Enhanced Multiscale Pyramid Deep Image Prior for Unsupervised Remote Sensing Image Restoration

Abstract

The deep image prior (DIP) restores images without training data by exploiting the implicit bias of convolutional architectures, but suffers from slow convergence, limited texture fidelity, and inconsistent quality on satellite imagery. Enhanced multiscale pyramid DIP (EMSP-DIP) extends the DIP framework with: 1) coarse-to-fine hierarchical training with cross-scale consistency; 2) structured multifrequency input noise; and 3) a leak-safe compound loss combining gradient, edge, and boundary-transition terms, all evaluated on cloud-free pixels only. On RICE II images that span cloud coverage of 0.1%–87.2%, EMSP-DIP improves over baseline DIP by +5.46 dB PSNR and +0.043 SSIM (paired Wilcoxon <inline-formula> <tex-math notation="LaTeX">$p\lt 10^{-3}$ </tex-math></inline-formula>), in every coverage tier. When an exact cloud mask is available, classical biharmonic interpolation remains superior on pixel-level metrics; EMSP-DIP narrows this gap and overtakes it in the most heavily occluded scenes (36%–87% coverage), where boundary information becomes scarce. A complete <inline-formula> <tex-math notation="LaTeX">$2^{3}$ </tex-math></inline-formula>-component ablation is reported, together with a seed-paired comparison over 45 pairs and three seeds, which establishes that compound loss improves all four metrics (<inline-formula> <tex-math notation="LaTeX">$p\lt 0.05$ </tex-math></inline-formula> in each seed), a frequency-domain analysis of input noise, and a study of hyperparameter sensitivity.