12-12-2012, 05:10 PM
Semisupervised Biased Maximum Margin Analysis for Interactive Image Retrieval
Interactive Image Retrieval.pdf (Size: 2.39 MB / Downloads: 42)
Abstract
With many potential practical applications, content-
based image retrieval (CBIR) has attracted substantial attention
during the past few years. A variety of relevance feedback
(RF) schemes have been developed as a powerful tool to bridge
the semantic gap between low-level visual features and high-level
semantic concepts, and thus to improve the performance of CBIR
systems. Among various RF approaches, support-vector-machine
(SVM)-based RF is one of the most popular techniques in CBIR.
Despite the success, directly using SVM as an RF scheme has
two main drawbacks. First, it treats the positive and negative
feedbacks equally, which is not appropriate since the two groups
of training feedbacks have distinct properties. Second, most of the
SVM-based RF techniques do not take into account the unlabeled
samples, although they are very helpful in constructing a good
classifier. To explore solutions to overcome these two drawbacks,
in this paper, we propose a biased maximum margin analysis
(BMMA) and a semisupervised BMMA (SemiBMMA) for integrating
the distinct properties of feedbacks and utilizing the
information of unlabeled samples for SVM-based RF schemes.
The BMMA differentiates positive feedbacks from negative ones
based on local analysis, whereas the SemiBMMA can effectively
integrate information of unlabeled samples by introducing a
Laplacian regularizer to the BMMA.
INTRODUCTION
DURING the past few years, content-based image retrieval
(CBIR) has gained much attention for its potential applications
in multimedia management [1], [2]. It is motivated by the explosive growth of image records and the online accessibility
of remotely stored images. An effective search scheme
is urgently required to manage the huge image database. Different
from the traditional search engine, in CBIR, an image
query is described by using one or more example images, and
low-level visual features (e.g., color [3]–[5], texture [5]–[7],
shape [8]–[10], etc.) are automatically extracted to represent
the images in the database. However, the low-level features
captured from the images may not accurately characterize the
high-level semantic concepts [1], [2].
BMMA AND SEMIBMMA FOR SVM RF IN CBIR
With the observation that “all positive examples are alike;
each negative example is negative in its own way,” the two
groups of feedbacks have distinct properties for CBIR [20].
However, the traditional SVM RF treats the positive and negative
feedbacks equally.
To alleviate the performance degradation when using SVM
as an RF scheme for CBIR, we explore solutions based on the
argument that different semantic concepts lie in different subspaces
and each image can lie in many different concept subspaces
[20]. We formally formulate this problem into a general
subspace learning problem and propose a BMMA for the SVM
RF scheme. In the reduced subspace, the negative feedbacks,
which differ in diverse conceptswith the query sample, are separated
by a maximum margin from the positive feedbacks, which
share a similar concept with the query sample. Therefore, we
can easily map the positive and negative feedbacks onto a semantic
subspace in accordance with human perception of the
image contents.
CBIR SYSTEM
In experiments, we use a subset of the Corel Photo Gallery
as the test data to evaluate the performance of the proposed
scheme. The original Corel Photo Gallery includes plenty of
semantic categories, each of which contains 100 or more images.
However, some of the categories are not suitable for image
retrieval since some images with different concepts are in the
same category and many images with the same concept are in
different categories. Therefore, the existing categories in the
original Corel Photo Gallery are ignored and reorganized into
80 conceptual classes based on the ground truth, such as lion.
CONCLUSION AND FUTURE WORK
SVM-based RF has been widely used to bridge the semantic
gap and enhance the performance of CBIR systems. However,
directly using SVM as an RF scheme has two main drawbacks.
First, it treats the positive and negative feedbacks equally, although
this assumption is not appropriate since all positive feedbacks
share a common concept, while each negative feedback
differs in diverse concepts. Second, it does not take into account
the unlabeled samples, although they are very helpful
in constructing a good classifier. In this paper, we have explored
solutions based on the argument that different semantic
concepts live in different subspaces and each image can live
in many different subspaces.