SPDiff: a spectral-prior-guided diffusion model for infrared ship image super-resolution

Abstract

Infrared ship image super-resolution remains challenging because infrared imagery often exhibits sparse texture cues, weak small-target boundaries, and background noise amplification, particularly under large upscaling factors. To address these issues, we propose SPDiff, a spectral-prior-guided residual diffusion model that explicitly incorporates the spectral statistics of infrared images into both conditioning and uncertainty regulation. Specifically, a Spectral Statistical Prior Map (SS Prior Map) is constructed by combining training-set-level radial amplitude statistics with the phase information of the upsampled LR image, thereby providing global frequency-domain and structure-aware guidance. Further a Spectral Discrepancy-based Uncertainty Prior Map (SDU Prior Map) is designed to estimate spatially varying reconstruction difficulty from the discrepancy between the upsampled LR image spectrum and the high-resolution spectral template. The SDU prior adaptively modulates noise injection and reverse sampling variance, enabling the diffusion process to treat structural regions and smooth backgrounds differently. Experiments on the IRay and MassMIND datasets demonstrate that SPDiff achieves competitive or superior performance across multiple upscaling factors, with clearer ship contours, masts, and slender structures in high-magnification super-resolution. Cross-dataset results further indicate favorable generalization to unseen maritime infrared scenes.