foot on the ground.
DIGITAL IMAGE CORRELATION
Testing for similarity of digital images is accomplished by
use of numerical image correlation formulas. In these
formulas, equal sized density arrays are compared pixel for
pixel and a coefficient is computed indicating the
similarity of the images. Many different formulas are
available for this purpose. The ones tested included
euclidian distance, normalized euclidian distance,
normalized cross correlation, and the hadamarad trans
formation. The normalized formulas account for lens fall-
off and photo processing inconsistencies by subtracting
the mean density of the array from the individual elements.
This allows conjugate imagery having systematic density
differences to be correlated as long as the relative con
trast is the same. After testing it was determined that
the normalized cross correlation coefficient yielded the
best results and was therefore used in the development of
the automated system. Equation 1 shows the form of the
normalized cross correlation coefficient. In this equation
A and B are the i,j th elements of the two arrays being
compared and A and B are the mean densities of the arrays.
n m
(1)
C
The value of this coefficient ranges from -1 to +1 with +1
indicating a perfect match. For our purpose it was re
scaled to a range of 0 to 200 with lower values indicating
better correlation.
In order to find the location in the photo of the desired
image array, referred to as the target array, the target
must be compared at all possible locations in a much
larger search array. This is accomplished by lagging the
target array through the search array in a systematic
fashion with a coefficient being computed for each lag
position. The minimum coefficient thus obtained indicates
a best match to the target. This coefficient is tested
against a threshold value to see if the match is valid.
This is necessary since the target may not fall within the
search array and the minimum coefficient does not always
indicate an acceptable match. Testing the correlation
algorithm showed that values below approximately 40
indicated significant correlation with values less than
20 showing very strong correlation.
Due to the large number of computations necessary to
compute all of the correlations possible within the search
array, a systematic method of lagging the target array was
devised which allows corrections to be added to the coe
fficient at the previous lag position. The target is
shifted one row or column at a time with correction terms