|
Image registration refers to the mergence of two partially-
overlapped images,subject to rotation, translation and scaling,
into one big image.In this thesis, we first extend differential
operators derived from orthonormal compactely supported
wavelets to partial derivative operators for two-dimensional
signals. The partial derivative operators are then used to
compute the gradient of two images to be registered to each
other. The gradient-based edge detector are used to locate
feature points in both image.From these feature points, the
angle of rotation , the distance of translation and the factor
of scaling between the two original images can be computed,
with very good accuracy. The matching or correspondence
between the two sets of features points in the images to be
registered is computation intensive and prone to errors .In
this thesis,however, we take advantage of the color information
in images to achieve more accurate matching between these
feature points.An estimate of the matchingparameters are carried
out with knowledge of the vector relations between the the
refined matched pairs . It follows that an even more refined
match can besaught based on those newly computed matching
parameters.In fact,this refinementprocess can be repeated for a
few more times,if necessary.Finally,the differencein the color
of the two original images are reduced with histogram
specificationon the red,greed and blue planes,respectively.
Experiment with simulated and real images show that the proposed
method produces accurate registration of image even when the
factor of scale and the rotationbetween the images are large or
the overlap between them is small.Keywords: Wavelet, Image
Registration, Feature Points ,Differential Operator
Color Image.
|