The rapid growth of digital images through the widespread popularization of computers and the Internet makes the development of an efficient image
retrieval technique imperative.
Content-based image retrieval, known as
CBIR,
extracts several
features that describe the content of the image, mapping the visual content of the
images into a new space called the feature space. The feature space values for a
given image are stored in a descriptor that can be used for retrieving similar images.
The key to a successful retrieval system is to
choose the right features that
represent the images as accurately and uniquely as possible. The features chosen have
to be
discriminative and
sufficient in describing the objects present
in the image.
To achieve these goals, CBIR systems use three basic types of features:
color features,
texture features and
shape features. It is very difficult to achieve
satisfactory
retrieval results using only one of these feature types.
To date, many proposed retrieval techniques
adopt methods in which more than one feature type is involved.
For instance, color, texture and shape features are used in both IBM's QBIC and
MIT's Photobook. QBIC uses color histograms, a moment-based shape feature,
and a texture descriptor. Photobook uses appearance features, texture features, and
2D shape features. Other CBIR systems include SIMBA, CIRES, SIMPLIcity,
IRMA, FIRE and MIRROR. A cumulative body of research presents
extraction methods for these feature types.
In most retrieval systems that combine two or more feature types, such as color
and texture, independent vectors are used to describe each kind of information.
It is possible to achieve very good retrieval scores by increasing the size of the
descriptors of images that have a high dimensional vector, but this technique has
several drawbacks. If the descriptor has hundreds or even thousands of bins, it may
be of no practical use because the retrieval procedure is significantly delayed. Also,
increasing the size of the descriptor increases the storage requirements which may
have a significant penalty for databases that contain millions of images. Many presented methods limit the length of the descriptor to a small number of bins,
leaving the possible factor values in decimal, non-quantized, form.
The Moving Picture Experts Group (MPEG) defines a standard for content-based
access to multimedia data in their MPEG-7 standard.. This standard identifies
a set of image descriptors that maintain a balance between the size of the feature
and the quality of the retrieval results.
In this web-site a new set of feature descriptors is presented in a retrieval system. These descriptors have been
designed with particular attention to their size and storage requirements, keeping
them as small as possible without compromising their discriminating ability. These descriptors incorporate color and texture information into one histogram
while keeping their sizes between 23 and 74 bytes per image.
High retrieval scores in content-based image retrieval systems can be attained by adopting relevance feedback mechanisms. These mechanisms require the user
to grade the quality of the query results by marking the retrieved images as being either relevant or not.
Then, the search engine uses this grading information in subsequent queries to better satisfy users' needs.
It is noted that while relevance feedback mechanisms were first introduced in the information retrieval field,
they currently receive considerable attention in the CBIR field. The vast majority of relevance feedback
techniques proposed in the literature are based on modifying the values of the search parameters so that they
better represent the concept the user has in mind. Search parameters are computed as a function of the relevance values assigned by the user to all the images retrieved so far.
For instance, relevance feedback is frequently formulated in terms of the modification of the query vector and/or in terms of adaptive similarity metrics.
Also, in this web-site an Auto Relevance Feedback (ARF) technique is introduced which is
based on the proposed descriptors. The goal of the proposed Automatic Relevance Feedback (ARF) algorithm is to optimally readjust the initial retrieval results based on user preferences. During this procedure the user selects from the first round of retrieved images one as being relevant to his/her initial retrieval expectations. Information from these selected images is used to alter the initial query image descriptor.