For the first class of aggregate images, BCV performs
well, but it is not possible to use it for the second class of
aggregate images. In fact, it should be stressed that no
global technique can handle, for example, images of the
kind shown in Fig. 5a, because some part of the background
is darker than some objects. Visual inspection
may make a human believe that this is not the case, but
this is an illusion probably caused by the fact that our
eyes make local comparisons, cf. for example, Tseng and
Huang (1993), and, locally, objects are darker than background.
In order to threshold both classes of images correctly,
we developed a local thresholding algorithm based
on particles' size, shape and interior range of gray levels.
The algorithm repeats BCV thresholding of an image
until some kind of stop criterion is reached in the various
areas. The algorithm is described as follows.
The developed algorithm assumes1 that the gray
levels of local background are significantly higher than
those of the particles (note that an object may consist
of several particles touching each other to form a cluster); and that the range of the gray values in a particle
is not too large. In practice the new algorithm processing
sequence is (1) the BCV algorithm is applied to
the whole image for the initial thresholding round.
Then, (2) for area A, a shape factor S and the range of
gray levels Df for each object is calculated. (3) For one
object, if the area is too large, or the shape is `strange'
and the range of gray levels in the object is large
enough, perform BCV thresholding in the object
region. (4) Repeat the above step until no further
object can be thresholded, according to these rules.
Before formally presenting the algorithm, we will discuss
data characteristics.