Correlation vs SAD, SSD ??
David
I'd like to ask which method performs better for template matching?
Correlation or SSD (sum of squared differences)?
For better I mean in computational speed and in matching accuracy.
Are there any papers/articles discussing about this issue?
ps. Better are theoretical comparisons than empirical results.
Thanks a lot!
Merry Christmas,
David Liu
jundong liu
this method is efficient in term of speed. But it is limited to the
matching
of image pair which are identical except for noise.
Correlation:
Cross-correlation should be used for image types where the
relationship
between values in two imags is predominantly linear at registration.
Notice, you have to use normalized cross-correlation.
BTW, correlation coefficient is better than cross-correlation.
SSD:
I think, is only a more intuitive measure of cross-correlation.
David
I'd also like to ask if there are some papers (maybe some old ones) which
are related to this topic? Since I'm looking for *References*.
Merry Christmas!!
David Liu
andrew queisser
Lisa Gottesfeld Brown, "A Survey of Image Registration Techniques",
ACM Computing Surveys, Vol. 24, No. 4, December 1992
Andrew
"David" <dawenliu@hotmail._com_> wrote in message
news:91kfc8$qev$1...@gemini.ntu.edu.tw...
Leon Majewski
>I'd also like to ask if there are some papers (maybe some old ones) which
>are related to this topic? Since I'm looking for *References*.
I recently did something with the Maximum Cross Correlation Method for the detection of
advective surface currents in satellite imagery.
The best papers i've seen (that were helpful to me anyway) were by Emery and also Kelly
(see below)
The earliest mentioning of such a technique are in Leese et al. 1971 and also McGillem and
Svedlow, 1977,1978 (both referenced in Emery, also Gao and Lythe) - used to track clouds
and also ice.
---
straight out of my references:
DOMINGUES C, R., et al., 2000, Advective Surface Velocities Derived from Sequential
Infrared Images in the Southwestern Atlantic Ocean, Remote Sens. Environ., v73, n2,
p218-226
EMERY, W. J., et al., 1986, An objective method for computing advective surface velocities
from sequential infrared satellite images, J. Geophys. Res., v91, n11, p12,865 - 12,878
GAO, J., and LYTHE, M. B., 1996, The maximum cross-correlation approach to detecting
translational motions from sequential remote-sensing images, Computers & Geosciences, v22,
n5, p525 - 534
HUANG, W. G., et al., 1992, An automatic procedure for sea surface current determination
with sequential thermal infrared Imagery, in Space Oceanography, ed., Cracknell, A. P.,
World Scientific, Singapore
KAMACHI, M., 1989, Advective surface velocities derived from sequential images for
rotational flow field, J. Geophys. Res., v94, n12, p18,227 - 18,233
KELLY, K. A., 1989, An inverse model for near surface velocity from infrared images, J.
Physical Oceanography, v19, b12, p1,845 - 1,864
---
hope that helps
leon
-------------------------
Leon Majewski
Remote Sensing & Satellite Research Group
Curtin University of Technology, Perth, Australia
email: maje...@ses.curtin.edu.au
Jim Sehnert, Ph.D.
First, you must clear up what you mean by correlation. For robust use of
correlation, the image 'section' you are correlating with should be
normalized to a mean of zero and a norm, or length, of 1. If you don't do
the proper normalization, correlation can be maximized at locations that do
not match the template well - simply because that image section may be
unusually bright (or dark). You can find a concise and complete discussion
in Gonzalez and Woods "digital image processing" in Chapter 9.
Also, there are variants on SSD as well. Because the template and image can
often times be off be a constant shift factor, it may be better to measure
the variance of the data set T - IS (template - image section), as opposed
to true SSD. If IS and T differ by approximately a constant value, then the
likelihood that are they are similar objects is high and the variance should
be small.
There is also another article that deals with image registration, which is
similar to template matching, and you can find it on-line. Heres the author
and title (I'm unsure of the periodical, I have a pre-print).
Chia-Yen Chen "Image stitching - Comparisons and new techniques" (1998).
Good luck
Jim
"David" <dawenliu@hotmail._com_> wrote in message
news:91htmf$qva$1...@gemini.ntu.edu.tw...
k
javi...@my-deja.com
Maybe this is obvious for you, but I found people who didn't know that.
If you perform the correlation through FFT, it is computationally
cheaper than do all the sums, I think.
Javier
Sent via Deja.com
http://www.deja.com/