URsNet: Unsupervised Remote Sensing Stitching Network for Low-Altitude UAV Images

Abstract

Traditional feature-based image stitching methods depend heavily on the quality of feature matching, which leads to suboptimal performance when applied to drone remote sensing images with significant differences in viewpoint and depth of field. Concurrently, supervised learning paradigms have proven infeasible due to the paucity of labeled data. To address these limitations, this letter proposes an unsupervised stitching framework tailored for uncrewed aerial vehicle (UAV) images. The proposed framework consists of two stages: unsupervised image alignment and unsupervised image fusion. In the first stage, we propose a multiscale feature-based progressive transformation regression network, which performs high-precision alignment in complex scenarios via a global–local adaptive transformation. For the second stage, an enhanced unsupervised soft-mask constraint strategy is proposed to generate an artifact-free stitched image. To establish a training and evaluation benchmark, a drone-specific dataset for unsupervised deep image stitching is constructed by integrating multisource datasets. We also utilize multimodal large language models (MLLMs) to establish a stitching quality evaluation metric. Experiments demonstrate that the proposed method significantly outperforms state-of-the-art approaches in both quantitative metrics and visual quality; notably, it remains competitive even when compared to supervised methods.