Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 79

Table 3.5 Retrieval results for DB2, averaged over 30 query images
Avg. number of user RF
Method Avg. relative precision (%) for convergence (Iter.)
Non-adaptive method 49.82 -
Pseudo-RF 79.17 -
User-controlled RF 95.66 2.63
Semi-automatic RF 98.08 1.33
Column 2: average relative precision (%); column 3: number of user feedbacks
(iterations) required for convergence. The results were performed by ARBFN in
both automatic and semi-automatic modes

For relevance classification on the feature space F 2 ,theGWT[ 91 , 92 ] was again

adopted to characterize the retrieved images. The GWT was applied to the dominant
colors in each channel and the transform coefficients were used to construct feature
vectors. This method gives better characterization of texture information from
different color spaces.
Table 3.5 provides average relative precision (ARP) figures for thirty query
images. The ARP was defined as Pr , where Pr is the number of relevant

images over 12 retrieved images, and the maximum was taken across all methods
discussed. This provides an easy way to compare the relative performance among
different retrieval methods. In general, conclusions similar to those for the texture
database (DB1) can be drawn from these results, with regard to the retrieval per-
formance. The semiautomatic method consistently displayed superior performance
over the other methods discussed, showing improvement from 49.8 to 98.1 %, and
with the number of user feedbacks reduced by half to reach convergence.
The retrieval session for this database is shown in Fig. 3.9 a, b. Figure 3.9 ashows
the 12 best-matched images without learning, with the query image displayed in
the top-left corner. It is observed that some retrieved images are similar to the
query image in terms of texture features. Seven similar images are relevant. Based
on this initial information, the self-organizing system dynamically readjusts the
weight parameters of the ARBFN model to capture the notion of image similarity.
Figure 3.9 b displays the retrieval results, which are considerably improved after
using the automatic interactive approach. Figure 3.9 c shows the retrieval results
of the semiautomatic ARBFN in comparison to the user-controlled interactions
illustrated in Fig. 3.9 d.

/ max ( Pr )

3.4 Region-Based Re-ranking Method

The pseudo-RF learning represents the blind relevance feedback, where the machine
performs pseudo labeling. This process requires modeling image contents with
sufficiently accurate features for the characterization of perceptual importance.
This issue is especially pressing with automatic RF since, without providing
some form of knowledge to the relevance classification process from the external