Rob Reeves
pixel and DCT domain algorithms is the method by which the partial derivatives and their products are calculated for the
least squares adjustment A matrix. In the pixel domain, the first difference is used. In the transform domain, the method
effectively reconstructs the symmetrically extended continuous derivative, and samples it at the appropriate points. We
conclude that this difference in the way derivative estimates are calculated is a significant factor in the improved matching
quality of the DCT domain algorithm.
8 SUMMARY AND CONCLUSIONS
A hybrid least squares image matching algorithm was proposed, that performed least squares matching in the transform
domain, but resampled between iterations in the pixel domain. Experiments were conducted to compare this algorithm
with an algorithm operating completely in the pixel domain. There were two major differences between the algorithm,
apart from the domain used for the least squares adjustment. These were a window side two pixels smaller in the pixel
domain, and a difference in the method used to estimate the sampled partial derivatives used in the construction of the least
squares adjustment A matrix. Since the energy compacting property of the DCT concentrates the signal energy into the
low frequency DCT coefficients, it was proposed to match in the DCT domain with only low frequency DCT coefficients,
The size of the A matrix could then be significantly reduced, without discarding any significant signal structure important
for matching, and hence maintaining the quality of the matching process. Since the time taken to match is dependent
on the size of the A matrix, it was hypothesized that this could be used to increase matching speed. While there was
no improvement in matching speed for an 8 x 8 match window, the parameters which measured match quality, (the
number of converging windows, and the mean and standard deviation of the disparity errors), were comparable or showed
improvement in the transform domain when about 50% of the DCT coefficients were used in the matching process.
For 16 x 16 and 32 x 32 windows, there were useful gains in matching speed, while the other measures of matching
quality were comparable or better. The optimum percentage of DCT coefficients for these window sizes were 20% and
10% respectively. The approach also incorporates derivative estimates that result in better accuracy than can be achieved
using the first differences of our pixel domain approach. In addition, our results suggest that neglecting the higher order
DCT coefficients, equivalent to low pass filtering, may increase the number of converging match windows in areas with
repetitive regular texture.
REFERENCES
Ackermann, F., 1984. Digital image correlation: Performance and potential application in photogrammetry. Photogram-
metric Record 11(64), pp. 429—439.
Forstner, W., 1982. On the geometric precision of digital correlation. International Archives of Photogrammetry and
Remote Sensing 24(3), pp. 176-189.
James, G., Burley, D., Dyke, P., Searl, J., Steele, N. and Wright, J., 1993. Advanced Modern Engineering Mathematics.
Addison-Wesley, Wokingham, England.
Martucci, S. A., 1994. Symmetric convolution and the discrete sine and cosine transforms. IEEE Transactions on Signal
Processing 42(5), pp. 1038-1051.
Rabbini, M. and Jones, P., 1991. Digital Image Compression Techniques. SPIE Optical Engineering Press, Bellingham,
WA.
Rao, K. and Yip, P., 1990. Discrete Cosine Transform - Algorithms, Advantages, Applications. Academic Publishers,
San Diego.
Reeves, R., 1999. New shift, scaling and derivative properties for the DCT. In: Visual Communication and Image
Processing '99, San Jose, CA, USA.
Reeves, R. and Kubik, K., 1998. Least squares matching in the transform domain. International Archives of Photogram-
metry and Remote Sensing 32(3/1), pp. 168-176.
Wallace, G. K., 1991. The JPEG still-picture compression standard. Communications of the ACM 34(4), pp. 31-44.
Wang, Z., 1990. Priciples of Photogrammetry (With Remote Sensing). Press of Wuhan Technical University of Surveying
and Mapping, Beijing.
768 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
KE
AB
Ab
of |
con
The
des
rela
give
hon
squ.
inte
The
of t
Coo
feat
leas
The
acci
whi
Mai
obs
tern
The
unb
(‘le
i ke |
poir
Obs:
som
For
line
the
tedi
In tl
non
repr