ROBUST REPRESENTATION OF SHAPE WITH DAGS - Video Object Extraction and Representation: Theory and Applications - page 124

dependent computation of each edge value. When the recursion returns,
its output is used in the edge value representation of the macro-structure.
After the DAG structure is derived and all the edge values have been
computed, the DAG is compared against a reference DAG with the DAG-
Compare operation. A DAG-Compare can perform any operation that

can be expressed as a path comparison between two DAG o s such as

grammar checks or dictionary searches.

BASIS
As shown in Figure 5.12, the basis of these systems is a special case
of the inductive step where the datastream does not require partitioning
to be recognized. When the datastream cannot be decomposed any fur-
ther, a single functional block handles the recognition of the datastream.
This functional block can be complex and / or adaptive, such as a HMM,
TDNN, Kohonen self-organizing feature map or as simple as the low fre-
quency points of a DFT. The classification results from this functional
block are passed back up. A recursive design philosophy would suggest
a basis to be as simple as possible.

APPLICATION: CURSIVE HANDWRITING
WORD RECOGNITION
We have implemented a simple cursive word recognition system [Lin
and Kung, 1997b] that has a simple design with two levels of recur-
sive structure with a simple Discrete Fourier Transform (DFT) curve
recognizer as lowest level (or basis) system. As shown in Figure 5.13,
the recursive structure of the cursive word recognizer integrates a curve
matching algorithm, letter recognizer and a full dictionary search. In
cursive handwriting, breaks exist at different levels of representation:
loop/line, character, word, and sentence level. Each level of representa-
tion corresponds to a level within the recursive architecture. This struc-
turally recursive architecture and its implications on recognizer design
are discussed in detail in [Lin and Kung, 1997a].