
530 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 17, NO. 4, APRIL 2008
then the following inequality holds:
(12)
Proof: See Appendix I.
B. Property I
Let
and . Hence,
then the following inequalities hold:
(13)
Proof: The left inequality can be easily derived from the
application of the Cauchy–Schwarz inequality to each subvector
pair defined by subsets
. The right inequality
can be obtained directly from successive applications of the pre-
vious lemma.
Property I states that an upper-bound of , the cross correla-
tion between vectors
, can be obtained by applying -times
the Cauchy–Schwarz inequality to the subvector pairs defined
by subsets
, and that this bound is tighter than the upper bound
attainable by applying the inequality to the original vectors
.
C. Property II
Let
and . Hence,
then the following inequalities hold:
(14)
Proof: Similar to Property I, the left inequality derives
from the application of the Cauchy–Schwarz inequality to each
subvector pair defined by subsets
.
The right inequality is easily proved by applying the
Fig. 1. Generic partitioning of template and current image subwindow.
Cauchy–Schwarz inequality to the subvector pair defined
by subset
.
Property II tells us that given an upper bound of
obtained
by partitioning
into subvectors as defined in Property I, a
tighter upper bound can be obtained by replacing the product-of-
norms term related to the subvector pair defined by
with the
corresponding cross correlation term. Successive applications
of Property II yield increasingly tighter upper-bounding func-
tions of
, each step of the succession requiring the computa-
tion of a new cross correlation term associated with a subvector
pair so as to replace the corresponding product-of-norms term.
V. C
ORE EBC ALGORITHM
This section describes the core EBC algorithm, which relies
on the mathematical properties presented in Section IV. First
of all, both the template
and the current image subwindow
are seen as vectors belonging to a -dimensional
space and Fig. 1 shows a generic partitioning of such vectors
into
subvectors, as required by Properties I and II. In general
a subvector can consist of disjoint sets of pixels, as is the case
of subvector 4 in Fig. 1.
To deploy the properties of Section IV within the bounded
correlation framework outlined in Section III, we define a parti-
tioning of vectors
and apply Property I at first, to ob-
tain an initial upper bound
at each image position
and check the associated skipping condition (7) that does not re-
quire calculation of any partial correlation term. If such a initial
condition is not verified, we then apply Property II in successive
steps. At each step, the product-of-norms term of a subvector
pair is replaced by the corresponding cross-correlation term, to
obtain a tighter bounding function and associated skipping con-
dition (7).
For reasons of computational efficiency, in a practical deploy-
ment of the EBC principle it is preferable to adopt a kind of
“regular” partitioning scheme to be applied to
and .
In our implementation
and are partitioned into sub-
vectors made out of successive rows, as shown in Fig. 2. All
subvectors are chosen to have the same number of rows,
,ex-
cept for the last one (e.g., in Fig. 2, subvector
). Hence, with
our partitioning scheme the first
subvectors have
elements and the last one has elements.