Remote sensing image denoising based on Fourier embedding

Aiming to address the issue of limited model expression ability due to the loss of detailed features and insufficient extraction during remote sensing image denoising, this paper proposes a Fourier-embedded remote sensing image denoising network. By analyzing the magnitude spectrum and phase spectrum of Fourier coefficients, it is found that the noise features of the image mainly exist in the phase spectrum. In order to fully utilize this property, Fourier features are learned and their connections are embedded into the detail information recovery block to efficiently capture noise points in the image and recover detail features; local and global contextual semantic information of the image is extracted and integrated from different perceptual dimensions by the multi-scale residual block; and the selective feature fusion block adopts an elemental weighting approach to integrate the feature representations at different levels and selectively highlights useful feature information. In addition, in order to enhance the visual perception of the denoising results, the loss function used to train the model incorporates a combination of perceptual loss, joint mean absolute error, and mean square error. The experimental results show that the proposed model achieves satisfactory results on both synthetic noise data and real noise data.