Digital Signal Processing Reference
In-Depth Information
Effective image description method makes base for image data management. During
the recent few years people have been doing abundance of image description related
researches. These researches are different from the traditional method which merely
takes one aspect as describing target, they pay more and more attention to the
combination of semantic and content, they help to obtain high-level semantic indirectly
from the low-level visual features and annotate the image automatically.
Automatic annotation is a challengeable but important task for the content-based
image retrieval work. Using the annotated image collection, it will figure out the
relationship model between the semantic concept space and the visual feature space
automatically, and then use this model to annotate the unspecified target images, more
specifically to say, it's trying to build a bridge between high-level semantic and low-
level visual features, therefore, to some extend, this method will help to handle the
mostly semantic gap problems based on the using of different image retrieval methods.
If this automatic annotation method is feasible, then the current image retrieval
problems will be translated into text information retrieval problems, which has been
maturely mastered by people, and it will be widely used in fields like biomedical
sciences, business, military, education, digital library and internet retrieval, and so on.
The outline of the paper is as follows. We discuss related work in section 2.
Section 3 describes the automatic image annotation algorithm in detail.
Extensive experimental results are shown in Section 4. Finally, conclusions are given
in Section 5.
2
Related Work
With the development of the Internet, bringing new challenges to annotation image
technology. There have been a lot of algorithms presented for automatic image
annotation, which seeks to find keywords that best describe the visual content of an
image. The research on automatic image annotation has proceeded along two
categories, probabilistic modeling method and classification method. The basic idea
of the former is to determine the joint probabilities between annotations and image
visual content. The representative works are Cross-Media Relevance Model (CMRM)
[1], Continuous Relevance Model (CRM) [2], and Multiple Bernoulli Relevance
Model (MBRM) [3]. The method of the second category attempts to infer the
correlations between semantic words with images by learning classifiers, which
include content-based annotation methods with Support Vector Machine [4],
estimating the visual feature distributions associated with each keyword [5], and
asymmetrical support vector machine-based MIL algorithm [6].
In the last few years, segmentation has been used as a preprocessing tool for image
annotation. Images can be segmented in regions, cluster similar regions and then use
these regions as the region semantics. For instance, general purpose segmentation
algorithms like Blobworld [7] and Normalized-cuts[8] to extract regions. These
algorithms do not always produce good segmentations, but are useful for building and
testing models. Duygulu et al. [9] segmented images into blobs, extracted features from
them, and trained the correspondence between features and keywords with the EM
