SAR Target Recognition from Subband and Subaperture Images via Complex Valued Convolutional Neural Networks

Synthetic aperture radar (SAR) images are different from optical images, as they reflect the scattering characteristics of targets and are complex data. Traditional real valued neural networks can only handle amplitude information and lose phase information. In addition, SAR images obtained by traditional imaging algorithms can only represent spatial scattering characteristics of targets, while the scattering characteristics of frequency and azimuth angle embed in two-dimensional SAR images, which cannot be utilized when directly using SAR images as inputs for neural networks. In response to the above issues, a method of SAR target recognition from subband and subaperture images via complex valued convolutional neural networks is proposed in this paper. The method uses subband and subaperture images at different frequencies and azimuth angles as multi-channel inputs for the complex valued neural convolutional network to achieve target recognition. By this method, not only is the spatial scattering information of the target utilized, but also the scattering information about frequency and azimuth angle is utilized. Not only is the amplitude information of the image utilized, but also the phase information of the image is utilized. Experiments on the MSTAR dataset have verified the effectiveness of this method.