化SAR异质场景分类

项目名称： 基于特征量统计与联合稀疏迁移学习的极化SAR异质场景分类

项目编号： No.61472306

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 缑水平

作者单位： 西安电子科技大学

项目金额： 81万元

中文摘要： 针对极化SAR图像中多通道、强噪声和高分辨率等原因引起的混合散射体异质场景难以准确分类问题，采用局部高斯概率模型对特征空间的特征值、功率和特征向量等的分布进行动态统计，建立一种对极化SAR图像中纹理和噪声的自适应非高斯统计模型，该模型能够随着图像空间变化而变化地表征散射特征的动态参数；进而对于极大似然估计的偏向问题实施参数迁移，以降低估计偏差对分类的影响；通过对三个通道数据联合稀疏表示，自动提取出混合散射体的本质特征；针对不同时相极化SAR数据将联合稀疏与迁移学习结合，充分利用已有先验信息进一步提高迁移学习方法的泛化能力，并将其用于散射回波随机波动的极化SAR图像异质场景高效分类。期望在解决不满足独立同分布且泛化能力差的机器学习方法及其应用于极化SAR图像的异质场景分类等问题上，有实质性进展，研究成果可应用于军事、农业和地址灾害损毁评估等相关领域，具有较好的理论和应用价值。

中文关键词： 极化SAR分类；异质场景；特征统计；联合稀疏；迁移学习

英文摘要： In view of the low classification accuracy for the clutter of heterogeneous scene which is caused by multi-channel, speckle noise and high-resolution of polarimetric SAR images, and an adaptive Statistical non-Gaussian model is made for texture and speckle of polarimetric SAR images. This model is made on the features space using local Gassian distribution and can represents dynamic parameters of scatter object feature because it changes with the image space movement. Another, parameter transfer learning is used to correct the bias of maximum-likelihood estimator in order to reduce the effect on the classification. And then, essential characteristic of the clutter is extracted automatically by using Joint Sparsity representation for three channels data. Furthermore, the generalization ability of the transfer learning is improved by combining joint sparse representation and transfer learning, and this learning style is used to heterogeneous scene classification of polarimetric SAR images with random scatter waves. We expect there have substantive progress in the solution of unsatisfied generalization ability due to non-independent identity distribution and heterogeneous scene classification for polarimetric SAR images. The research achievements could be applied in military, civil and so on related domain and it has good theory and the application value.

英文关键词： Polarimetric SAR Classification;Heterogeneous Scene;Feature Statistical;Joint Sparsity;Transfer Learning