Active Re ranking for Web Image Search
Active
Re ranking for Web Image Search
Abstract
Image
search re ranking methods usually fail to capture the user’s intention when the
query term is ambiguous. Therefore, re ranking with user interactions, or
active re ranking, is highly demanded to effectively improve the search
performance. The essential problem in active re ranking is how to target the
user’s intention. To complete this goal, this paper presents a structural
information based sample selection strategy to reduce the user’s labeling
efforts. Furthermore, to localize the user’s intention in the visual feature
space, a novel local-global discriminative dimension reduction algorithm is
proposed. In this algorithm, a sub manifold is learned by transferring the local
geometry and the discriminative information from the labeled images to the
whole (global) image database. Experiments on both synthetic datasets and a
real Web image search dataset demonstrate the effectiveness of the proposed
active re ranking scheme, including both the structural information based
active sample selection strategy and the local-global discriminative dimension
reduction algorithm.
Existing
System
Although text-based
search techniques have shown their effectiveness in the document search, they
are problematic when applied to the image search. There are two main problems.
One is the mismatching between images and their associated textual information,
resulting into irrelevant images appearing in the search results. For example,
an image which is irrelevant to “panda” will be mistaken as a relevant image if
there is a word “panda” existing in its surrounding text. The other problem is
that the textual information is insufficient to represent the semantic content
of the images. The same query words may refer to images that are semantically
different, e.g., we cannot differentiate an animal panda image from an image
for a person whose name is Panda, just with the text word “panda”. Because the
textual information is insufficient for semantic image retrieval, a natural
recourse is the visual information. Recently a dozen of image/video re ranking
methods have been proposed to exploit
the usage of the visual information for refining the text-based search result.
Most of these re ranking methods utilize the visual information in an
unsupervised and passive manner. Unsupervised reranking
methods, can only achieve limited performance improvements. This is because the
visual information is insufficient to infer the user’s intention, especially
when the query term is ambiguous.
Disadvantages:
- user intention is not considering in the already existing system.
- reranking methods usually fail to capture the user’s intention when the query term is ambiguous.
- text-based search techniques are problematic when applied to the image search bcz of the mismatching between images and their associated textual information
- Textual information is insufficient to represent the semantic content of the images
Proposed
System
In this paper, reranking with user’s
interactions is named as active
reranking. In active re
ranking, the essential problem is how to capture the user’s intention, i.e., to
distinguish query relevant images from irrelevant ones. Different from the conventional
learning problems, in which each sample only has one fixed label, an image may
be relevant for one user but irrelevant for another. In other words, the
semantic space is user-driven, according to their different intentions but with
identical query keywords. Therefore, we propose to target the user-driven
intention from two aspects: collecting labeling information from users to
obtain the specified semantic space, and localizing the visual characteristics
of the user’s intention in this specific semantic space, respectively. Although
Intent Search can be deemed as a simplified version of active re ranking, i.e.,
the user’s intention is defined by only one query image, it can not work well
when the user’s intention is too complex to be represented by one image. the
query relevant images for “Animal” vary largely both in visual appearance and
features, thus we cannot represent “Animal” only with one image. Instead, our
proposed active re ranking method can learn the user’s intention more
extensively and completely.
Advantages:
Ø reranking with user interactions, or active
reranking is introduced in this system.
Ø collecting
labeling information from users to obtain
the specified semantic space.
Ø
localizing the visual characteristics of
the user’s intention in this specific semantic space
Ø
A new structural information (SInfo)
based strategy is proposed to actively select the most informative query
images.
Ø
To localize the visual characteristics
of the user’s intention, we propose a novel local-global discriminative (LGD)
dimension reduction algorithm.
Architecture:
Architecture Diagram
Software
Requirements Specification:
Software
Requirements:
Front End
: Struts
Framework
Back
End : Oracle 10g
IDE : my
eclipse 8.0
Language : java (jdk1.6.0)
Operating
System : windows XP
Hardware
Requirements:
System : Pentium IV 2.4 GHz.
Hard Disk
: 80 GB.
Floppy Drive : 1.44 Mb.
Monitor
: 14’ Colour Monitor.
Mouse
: Optical Mouse.
Ram : 512 Mb.
Keyboard
: 101 Keyboards.
Modules:
- Label information collection (Textual information search)
- Visual characteristic localization
- Re ranking implementation
- Active sample selection
Active
User’s Labeling Information Collection:
To
collect the labeling information from users efficiently, a new structural information (SInfo) based
strategy is proposed to actively select the most informative query images. it
is essential to get the necessary information by labeling as few images as
possible. In active reranking, however, only a few images will be labelled by a
user. To avoid or alleviate the influence of the small sample size problem, we
proposed SInfo sample selection strategy.
Visual
characteristic localization:
To localize the
visual characteristics of the user’s intention, we propose a novel local-global discriminative (LGD)
dimension reduction algorithm. Basically, we assume that the query relevant
images, which represent the user’s intention, are lying on a low-dimensional
sub manifold of the original ambient (visual feature) space. LGD learns this
sub manifold by transferring both the local geometry and the discriminative
information from labelled images to unlabelled ones. The learned sub manifold
preserves both the local geometry of labelled relevant images and the
discriminative information to separate relevant from irrelevant images. As a
consequence, we can eliminate the well-known semantic gap between low-level
visual features and high-level semantics to further enhance the re ranking
performance on this sub manifold.
Re ranking
implementation process:
To
verify the effectiveness of the proposed active reranking method, we apply the
SInfo active sample selection strategy and the LGD dimension reduction
algorithm to reranking. In this paper, we take the Bayesian reranking as the basic reranking algorithm for
illustration. When applying the Bayesian reranking for active reranking,
modifications will be made to incorporate the new obtained information. The
final reranking result is obtained by sorting the images according to in a
descending order.
more representative samples are preferred for
labeling. In SInfo , the ambiguity of an image is
Algorithm:
Local global discriminative (LGD) dimension reduction algorithm.
Categories:
where is the architectural diagram men .........
ReplyDelete