We use B-spline snakes [3][7] to represent image tag lines. The spline is given by the following expression:
where is a column vector of powers of
, the spline
parameter,
is a sequence of control points, and
is a matrix which blends the control points.
The approach is to minimize the following expression
along tags:
In equation (6), the first term maximizes
along the length of the snake, and second
and third terms attract the snake to the endpoint
of tag lines.
The discrete form of , for a quadric
spline may be written as:
where are the B-spline control points. Dynamic
Programming (DP) [1] may be
used to optimize the curve in the control
point space to minimize
using the following
recurrence
for , and
.
In general, for an order
B-spline,
is a function of
control points. Also,
note that the minimization yields the optimal
open spline, as is the case for a tag line.
In order to find a closed snake,
one performs
applications of the above recurrence,
where
is the number of possible choices for
the endpoint
, and for each optimization fixing
the end point to be one of the
choices,
repeating for all
possibilities,
and finally choosing the minimum.
Figure 2 shows the initial star-burst
image and results from the DP
algorithm. Figure 3 shows the energy field
for a diagonal tag line, and for one endpoint of
the same tag line.