This paper presents the difference between object recognition for satellite image and general object recognition. To
overcome the difficulties of object recognition for satellite image, two methods are proposed in this paper. First, it is
shown that traditional moment invariants are no longer invariant while object shape is distorted or contour is blurry. This
paper proposes an improved invariable moment algorithm, which reconstructs seven invariants moment formulas, and
weakens the sensitivity on distorted border. Moreover, a 3D model database including all interested target's 3D model
can be established. When starting to recognize the target in satellite image, several 2D projection images can be retrieved
from the model database, using these images as recognition templates will be more accurate than using a single target
image. Experimental results show that the methods presented in this paper are effective on satellite image recognition.
|