抄録
We have proposed an automated image registration method ART for multi-temporal processing between remotely sensed multi-spectral images. ART had a binary triangle division process, in which triangle was recurcively divided into two sub-triangles so that discrepancy in it was removed by a linear transformation. ART achived highly efficient processing in registration of remotely sensed images. ART, however, had redundancy in triangle division. When a new control point P1 was selected on a common side between triangles T1 and T2, the corresponding point was searched twice for division of T1 and T2, respectively. As the searching was done independently using an index of spatial correlation in the triangles, selected corresponding points were not always identical. Due to the redandancy, one to one correspondence between the reference image and the registered one was not satisfied. Gap or overlap between neighboring triangles decreased the accuracy of the registration.
In this paper, we propose an improved method TART (Improved ART) to realize both high accuracy and high efficiency. For reducing the redundancy in the triangle division, TART uses a forced division in the process. The forced division removes the gap or overlap and yields more triangles without the redundant searching process. The validity of TART is confirmed by a numerical simulation and by processing actual remotely sensed images. Robustness of IART for miss-selection of initial GCP pairs is also shown.
