Moving and stationary targets separation in SAR signal domain using parallel convolutional autoencoders with RPCA loss

Moving targets cannot be simultaneously focused with stationary background scenes in synthetic aperture radar (SAR) images. They appear defocused and azimuthally displaced like blurred areas in SAR images. The well-known robust principal component analysis (RPCA) theory is formulated and implemented in the loss function of our proposed framework, which is based on two parallel convolutional autoencoders. These two autoencoders are trained in a self-supervised manner by a simulated dataset. Each instance of the dataset contains the range-compressed signal of a moving target, with unknown motion and coordinates parameters, immersed in ground clutter. The trained model can extract the low-rank (ground clutter) and sparse (moving target) components in the range compressed domain. Even for stationary scene imaging, this separation of moving and non-moving targets is vital since there might always be some unwanted moving targets appearing in the illuminated area, in particular, in urban environments where cars are expected. Computer simulations have been performed to evaluate the performance of the proposed method and validate the theoretical discussions.