点特征相似与卷积神经网络相结合的SAR图像分类算法研究

Research on SAR Image Classification Based on Point Feature Similarity and Convolutional Neural Network

基于CNN的像素级SAR图像分类利用了输入图像块的邻域信息,但没有凸显出邻域像元对中心像元分类结果的影响力,导致在高噪声条件下中心像元易出现类别误判。针对该问题,该文提出了一种基于点特征相似性的卷积神经网络(Point feature Similarity-based Convolutional Neural Network,PSCNN),并将其用于SAR图像分类,以凸显邻域像元对中心像元分类结果的影响力,从而减小误分,提升分类精度。实验结果表明,相比传统基于CNN的SAR图像分类方法,该算法一方面能更充分利用图像块的邻域信息,有效抑制相干斑的影响,提升匀质区域的分类精度;另一方面借助块匹配方式,能够充分保留图像块的结构信息,有效提升边界定位精度。

Pixel-level SAR image classification based on CNN utilizes the neighborhood information of the input image patches,but it does not highlight the influence of neighborhood pixels on the classification result of central pixels,which leads to the misclassification of central pixels under high noise conditions.This paper proposed a novel point feature similarity-based convolutional neural network(PSCNN) for SAR image to improve classification accuracy.Firstly,the image window distributions between central pixel and neighboring pixels are extracted by non-local mean block matching.Secondly,the similarity between neighborhood pixels and center pixel is established by comparing the distributions of image window,and then the weighting factors of neighborhood pixels are generated based on their similarity.Finally,the weighted image patches are sent to the CNN network for feature extraction and classification.Compared with the traditional CNN-based SAR image classification,the proposed method can make full use of neighborhood information,so as to suppress the influence of speckles more effectively and improve the classification accuracy of homogeneous regions.On the other hand,it can fully retain the structural information of the image patches by means of block matching to improve the positioning accuracy of the boundary.The experimental results demonstrate the higher overall classification accuracy and boundary positioning accuracy of the proposed method comparing with the traditional methods.