Imagine you are a designer working on the next
Lord of the Rings movie. You have seen thousands of images, graphics, and
photos. However, you can only recall a few characteristics of the images
(perhaps it had a blue sky or sand dunes, etc.). How do you find the visual
similarities? Perhaps you are a journalist who needs to make a comparison of
the New Year celebrations from around the world. How do you find the right
video shots? Visual Information Retrieval is focused on finding visual imagery.
As databases are more and more popular, the
claim for storing and querying images from databases has also appeared. Though
special tools for solving these problems cannot be used in all cases. For
example, when a given image database is only an extension of an existing large
database containing text data (e.g., police registration). Storing images in
legacy database is more cost-effective than procuring special new database
engines only for storing images. Obviously, in each case (mainly in the
latest), retrieval of images is based on a given matching strategy, which
contains a given algorithm in most cases.
Storing of images and, mainly, their retrieval
from databases differs from the storing and retrieving of other
non-multimedia-like data. By the spread of the new Object-Relational or fully
Object-Oriented databases, further possible solutions have occurred.
Nevertheless, matching algorithms and strategies used for retrieving images
from the databases at present time do not really support the usage of complex
matching.
There are several retrieval paradigms used in
Visual Information Retrieval. When text annotation is available, it can
directly be used for keyword-based searches. In many situations, text
annotation does not exist or it is incomplete.
When text annotation is unavailable, we must
turn to content-based retrieval methods. When using content-based retrieval
methods, search is performed on features derived from the raw visual media such
as the colour or texture. The VIR paradigms include querying for similar
images, sketch queries and iconic queries. In similar image queries the user
selects a query image, and the system gives a set of similar images to the
query image. In sketch-based queries, the user manually sketches a skeleton,
which will be the base of the query. When using iconic search, the user places
symbolic icons where the visual features should be.
In my talk, I will
introduce strategies and paradigms used by the most popular image querying
systems. Based on my recent results, some possible directions for future
research will be discussed as well.