Information Technology Reference
In-Depth Information
Efforts have been taken to find good weights for particular visual features.
As a matter of fact, the problem of aggregating the features has been trans-
formed into the problem of finding suitable weights. As suggested in the
experiments presented in Subsection 11.2.5.2, visual features perform dif-
ferently on different queries. A promising solution would be to derive the
variable weights from the query for each query.
From one query image, it is difficult to determine the query concept,
which is what users look for by presenting the example. The similarity-based
retrieval using one example simply finds the close match of the means of all
features of images. The weights are either equal (no weighting) or preselected.
Query by one example cannot realistically lead to scalable, satisfactory query
performance [3]. Inspired by the weighting technique in relevance feedback
techniques [10], the weights are normally set to be inversely proportional to
the variances of the feature distances of multiple example images. The prin-
ciple of this strategy is to assign higher weights to features that are close to
each other in examples and considered characterizing the more significant
features of the query.
Figure 11.7 shows a comparison of unweighted and weighted linear
aggregation of features that are proposed in Ren [9]. Figure 11.7a and
Figure 11.7b are the query examples and the groundtruth image set, respec-
tively, which are the relevant images available in the image collection.
Figures 11.7c and 11.7d are the results with equal weights (unweighted)
and the proposed weights, respectively, where the groundtruth images are
highlighted with dark bars. The weighted system is able to retrieve all seven
groundtruth images in the top 10 output images while the unweighted sys-
tem can only find five groundtruth images in the top 20 output images as
shown in the figures.
The difficulty of the linear combination of features is not only the diffi-
culty to determine the weights but also the loss of the visual significance in
summing distances of individual visual feature.
11.3.3 Classifier Combination
The common combinatorial operations are summarized as follows. Let
C
1
, C
2
,
…
, C
n
denote the individual features,
s
Ci
(
x
) be the similarity
defined for the
i
ith feature of the image
x
and
s
(
x
) is the aggregated overall
s i m i l a r it y.
•
MAX (Maximum)
The maximum feature similarity is chosen as the overall image
similarity. This operation ensures that two images are considered
similar if all features are similar.
s
MAX
(
x
) = max {
s
C1
(
x
), … ,
s
Cn
(
x
)}.
(11.24)
