Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 66

Then, the winning node is obtained:

j ∗ = arg min

j

d ( v i , w j ) (3.11)

The input vector v i is assigned with the label, y i according to the following
assignment,

1

w j ∈

S k ( v q )

y i = (3.12)

0 Otherwise

The application of Eqs. ( 3.10 )-( 3.12 ) to all samples v 1 , v 2 ,..., v N , results in the

N
i = 1

set of labels { y i } , y i ∈{ 0 , 1 } . This constitutes the label set for the corresponding

N
i

samples x 1 , x 2 ,..., x N , which forms the training set { x i , y i }

1 for the relevance =

feedback modules for the next retrieval step.

3.2.4 Experimental Result

The experiment was conducted to compare the performance of SOTM with SOM
for obtaining pseudo-RF in the adaptive retrieval process. This was carried out
using a subset of the Corel image database [ 73 ] consisting of nearly 12,000 images,
covering a wide range of real-life photos, from 120 different categories, each
containing 100 images. Three set of query images were constructed for testing and
evaluation: Set A, Set B, and Set C. In each set, one random sample was selected
from each class; thus, one set of tests quires included an example for every class
(120 in total). For feature space

F

1 , color histograms, color moments, wavelet
moments, and Fourier descriptors were used, while Hu's event moment [ 74 ] and
Gabor descriptors accompanied with color histograms and color moments were used
for feature space

F

2 .
The system architecture of pseudo-RF discussed in Fig. 3.3 was implemented.
In the SOM and SOTM algorithms, the maximum number of allowed clusters was
settoeight.A4

×

2 grid topology was used in the SOM structure to locate the
eight possible cluster centers (fixed topology). Table 3.1 shows the retrieval result
obtained by the pseudo-RF as compared to the initial retrieval result (at first search

Fig. 3.2 (continued) Comparison of SOFM vs. SOTM. SOFM ( left ) was run to completion with
4 different grid lattices ( top to bottom :2 × 2, 3 × 3, 4 × 4, 5 × 5); SOTM ( right ) is the result of a
single run, shown at equivalent stages of node generation ( top to bottom : 4 nodes + trigger site for
next node, 9 nodes, 16 nodes, and 25 nodes). SOTM shows efficient allocation of nodes to regions
in which data exists. The broken circles indicate the hierarchical control function (threshold about
each node beyond which data spawns new nodes). The circles in the final plot indicate cluster
densities