depends upon the number of image grey levels (typically
256) and whose element Aj,i gives the number of pixels
of the original image which have grey level i and local
mean ,j. A clustering technique applied to the points of
A can be used to detect the presence of a ship in the
image. If this presence is confirmed with a high
confidence, the value of the binarization threshold can
be computed by means of the following algorithm.
Starting from the matrix A, a predetermined percentage
p% of the points of the original image that are
closest to the main diagonal is selected. The histogram
Hl of these selected points is evaluated. In our
experiment, the value of p has been chosen equal to
5%, according to Kirby and Rosenfeld’. This procedure
spots the pixels which belong to ‘homogeneous’ regions,
and consequently, the histogram Hl exhibits deeper
valleys than the original histogram. It should be noted
that this technique is particularly effective in the case of
radar images, because ships are more homogeneous
than the background.
To smooth the histogram profile and make the valley
bottom easier to select, histogram Hl is filtered by a
median filter with a suitable mask, and histogram H2 is
generated. In this paper, the mask length of the median
filter is assumed to be equal to 13, but the choice of this
parameter is not critical. The threshold value T is
chosen equal to the local minimum of H2 with the
highest grey level. The value of this minimum must
exceed a given value (fixed at 120 in our experiments) to
once more decrease the probability of a false alarm.
Figure 3 shows histograms of the images of Figure 1
after transformation and filtering. For each example,
the value of the threshold T is indicated on the greylevel
scale in Figure 4.