Exploiting Thermal Self-Similarity for Robust Blind Super-Resolution of Real-World Infrared Images

Abstract

Infrared image super-resolution (IISR) is challenging because real infrared (IR) sensors introduce composite degradations. Optical blur couples with Gaussian noise, fixed-pattern stripe noise, and low-frequency non-uniformity, and their composition and intensity vary with the operating environment and sensor operating time. Existing IISR methods are trained mostly under a single bicubic degradation or handle only a subset of noise types, so they degrade sharply under these diverse real-world conditions. In this paper, we propose a blind super-resolution framework that restores infrared images robustly across the full range of real-world degradations. We first build a physics-based composite degradation model that reproduces the coupling between optical blur and sensor-specific noise, providing a realistic training distribution for blind IISR. We then design a non-local aggregation network (NLAN) that suppresses sensor noise by averaging redundant observations of the same scene content. Thermal scenes are texture-sparse and self-similar, so a given region recurs across the image. NLAN clusters these recurring regions and lets each group exchange information, which cancels the independent sensor noise while preserving the shared structure. This realizes a learnable non-local means that is robust to the composite, out-of-distribution noise of real IR sensors. Finally, a pretrained generative diffusion prior takes the clean restoration and adds back fine textures. Extensive experiments on synthetic benchmarks and real infrared images show that the proposed framework restores images robustly across degradation conditions ranging from weak to severe and produces clean, highly legible restorations of high perceptual quality.