Information Technology Reference
In-Depth Information
interpretation of images should be further developed. Some improvements have
been achieved using relevance feedback techniques. However, these techniques
are still limited, and demand to gather information from sequences of relevance
feedback iterations from different users. Relevance feedback approaches saturate
the provided gain within a few number of iterations and are not automatic. It
makes this approach susceptible to the problem of subjectivity and inconsistency
caused by the typical human restrictions of time, interest or tiredness. However,
the use of relevance feedback brings the human to be part of the process, and if
well implemented, it can really diminish the semantic gap.
On the other hand, development in the field of association rule mining can
effectively help to reduce the semantic gap automatically, where patterns relating
semantic meaning to low-level representation can be found. Advances in feature
selection techniques can also help to reduce the semantic gap by determining
the minimal subset of features that effectively represents semantic information
embedded in complex data.
Another question that should be answered is: how to make computers to
recognize objects in a scene (image) in the same way as humans do? Researchers
from several areas, including medicine, computer science, physics and electrical
engineering, are working together to answer such question. In the future, the
result of this effort might facilitate the human life in several aspects. For instance,
small computer devices that reproduce the electrical stimuli occurred in the brain
when a person is looking to a scene could be developed. Such devices could then
be implanted in blind people's brains allowing them to see a scene, and even
having the perception of colors and depth.
In the future, the image mining field, which is a special case of a complex data
mining field, should follow two main directions:
•
Mining for patterns:
mining of images aimed at finding interesting pat-
terns in a image, scene or sequence of scenes;
•
Mining for search:
mining to support content-based retrieval of images.
The development in the
mining for patterns
direction can lead to the discovery
of other types of data mining tasks. For example, a new future task of data mining
can be the
mapping task
.
Mapping task
can be a future field of data mining
specialized in mapping data, such as space, time and behavior, in trends, objects
and evolution lines. In addition, progress in
mining for patterns
direction will
improve in medical care, agriculture, climate forecast, and space exploration. A
recent demand for the
mining for patterns
research is regarding the improvement
of surveillance and security.
With the development of the Internet, improvements in the
mining for search
direction are mandatory. Nowadays, a small potential of the Internet is effectively
explored. There are very few and limited mechanisms of content-based search
available in the Internet. The users desire to search for images, scenes or movies
having a given subject. Today, well-known problems forbid them to perform
such searches: (a) inconsistence in image descriptions; and (b) low-level features
are not associated to their semantic meaning. New mining techniques should
