Database Reference
In-Depth Information
1
,
2
,...,|
I
|}
, where
|
I
|
is the number of query images, I i
C
,
i
=
1
,
2
,...,|
I
|
.Using
Bayes' theorem, the a posteriori probability can be written as
P
(
c
|
x
,
I
)
p
(
x
|
c
,
I
)
P
(
c
|
I
) ,
(2.81)
with the equality replaced by the proportionality due to the unimportance of
the probability density function (PDF) of an observation, i.e. p
, when the
theorem is employed to solve a classification problem. Based on the meaning of the
background information I , we can assume the conditional independence between the
observation x and I given the class label of the observation, i.e. x
(
x
|
I
)
I
|
c . Therefore,
the a posteriori probability in Eq. ( 2.81 ) can be calculated through
P
(
c
|
x
,
I
)
p
(
x
|
c
)
P
(
c
|
I
)
(2.82)
The first term on the right-hand side of Eq. ( 2.82 ) is the PDF of the feature
vector of the class c , which is considered as the content model characterizing the
visual properties of that class. The second term is essentially a distribution of one
class or candidate image, say c , conditional on a set of other classes or query
images, collectively represented by I . This is exactly the contextual information
that characterizes the statistical relation between different classes or images. It will
be shown that such contextual information can be learned from past user feedback
for image retrieval. According to Eq. ( 2.82 ), the content and contextual information
are integrated through the decision-level fusion in a multiplicative fashion.
The Bayesian framework is applied to tackle the semantic gap of image retrieval
by integrating short-term relevance feedback (STRF) and long-term relevance
feedback (LTRF). STRF refers to the user interaction during a retrieval session
consisting of a number of feedback iterations, such as query shifting and feature re-
weighting. On the other hand, LTRF is the estimation of a user history model from
past retrieval results approved by previous users. LTRF plays a key role in refining
the degree of relevance of the candidate images in a database to a query. The STRF
and LTRF play the roles of refining the likelihood and the apriori information,
respectively, and the images are ranked according to the a posteriori probability.
By exploiting past retrieval results, it can be considered as a retrieval system with
memory, which incrementally learns the high level knowledge provided by users.
The underlying rationale of applying the Bayesian framework to image retrieval
can be illustrated using Fig. 2.6 , of which the gist is to boost the retrieval per-
formance using some information extracted from the retrieval history. The two
types of similarity measure are complementary to each other. Specifically, the
similarity measure by the content-based component illustrated by the low-level
feature space in Fig. 2.6 a suffers from the semantic gap which can be alleviated
using the contextual information. The links between relevant images in Fig. 2.6 b
are estimated by utilizing the co-occurrence of relevant images in the past retrieval
results. At the same time, the contextual information can only be acquired by
learning from the knowledge accumulated through the content-based component.
The retrieval system, illustrated in Fig. 2.7 , seamlessly integrates the content-based
Search WWH ::




Custom Search