小训练样本下的合成孔径雷达图像分类研究

针对纹理是合成孔径雷达(SAR)图像目标分类的一个重要因素,SAR图像的过完全小波分解产生大小不变的子图像,具有移不变特性,可在不同尺度下表征纹理。利用图像灰度均值与细节图像能量特征组成特征矢量,对SAR图像有好的表征效果。与完全由图像分解子图能量得到的特征矢量相比,目标间的纹理特征差异更明显。神经网络具有高度非线性判决性能,可将所提出的过完全小波分解纹理能量特征(OWATF)与径向基函数(RBF)神经网络相结合对SAR图像面目标进行分类。实验证明,在小训练样本条件下,RBF神经网络与OWATF特征相结合对SAR图像进行分类能够很好地体现目标的整体特性。

Synthetic aperture radar image classification with a few training samples

In synthetic aperture radar image, texture is an important feature. Transformed SAR images by over-complete wavelet analysis are of the same size as the original ones, and over-complete wavelet analysis is shift-invariant. OWATF vectors are composed of mean gray level of the image and texture features of detail images. Differences of three objects' vectors in texture feature are much larger than those composed of sub-images' energy features only. Artificial neural network is always used in image classfication, so radial basis function (RBF) network is used to classify the objects in SAR images, with small training samples and OWATF vectors as the inputs. The result of classification can show the same object obviously. It is concluded that RBF network and OWATF feature based SAR image classification is a good way to SAR image classification with a few training samples.