REGISTERING INTERFEROMETRIC SAR IMAGES USING MARR- 
HILDRETH ZERO-CROSSINGS 
S. Mukherji 
Centre of Studies in Resources Engineering, Indian Institute of Technology, Bombay, Mumbai — 400 076, INDIA. 
shyamali@csre.iitb.ac.in 
KEYWORDS: Registration, SAR, Edge, Image, Digital. 
ABSTRACT: 
Leberl (Leberl, 1990) proposed an edge-based algorithm for finding matching pairs of points in SAR images. In this 
paper, we show that the algorithm can be used to register interferometric SAR images which requires sub-pixel 
precision. This precision is achieved by interpolating the discrete values of the metric used for matching, in this 
algorithm, using cubic B-splines. The interferogram obtained is comparable to that obtained using cross-correlation. 
1. INTRODUCTION 
In the algorithm proposed by Leberl, the Marr- 
Hildreth operator is used to smooth the images and 
then find the laplacian. The resulting images have 
positive and negative values. These images are 
thresholded at zero and the resulting binary images 
used to find the matching points. The metric used for 
matching is the number of matching pixels in 
corresponding windows of the images. 
2. A FAST ALGORITHM 
The computation of the metric can be made efficient 
by finding the number of matching pixels at one offset 
between the windows and then adding the change in 
this number as one of the windows is translated across 
the other. The scan-lines and columns of the windows 
are run-length encoded. Now, the change in the 
number of matching pixels can be attributed solely to 
the right-hand pixels of the runs of one of the 
windows. Furthermore, this change increases or 
decreases only when these right-hand pixels cross over 
the left-hand pixels of the runs of the other window; 
this speeds up the computation. 
3. RESULTS AND CONCLUSIONS 
These concepts have been implemented in a 
preliminary fashion by Leberl. We have implemented 
the algorithm fully. We have also applied the Marr- 
Hildreth operator to the windows of the master image, 
that are to be matched, instead of the entire image. We 
have registered an ERS-1/2 tandem pair of images 
(Fig. 1) using this algorithm to find the corresponding 
pairs of points in the images. A low value (sigma — 1) 
for the standard deviation of the Gaussian function in 
the Marr-Hildreth operator works well. Since 
matching with sub-pixel precision is necessary for the 
816 
registration of interferometric images, we have 
interpolated the number of matching pixels (for 
various offsets between the windows) using cubic B- 
splines to achieve sub-pixel precision. The 
interferogram that we have thereby obtained (Fig. 2) is 
comparable to that obtained using cross-correlation 
(Fig. 3). 
REFERENCES 
Leberl, F.W., 1990. Radargrammetric | Image 
Processing, Artech House, Massachusetts, pp. 250- 
259. 
Interne