Complex-valued deep-learning model for 3D SAR modeling and image generation

Synthetic aperture radar (SAR) collects samples of the 3D Spatial Fourier transform on a two-dimensional manifold corresponding to the backscatter data of wideband pulses launched from different look angles along an aperture. Regularized inversion methods typically yield a sparse set of points and artifacts due to limited measurements in the elevation domain. We present a coordinate-based multi-layer perceptron (MLP) that enforces the smooth surface-prior and models the complex-valued scattering coefficients. The 3D geometry is represented using the signed distance function, while the scattering coefficients are represented using real and imaginary channels. Since estimating a smooth surface from a sparse and noisy point cloud is an ill-posed problem, in this work, we regularize the surface estimation by sampling points from the implicit surface representation during the training step. The loss function also incorporates the consistency between the predicted complex-valued scattering coefficients and the true scattering coefficients. We validate the model's reconstruction ability using the Civilian vehicle data domes.