基于小波域两向二维主分量分析的 SAR 目标识别

摘要

针对合成孔径雷达图像目标识别在图像域进行特征提取时空间维数较高、计算复杂度较大、识别效率较低等问题，提出基于小波域两向二维主分量分析和概率神经网络的 SAR 图像目标特征提取与识别方法。该方法首先引入二维离散小波变换将预处理后的 SAR 图像变换到小波域，得到可充分表征目标特征信息的低频成分。然后提取低频子图像的两向二维主分量分析低维特征作为训练样本和测试样本的目标特征，最后由概率神经网络分类器完成目标识别。MSTAR 数据实验结果表明，在特征矩阵维数低至6×3（原始图像128×128）的情况下平均识别率高达99．32％，且最高可达99．83％，该方法不但能够有效压缩目标特征维数和提高识别率，还对目标的方位信息具有很强的鲁棒性，可有效应用于 SAR 图像目标特征提取和识别。%Considering various problems in SAR image target recognition,such as high feature dimensions, large computational complexity and low recognition efficiency,a novel SAR image target recognition method based on two-directional two-dimensional principal component analysis((2D)2 PCA)in wavelet domain and probabilistic neural network(PNN)is proposed.First,the pre-processed SAR image is transformed into wavelet domain by two-dimensional discrete wavelet transformation(2DDWT).And the low-frequency sub-image of the beat decomposition series is obtained which fully represents the target information.Then,it takes the (2D)2 PCA feature of the low-frequency sub-image as testing and training sample feature.Finally,target recognition is achieved by PNN classifier.Experimental results based on MSTAR SAR data show that,in the case of feature dimension as low as 6×3(original image 128×128),the average recognition rate with the proposed method is up to 99.32% with the highest value of 99.83%.The method not only can compress feature dimension and improve recognition rate effectively,but also has strong robustness to the azimuth information of target.And the proposed algorithm shows a better performance of SAR image target recognition.