GLFR: Progressive Global–Local Feature Representation Network to Approximate Taylor Expansion for Remote Sensing Image Denoising

Prior-based denoising methods estimate the latent clean image by constraining the solution space through regularization terms, which have powerful interpretability but overly rely on prior knowledge, are sensitive to regularization parameters, and have a heavily computational burden. To address these issues, this letter analyzes the potential mathematical relationship between image denoising and Taylor expansion approximation, and reports a progressive global-local feature representation (GLFR) network to approximate Taylor expansion for remote sensing image denoising. Specifically, we use a lightweight convolution to estimate the base layer and a U-shaped network to approximate the derivative layers. In the U-shaped network, a dual-domain feature fusion (DDFF) block is introduced as a backbone, including a global-local feature extraction (GLFE) block and a 3D feed-forward network (3DFFN). Finally, the outputs of the base layer and the derivative layers jointly contribute to the Taylor expansion approximation solution (the latent clean image). In experiments, we first conduct ablation studies to verify the effectiveness of each block. Then, both objective and subjective results are compared and demonstrate that the proposed GLFR achieves competitive denoising performance, and are even better than the existing state-of-the-art (SOTA) denoising methods.