ALIPR annotates images based on content.
First, it has to learn to recognize the meaning of a tag before
suggesting the correct labels.
As part of the learning process, the researchers feed the computer
hundreds of images of the same topic, for example "sunset." The
computer analyzes the pixels and extracts information related to color
and texture. It then stores a mathematical model for "sunset" based on
the cumulative data.
Later, when a user uploads a new picture of a sunset, the computer compares the pixel information from the pre-computed models in
its knowledge base. In just about a second, ALIPR suggests a list of
15 possible tags. The user has only to check off those appropriate.
"About half the time, the computer's first tag
out of the top 15 tags is correct, and a vast majority of images have
at least one correct tag," said Wang.
If ALIPR does not suggest a helpful tag, the user can type a word in a
space provided, improving the computer's ability to identify tags for future images.
