IEEE Geoscience and Remote Sensing Magazine - December 2019 - 14

composite sequence and the requirement for intermediate
treatment. In contrast, multiframe joint methods are more
efficient, as shown in Figure 6. These methods can simultaneously compute a seamline network for multiple images. In other words, the seamlines for all of the images are
detected at the same time. The multiframe joint methods
were developed on the basis of frame-to-frame methods in
recent years, so they have on advantages in quantity.
Several methods have been developed for generating
the seamline network based on a Voronoi diagram, as
shown in Figure 7. Hsu et al. [94] first proposed a local-toglobal method using ordinary Voronoi diagrams to generate a network of seamlines. However, it cannot be ensured
that the seamline always lies in the overlapping areas.
More effectively, based on area Voronoi diagrams with
overlap (AVDO), a seamline network was formed automatically and effectively by Pan et al. [95]. This network is
globally generated and can be further refined by the radiometric difference in the overlapping areas. Pan et al. [96]
further improved the AVDO algorithm by including the
detection of valid regions, providing a more general algorithm for generating bisectors, and refining the seamline
network by combining the bottleneck model and Dijkstra's algorithm. Owing to the watershed segmentation algorithm, the bounded Voronoi diagrams [97] for a global
seamline network generation are improved. The Voronoi
diagram provides an excellent way for multiframe joint
seamline detection. To obtain an outstanding seamline
network that bypasses the integrated objects, preliminary
results of the Voronoi-diagram-based methods should be
further optimized.
The framework of graph cuts has also been introduced
into the multiframe joint optimization strategy for seamline detection [98]. This method allows a simple human-
computer interaction to constrain the image regions the
seamlines will or will not pass through. Based on the optimal network vertices, the seamlines with the shortest
paths between vertices can be detected by a graph-based
approach [99]. Pan et al. [100] proposed an initial seamline
network generation method based on improved seeded region growing, where the boundary of the overlapping areas is selected as the seed of the algorithm. This method is
raster based and can address concave polygonal overlapping regions.

Compared with frame-to-frame methods, multiframe
joint methods significantly increase the efficiency of
seamline detection by omitting intermediate treatments.
At the same time, the result is improved by the replacement of serial detection with global optimization. Based
on only the information in the images themselves, multiframe joint methods can achieve a great effect. However,
their efficiency can be further improved by introducing
external data.
EXTERNAL-DATA-BASED METHODS
Traditional methods of seamline detection are based on
only the images themselves. In recent years, methods
based on external data have attracted increasing attention. These external data can guide the seamline to bypass ground objects (e.g., buildings), which brings a new
concept to seamline detection. The main works are listed
as follows.
First, lidar point clouds were used for seamline detection in [101]. Wang et al. [102] adopted a vector road map to
generate seamlines; here, the vector road data are applied to
construct a weighted graph, and Dijkstra's algorithm finds
the lowest-cost path as the seamline. Similarly, Wan et al.
[103] used vector road data to search for the seamline, but
they instead applied the Floyd-Warshall algorithm [104]
to find the lowest-cost path. A region-based saliency map
[105] generated by a human attention model [106] has also
been used to guide the seamline, which is also guided by
pixel-based image similarity and location constraints.
To avoid discontinuity in the mosaic, the information of
ground object classification has been imported for seamline detection [107], where the object classes are obtained
by the normalized difference vegetation index and the
morphological building index.
A digital surface model (DSM) is commonly applied to
detect seamlines. For example, Chen et al. [108] used the
elevation information from a DSM to guide a seamline
toward a low area by stereo matching. Based on the initial seamline network of the Voronoi diagram algorithm,
Zheng et al. [109] and Zheng et al. [110] used the DSM
to detect the edge diagram, which is finally refined by
the weighted A* algorithm. To avoid buildings, on the
one hand, a DSM can be applied using a gradient operator [111]; on the other, object heights can be derived from

V(A1)

V(A3)

V(A2)

FIGURE 6. A multiframe joint method for seamline (denoted by the

red lines) detection.
14

FIGURE 7. Seamlines based on a Voronoi diagram.
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2019

IEEE Geoscience and Remote Sensing Magazine - December 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2019