Information Technology Reference
In-Depth Information
The same study has shown that most descriptors have peaks
and holes in the distribution of descriptor elements that can be
compressed.
11.2.5.2 Perceptive Analysis
A comparative study on physical similarity measured by descriptors and
perceived similarity by human subjects is important to show the expressive-
ness, or effectiveness, of various descriptors for indexing and retrieval appli-
cations. Experiments have been carried out for MPEG-7 color and texture
descriptors [13].
The groundtruth sets of the queries are generated from the human sub-
jects for each query. Then the MEPG-7 descriptors are applied to retrieve
the images and the performance is measured by precision and recall, which
will be described in Section 11.3. Figure 11.3 shows the results for three dif-
ferent queries named Bush, Opera House, and Party. It shows that different
descriptors perform differently for different queries. For the Bush query, the
color structure works best and the EHD works significantly better than color
layout and homogeneous texture descriptors while for the Opera House
query, the color layout and edge histogram descriptors do equally better than
the color structure and homogeneous texture descriptors. However, for the
Party query, all descriptors perform equally poor for almost all of the recall
range. There is no one descriptor that can perform satisfactorily for all que-
ries. These observations have led to the research on methods to aggregate
the individual feature similarities and to design the aggregation strategies
based on individual queries that will be discussed in Section 3.
11.2.5.3 Effectiveness for Indexing
To evaluate the effectiveness of MPEG-7 descriptors for indexing applica-
tions, experiments have also been conducted to describe the evolutional
changes in image time sequences [12]. It shows the applicability of MPEG-7
for measuring indexing images with gradual changes. Figure 11.4 shows the
visual changes of a rotting banana over time and Figure 11.5 shows the nor-
malized distances of these changes measured by MPEG-7 color and texture
descriptors using their respective distance functions presented in Section
11.2 with the first image as the reference. The results demonstrate that all four
descriptors, dominant color, color layout, scalable color, and homogeneous
texture are able to index the changing images correctly with their monotonic
increasing distances. It is also found that the distance functions of the color
layout and SCDs are more consistent to the degree of the perceived visual
changes, noting that the small visual change of the banana from time point
1 to time point 2 are measured by these two descriptors with small distances
rather than by the other two.
