Interferometric image denoising network SEVReNet

Interferometric imaging is often accompanied by complex noise contamination, primarily manifested as Gaussian noise introduced by the thermal motion of detector electrons and speckle noise produced by the interference of coherent light and multipath scattering. However, most existing deep-learning-based denoising networks are typically designed to model and learn a single noise distribution. Consequently, when faced with real scenarios in which multiple noise sources are superimposed, these models often have limited generalization ability and struggle to suppress different noise components simultaneously. To address this problem, we propose SEVReNet, a self-supervised network for denoising interferometric images. On top of ordinary convolution, the proposed network introduces a scale-equivariant module and a rotation-equivariant module, forming a three-branch architecture that simultaneously leverages the advantages of translation, rotation, and scale equivariance. This design exploits the differing preferences of the three modules for different types of information and performs adaptive fusion through weighted integration, thereby achieving better separation of structural information and noise while preserving fine texture details, bringing into play the maximum advantages of the three modules for the denoising task. We conducted extensive experiments on both synthetic and real interferometric datasets contaminated by Gaussian and speckle noise. The results show that, compared with BM3D, U-Net, Restormer, and AdaReNet, SEVReNet achieves superior denoising performance, with an average PSNR of 31.48 dB and an SSIM of 0.924, significantly outperforming competing methods. These results verify the robustness and effectiveness of SEVReNet under complex noise conditions and provide new insights into noise modeling and image restoration in optical imaging.