CDP-Net: A Cross-Domain Deterministic-Guided Probabilistic Image Denoising Network

Deep learning has driven significant progress in satellite image denoising; however, most methods rely on deterministic mappings, failing to capture the diversity and uncertainty of degradations. To address this, we propose a Cross-Domain Deterministic Guided Probabilistic Network (CDP-Net) integrating structural priors with probabilistic modeling. CDP-Net employs a Swin Transformer branch to provide global priors that guide a CVAE-based probabilistic branch in generating diverse denoised outputs. Within this branch, a Frequency–Spatial Fusion Module (FSFM) enhances texture representation and suppresses artifacts, while a Cross-domain Bidirectional Attention Mechanism (CM-BAM) enables complementary interactions between branches. Experiments on the NWPU VHR-10 dataset show that CDP-Net achieves superior performance, reaching a PSNR of 38.82 and an SSIM of 0.9774 under salt-and-pepper noise, outperforming existing methods and demonstrating its practical value for satellite image analysis.