1. EVEN MORAVEC using 32x32 pixel partitions, 3 
candidate points per partition, a 5x5 interest operator 
window and an interest value minimum threshold of 500. 
• Original image CCPs after Moravec filter 153 
• Skewed image CCPs after Moravec filter 160 
• Original image CCPs after eliminating adjacent pts.. 53 
• Skewed image CCPs after eliminating adjacent pts.. 56 
2. TIN HORN to match the candidate point lists, given a 
distance threshold of 20 pixels and an AIV maximum 
threshold of 500. 
• CCP pairs after correspondence 44 
3. HARDCORR using a 5x5 correlation window into an 
11x11 search window, and a correlation coefficient 
minimum threshold of 0.75. 
• CCP pairs after correlation 31 
4. COORD2 (an ERDAS program) took the final list and 
calculated the transformation parameters via a least-squares 
fit. The RMS tolerance threshold was set at 1.0 pixel. 
• CCP pairs after least-squares fit 25 
Using the GCP list as is, it took 7 iterations to bring the 
RMS below tolerance. Six points were discarded by 
COORD2 in the process. These points were shown to have 
errors in X and/or Y of at least 10 pixels. Without the errant 
control points the RMS values were: 
X = 0.867 pixels 
Y = 0.296 pixels 
The skewed image was resampled using the transformation 
parameters. The results were visually tested by displaying 
the original image in blue and the restored image in red. The 
composite image yielded a sharp pink image, confirming that 
both images were in register. By comparison, the original 
image overlayed on the skewed one had the distinct 
appearance of a "double" image. So the package appears to 
work correcdy given a skewed image to match. 
There have been other tests of the package - mainly 
duplicating the above test with different degrees of relative 
transformation and/or with different sizes and types of 
images. There were similar CCP attrition rates during these 
tests. 
DIRECTION OF FUTURE WORK 
The testing has not gone beyond restoration of a skewed 
image because of major problems encountered with the 
correspondence algorithm. As mentioned earlier, TIN 
HORN cannot match CCPs if there is too large a difference 
in the orientations of the images to match. The degree of 
difference at which the algorithm fails was found to depend 
mainly on the size of the images. The larger they are, the less 
likely they will be able to be matched. This is because 
rotational differences result in larger displacements the 
further a CCP is from the center of rotation. Therefore, small 
images were able to be matched accurately in tests, while 
larger images, using the same transformation matrix to create 
the skewed subject image, were not. 
"La Grange," as it currently stands, is not able to co-register 
full Landsat scene-sized images unless they are already 
nearly identical in geometry. Even relatively small images 
(512x512) cannot be matched if there is more than a few 
degrees rotation and/or a few pixels of translation between 
them. User removal of most of the translation and rotation 
difference between the images during preprocessing is not 
really a solution since it would defeat the purpose of an 
automated image matching system. 
We have decided that the best way to overcome the 
correspondence problem is to develop a method to 
automatically calculate approximate transformation 
parameters in the preprocessing phase. The calculated 
transformation matrix would be passed to TIN HORN to 
identify corresponding areas in the images. 
To do this, we are going to attempt to match clusters of 
Moravec operator-extracted points. As mentioned earlier, if 
the Moravec operator is not modified by image partitioning 
and elimination of adjacent points, it will tend to extract 
points in clusters about areas of high contrast if given a low 
enough threshold. Preliminary investigations have shown 
that similarly shaped clusters are generated in separate 
images. Using strategies as outlined in Ventura et al. [1990] 
and Tom and Jain [1989], it may be possible to match 
Moravec clusters to determine approximate transformation 
parameters. 
CONCLUSION 
We have shown that the concept of selecting, matching, and 
fine tuning automatically-selected control points can co- 
register images with small degrees of transformation. 
However, the system needs to be significantly improved 
before a general automated image matching system can be 
realized. With an improved correspondence algorithm, 
matching may be possible regardless of image rotation and 
translation. It is hoped that such a process can be 
demonstrated to be effective for most image registration 
applications. 
REFERENCES 
Ehlers, M., 1985. "The Effects of Image noise on Digital 
Correlation Probability", Photogrammetric Engineering and 
Remote Sensing, vol. 51, no. 3, March 1985, pp. 357-365. 
Ehlers, M. and R. Welch, 1987. "Stereocorrelation of 
Landsat TM Images", Photogrammetric Engineering and 
Remote Sensing, vol. 53, no. 9, Sept. 1987, pp. 1231- 
1237. 
Forstner, W., 1986. "Digital Image Matching Techniques 
for Standard Photogrammetric Applications", Technical 
Papers, ACSM-ASPRS Annual Convention, Washington 
D.C., vol. 4, pp. 210-219. 
Luhmann, T. and M. Ehlers, 1984. "AIMS: A System for 
Automatic Image Matching", pp. 971-979. 
Luhmann, T. and G. Altrogge, 1986. "Interest-operator for 
Stereo Matching", ISPRS Commission III Symposium, 
Rovaniemi, Finland, ISPRS XXVI, vol. 3/2, pp. 459-474. 
448