Frequency-Aware Visible-to-SAR Cross-Modal Image Generation

Abstract

Synthetic aperture radar (SAR) imaging is unaffected by illumination and cloud cover, but its characteristic speckle texture makes visible-to-SAR cross-modal generation difficult. Existing methods mostly constrain the result at the pixel level, so the spectral distribution and speckle statistics of generated images can deviate from real SAR, limiting their usefulness in registration and recognition. Within the PearlGAN framework, this paper proposes a frequency-aware framework with Gaussian-prior decomposition, wavelet supervision, local/global radial spectral constraints, and speckle/ENL modeling. The evaluation includes controlled retraining, scene/source-domain analyses, and downstream SAR detection and classification. The results reveal a trade-off: resub_11 is the recommended balanced configuration with the lowest spectral errors, resub_23 best matches local CV/ENL statistics, and resub_40 gives the lowest GammaKS but remains below the baseline in PSNR, SSIM, and parts of the downstream evaluation. This clarifies the framework’s advantages and applicability boundaries.