Generalized incremental image mosaicking with a coarse-to-fine framework via graph cuts

Abstract

Image mosaicking aims to expand spatial coverage by integrating multiple Digital Orthophoto Maps (DOMs) into a unified whole, playing a crucial role in large-scale surface state observation. Optimal seamline detection is a critical process that minimizes intensity differences along effective boundaries, thereby ensuring seamless mosaicking. Recent research has primarily focused on multi-frame joint methods that generate an initial seamline network, followed by the refinement of individual seamlines. However, the simultaneous preparation of all images is not always guaranteed due to the inherent spatial and temporal attributes of the imagery. In contrast, existing frame-to-frame methods perform incremental mosaicking by solely considering simple overlapping relationships within image pairs, without adequately addressing the complexities posed by multi-source images that differ in resolution, size, or topology relative to historical results. Meanwhile, efficiency remains a significant concern, particularly for large-scale and latency-sensitive applications. To address these challenges in a unified manner, we propose an incremental image mosaicking framework capable of processing generalized inputs while effectively bridging the connections between historical and newly acquired imagery. Furthermore, our approach incorporates a graph-cut-based seamline detection method in a coarse-to-fine manner, providing high scalability and adaptability to varying runtime demands through controllable processing granularity. Extensive experiments demonstrate that the seamlines detected by our method exhibit higher quality compared to state-of-the-art commercial software. Moreover, the processing time for aerial images can reach speeds as fast as 2–3 s per task, meeting the requirements for real-time onboard processing. The software is available at https://github.com/pq-chen/GIIM.