Efficient Seamline Network Generation for Large-Scale Orthoimage Mosaicking

Seamline network generation from multiple aerial images with overlapping regions is a key issue for creating seamless and large-scale digital orthophoto maps (DOMs). In this letter, an efficient algorithm is proposed that can find adequate networks from hundreds of aerial images in several minutes. It can also be ensured that the seamline passes through weakly textured regions as much as possible. The proposed algorithm consists of two steps. First, the regions overlapped by the images are triangulated into a mesh. Second, a labeling
$l$
is computed that assigns an image
$l_{i}$
as a texture for each triangle using a Markov random field energy function. Triangulating the overlapping area of the images can greatly improve the efficiency of the seamline network generation, and by designing a smoothing term with a divergence difference between images, the seamline network passes through the weakly textured area as much as possible. In addition, we also use a hierarchical optimization strategy to improve efficiency in large-scale scenes. The proposed method can be easily combined with other data information (e.g., a digital surface model), or other auxiliary methods (e.g., building segmentation), to generate high-quality seamline networks. Experimental results on numerous challenging datasets show that the proposed algorithm can achieve high-quality seamline networks efficiently. Moreover, it outperforms several existing methods and software.