Self-Organizing Maps and Their Applications in Image Processing, Information Organization, and Retrieval - Nonlinear Signal and Image Processing: Theory, Methods and Applications - page 420

TABLE 11.1
Performance of the LBG, SOM, and MMSOM
in Quantizing Several Frames of Color Image
Sequence “Trevor White” a
Recall MSE

Frame No. LBG SOM MMSOM

10 289.65 313.46 240.32

50 283.27 298.84 222.38

100 295.43 312.01 242.52

120 285.27 308.36 226.46

129 282.22 307.14 224.05

a When the learning procedure was applied on the
first frame in the presence of mixed additive Gaussian
and impulsive noise, Codebook size = 32.

after the completion of the first layer, clusters the documents of the train-
ing corpus into classes that contain relevant documents with respect to their
context. Accordingly, the WEBSOM architecture is a prominent candidate for
document organization and retrieval. The MMSOM has replaced the SOM in
the aforementioned architecture. 41 To assess the quality of the document or-
ganization (i.e., segmentation) that results by the SOM and the MMSOM, the
average recall-precision curve 59 has been derived for a test (unseen) document
query of known categorization. Figure 11.9 demonstrates that the MMSOM
improves the precision rate for a range of recall rates. The frequencies of bi-
grams can be sorted into descending order and signed ranks can be assigned
to them. A novel metric that incorporates these signed ranks has been pro-
posed for evaluating the context similarity between document pairs. A novel
SOM variant, the Wilcoxon SOM ,isbuilt based on this metric. 46 Figure 11.10
depicts the average recall-precision curve for the SOM and the Wilcoxon SOM
for the “ACQ” category of the Reuters-21578 corpus. 60

11.7 A Class of Split-Merge Self-Organizing Maps

One common feature of both SOM and LBG algorithms is that they rely on
the assumption that the number of output neurons or the size of codebook N
is known in advance or is preset to a desired value, usually a power of 2. Al-
though several VQ design techniques, such as the pairwise nearest neighbor
algorithm 39 , 61 or the tree-structured VQ, 62 employ splitting criteria that might
be useful in determining how many output neurons or codevectors ade-
quately approximate the training set provided in the input of VQ, the avail-
ability of splitting criteria is not sufficient on its own. Because a quantization
in terms of N

1 output neurons always is expected to yield a lower MSE
than a quantization in terms of N output neurons, 63 the issue of deciding what

+