Researchers of Penn State has developed a statistical approach called Automatic Linguistic Indexing of Pictures in Real Time or ALIPR which can be a next major step insearching for pictures on Internet.
Technology Used By ALIPR
This technology teaches computers to recognize contents of pictures, such as peoples, landscapes, buildings, parks etc. opposite to the current technology of image-retrieval in which photographs are searched by keywords in the surrounding text such as ALT text. The researchers are hoping that soon this technology can be used for automatic tagging as a part of Internet search engine.
Jia Li - Associate Professor of Statistics at Penn State gives explanation about their approach.
The basic approach is a take a large number of photos (they have started with 60,000), and tag these photos with with a variety of keywords, manually. Let's say, take 100 photos of national parks and tag them with keyword: National Parks, Landscape and Trees.
After that they would build a statistical model which will teach computers to recognize color and texture pattern in these 100 photographs and then assign these keywords to those pictures that seems to contain parks, landscapes and trees. Eventually the process will be reversed so that a internet surfer can use keywords to search the World Wide Web for relevant images.
Problem With Current Image Retrieval Systems
Most of the image-retrieval systems used today, search for keywords in the text associated with the photo or in the name that was given to the photo. But with this technique the surfer often misses appropriate photos and gets inappropriate images.
This new technique of Jia Li can train computers to recognize the semantics of images based on pixel information alone.
Keywords Approach and Accuracy
According to Jia Li- developer of ALIPR says that their approach appropriately assigns to photos at least one keyword among seven possible keywords about 90 percent of the time. But, the accuracy rate really depends on the evaluator. "It depends on how specific the evaluator expects the approach to be," she said. "For example, ALIPR often distinguishes people from animals, but rarely distinguishes children from adults."
Now the team is working on improving the accuracy of ALIPR, but according to Li it is not easy to achieve 100% accuracy. As there are so many images on Internet and they have so much variations that it is not possible for ALIPR to be 100% accurate each time.
Reasons for Inaccuracy
We can understand this by a simple example. Suppose in an image there is a Cat wearing a Red Coat.....then Red Coat will lead ALIPR to tag the photo with words irrelevant to the Cat. There is too much variations in the images that will cause problems for ALIPR. But Li is working on some new ideas to achieve better recognition of image sementics.
