采用深度学习的多方位角SAR图像目标识别研究

Research on Multi-Aspect SAR Images Target Recognition Using Deep Learning

合成孔径雷达(Synthetic Aperture Radar,SAR)在对地面目标进行观测时,可以在多个不同的方位角获取到目标的SAR图像,但这些图像中目标的形态各不相同。考虑到SAR图像对观测方位角极其敏感和SAR图像数据规模小这两个因素,本文设计了一个利用多方位角SAR图像进行目标识别的卷积神经网络(Convolutional Neural Network,CNN),同一目标的3幅SAR图像被当作一幅伪彩色图像输入到网络中,充分利用了SAR图像数据的获取特点,同时用池化层替代了展平操作,降低了网络参数数量。实验结果表明,即便在小规模SAR数据集上,该卷积网络具有识别精度高的特点,对同类别不同型号的目标也具有出色的识别表现。

Synthetic Aperture Radar（SAR） can obtain SAR images of the target from a number of different azimuths when observing a ground target,but the shapes of the target in these images are different. In view of the fact that SAR image is extremely sensitive to the observation azimuth and small scale of SAR image dataset,this paper designs a Convolutional Neural Network（CNN） for multi-aspect SAR images target recognition. Three SAR images of the same target are regarded as a pseudo-color image inputted to the network,which making full use of the acquisition characteristics of SAR image data.Instead of flattening,we use pool layer to reduce the number of parameters of network at the same time. The experimental results show that this convolutional network architecture has high recognition precision on small scale of SAR dataset,and has excellent recognition performance for different types of targets in the same category.