17-03-2014, 10:49 AM
Abstract—Adaptive Local Image Registration based on
adaptive filtering can register both grayscale images
and color images. Here the local distortions are
corrected without explicitly estimating the displacement
field. A filter of appropriate size convolves with
reference image and gives the pixel values
corresponding to the distorted image and the filter is
updated in each stage of the convolution. When the
filter converges to the system model, it provides the
registered image. In this method the 2-D image plane is
mapped into a 1-D sequence using space-filling curves.
Adaptive Local Image Registration can be implemented
in two ways as Pixel-By-Pixel Method and Block-Based
Method. This work did a comparative study between
these two methods. The two methods are tested using
different types of geometrically distorted images and
the quantitative performance evaluation is done using
Mean Squared Error (MSE), Mean Absolute Error
(MAE), Peak Signal-to-Noise Ratio (PSNR) and
Structural Similarity (SSIM) index. The Block-Based
method shows better performance than Pixel-By-Pixel
Method.
I. INTRODUCTION
Image registration is the process of determining
the coordinate transformation or mapping relating
different views of the same or similar objects. This
can also be stated as the process of overlaying two or
more images of same scene taken at different time,
from different viewpoints and or by different sensors.
The major registration purpose is to remove or
suppress geometric distortions between the reference
and sensed images, which were introduced due to
different imaging conditions, and thus to bring
images into geometric alignment. The general
applications of image registration include Medical
Imaging, Scene Analysis, Object Recognition,
Remote Sensing, Automated Monitoring, Industrial
Inspection, Robot Vision etc.
The broad classification of image registration is
on the basis of domain of transformation [3]. On the
basis of domain of transformation registrations can be
classified in to Global image registration and Local
Image Registration. Global Image Registration
methods employ parametric spatial transformation
while Local image registration methods can handle
spatially varying deformations.
Certain assumptions on the images being
registered are necessary in global transformations.
Invalid assumptions affect the accuracy of
registration process and so global methods may show
local distortions. Hence local registration methods are
needed. The important Local Image Registration
methods are Adaptive Local Image Registration [1]-
[2], Dense motion estimation methods [4]-[6], 2-D
mesh based approaches [7] and Methods employing a
3-D scene representation [8].
‘Adaptive Local Image Registration’ is
based on adaptive filtering [9], which can register
both grayscale images and color images. This method
registers the images without the explicit estimation of
the local displacement. Adaptive filters are able to
estimate unknown systems and so it can track smooth
changes in the system. For implementing this method
and to get spatial contiguity in images, mapping of
the 2-D image into a 1-D update sequence using
space-filling curves [10] such as Hilbert curve is
required. Here an adaptive filter convolves with
reference image and gives the predicted pixel value
for the current pixel value of the distorted image.
This is known as prediction step. After this the error
between the predicted output and the distorted image
is calculated and using this error and previous filter
parameter, the filter is updated. When the filter
converges to the system model, it provides the
registered image.
adaptive filtering can register both grayscale images
and color images. Here the local distortions are
corrected without explicitly estimating the displacement
field. A filter of appropriate size convolves with
reference image and gives the pixel values
corresponding to the distorted image and the filter is
updated in each stage of the convolution. When the
filter converges to the system model, it provides the
registered image. In this method the 2-D image plane is
mapped into a 1-D sequence using space-filling curves.
Adaptive Local Image Registration can be implemented
in two ways as Pixel-By-Pixel Method and Block-Based
Method. This work did a comparative study between
these two methods. The two methods are tested using
different types of geometrically distorted images and
the quantitative performance evaluation is done using
Mean Squared Error (MSE), Mean Absolute Error
(MAE), Peak Signal-to-Noise Ratio (PSNR) and
Structural Similarity (SSIM) index. The Block-Based
method shows better performance than Pixel-By-Pixel
Method.
I. INTRODUCTION
Image registration is the process of determining
the coordinate transformation or mapping relating
different views of the same or similar objects. This
can also be stated as the process of overlaying two or
more images of same scene taken at different time,
from different viewpoints and or by different sensors.
The major registration purpose is to remove or
suppress geometric distortions between the reference
and sensed images, which were introduced due to
different imaging conditions, and thus to bring
images into geometric alignment. The general
applications of image registration include Medical
Imaging, Scene Analysis, Object Recognition,
Remote Sensing, Automated Monitoring, Industrial
Inspection, Robot Vision etc.
The broad classification of image registration is
on the basis of domain of transformation [3]. On the
basis of domain of transformation registrations can be
classified in to Global image registration and Local
Image Registration. Global Image Registration
methods employ parametric spatial transformation
while Local image registration methods can handle
spatially varying deformations.
Certain assumptions on the images being
registered are necessary in global transformations.
Invalid assumptions affect the accuracy of
registration process and so global methods may show
local distortions. Hence local registration methods are
needed. The important Local Image Registration
methods are Adaptive Local Image Registration [1]-
[2], Dense motion estimation methods [4]-[6], 2-D
mesh based approaches [7] and Methods employing a
3-D scene representation [8].
‘Adaptive Local Image Registration’ is
based on adaptive filtering [9], which can register
both grayscale images and color images. This method
registers the images without the explicit estimation of
the local displacement. Adaptive filters are able to
estimate unknown systems and so it can track smooth
changes in the system. For implementing this method
and to get spatial contiguity in images, mapping of
the 2-D image into a 1-D update sequence using
space-filling curves [10] such as Hilbert curve is
required. Here an adaptive filter convolves with
reference image and gives the predicted pixel value
for the current pixel value of the distorted image.
This is known as prediction step. After this the error
between the predicted output and the distorted image
is calculated and using this error and previous filter
parameter, the filter is updated. When the filter
converges to the system model, it provides the
registered image.