Radiation-Tolerant Unsupervised Deep Image Stitching for Remote Sensing

Remote sensing image stitching is a critical technique for generating comprehensive and high-resolution geospatial datasets. However, radiation distortions caused by sensor noise, atmospheric interference, or environmental variability often degrade feature consistency, leading to visual misalignment and artifacts, i.e., seams in stitched results. To address this challenge, we propose a novel radiation-tolerant unsupervised deep image stitching (RT-UDRSIS) framework, which transforms the traditional stitching pipeline into two processes: radiation-tolerant warping (RTW) and seamless inpainting (SI). In the initial stage, our method leverages the Mamba architecture’s sequential modeling capabilities to robustly capture long-range geometric dependencies across global-to-local scales, even under radiation-induced discrepancies between image pairs. In the subsequent process, we introduce an effective unsupervised low-rank adaptation (LoRA) training approach that fine-tunes a stable diffusion model, employed for inpainting and iteratively refining features while suppressing radiation-induced noise. By integrating these two stages, our method achieves precise and seamless image stitching without the necessity for paired ground-truth datasets. Acknowledging the challenges in assessing the quality of stitched images, we introduce comprehensive metrics grounded in multimodal large language models (MLLMs). Extensive experiments conducted on benchmark datasets illustrate that our method surpasses state-of-the-art techniques in alignment accuracy and robustness, especially under conditions of severe radiation distortions.