基于分块CNN的多尺度SAR图像目标分类算法

针对合成孔径雷达（Synthetic Aperture Radar, SAR）图像目标分辨率差异大，多尺度SAR图像目标分类准确率不高的问题，提出了一种基于迁移学习和分块卷积神经网络（Convolutional Neural Network, CNN）的SAR图像目标分类算法。首先通过大量与目标域相近的源域数据对分块CNN的参数进行训练，得到不同尺度下的CNN特征提取网络；其次将CNN的卷积和池化层迁移到新的网络结构中，实现目标特征的提取；最后用超限学习机（Extreme Learning Machine, ELM）网络对提取的特征进行分类。实验数据采用美国MSTAR数据库以及多尺度SAR图像舰船目标数据集，实验结果表明，该方法对多尺度SAR图像的分类效果优于传统CNN。

Multi-Scale SAR Images Target Classification Algorithm Based on Block Convolutional Neural Network

Because of big difference of SAR image resolutions and low accuracy rate of multi-scale SAR image classification, an algorithm based on transfer learning and block convolutional neural network (CNN) is proposed. Firstly, the parameters of block CNN are trained by large number of source domain data which is similar to the target domain, and the CNN feature extraction networks under different scales are obtained. Secondly, the convolution and the pool layer are transferred to the new network structure to achieve the feature extraction. Finally, the extracted features are classified by extreme learning machine(ELM). The MSTAR database and multi-scale SAR ship image data sets are used as experimental data. The experimental results show the proposed method is superior to traditional CNN in the classification of multi-scale SAR images.