摘要:
高空间分辨率(简称"高分")SAR图像具有高维非线性特点,以高维空间蕴含的低维流形描述SAR图像,会更有利于目标识别.将流形学习应用到高维SAR目标识别的特征表达中,提出一种新的高分SAR图像建筑区提取方法.首先,对高分SAR图像进行预处理;然后,采用灰度共生矩阵(gray level co-occurrence matrix,GLCM)提取8种纹理特征,与灰度图像共同构建SAR图像的高维特征集;利用自适应邻域选择的邻域保持嵌入(adaptive neighborhood selection neighborhoods preserving embedding,ANSNPE)算法对高维特征集进行特征提取,提取出新的特征;最后,通过阈值分割及后处理提取建筑区,并进行精度评价.选择TerraSAR-X数据进行实验研究,结果表明,ANSNPE算法能够从高分SAR图像中有效提取建筑区,并具有较强的泛化能力;通过训练数据获得的投影矩阵可直接应用到新样本中,建筑区提取精度达85%以上.
Abstract:
The characteristics of high resolution SAR image is nonlinear and of high dimension. The description of SAR image in which a low dimensional manifold is embedded in high dimensional space is more useful for targets recognition. Therefore, a novel scheme of high resolution SAR image building area extraction is proposed by applying manifold learning to feature representation of a high dimensional SAR targets recognition. Firstly, the high resolution SAR image was preprocessed, and then eight texture features were extracted with gray level co -occurrence matrix ( GLCM ) so as to construct feature set with gray feature. Adaptive neighborhood selection neighborhood preserving embedding ( ANSNPE) algorithm was used to extract the new features from the feature set. Finally, the building area was extracted by threshold segmentation with the new features and post processing, and the accuracy was evaluated. Selecting TerraSAR -X as test data, the authors carried out the experiments. The results show that ANSNPE algorithm can effectively extract the building area from high resolution SAR image, and has strong generalization capability. The projection matrix obtained through the training data can be directly applied to the new samples, and the accuracy of building area extraction could reach higher than 85%.
下载: