中文摘要：

针对传统特征提取方法不能提取目标高层结构特征的问题,提出了一种基于软概率的池化方法,结合多层反卷积网络,学习目标的高层结构特征,并将其用于合成孔径雷达（SAR）图像分类。首先对SAR图像进行子块划分,然后对每个子块进行基于多层反卷积网络的特征编码,学习出不同层次上的图像特征,最后将该特征用于支持向量机（SVM）分类器,实现SAR图像的分类。在国内首批SAR数据上的实验表明,该算法获得了较高的分类准确率。

英文摘要：

Aim at the problem that the traditional feature extraction methods cannot get the high level structure features,this paper put forward a new soft probability pooling method,which is used in multilayer Deconvolutional Network,then high level structure features can be learned and be used for classification of SAR image.Firstly,the SAR image was divided into patches;then,the feature coding of each patch was obtained by means of multilayer Deconvolutional Networks,which can learn features suitable for image classification in different scale;finally,the SAR image was classified through the features used in SVM classifier.Experimental results on the first batch domestic PolSAR images show that the classification accuracy rate of the proposed algorithm is superior.