Improving internet information retrieval

Mar 23, 2007

Kettering Computer Science Lecturer Rajeev Agrawal works with colleagues from the University of Michigan-Dearborn and Wayne State University to refine internet searches that get it right the first time.

Once, the internet was called the information superhighway. But today, that superhighway is often gummed up with useless information derived from multiple sources that never quite get to what we're searching for.

Most readers are cognizant of this situation. After we punch a word or phrase into a searchable, online internet site such as Google or Yahoo, we're often faced with long lists of potential subject matter that require examination to narrow our specific search. This can be daunting and time consuming to say the least.

But based on the research efforts of a Kettering faculty member and his colleagues at the University of Michigan and Wayne State University, future searches on the internet could become much more efficient and helpful.

Rajeev Agrawal, a lecturer of Computer Science who is currently completing his Ph.D. at Wayne State University in Detroit, is developing a way to narrow the gap between content and context in multimedia internet searches. Working with William Grosky of the University of Michigan-Dearborn and Farshad Fotouhi of Wayne State University, Agrawal and his team hope to improve the gap between low-level visual features (color and texture, for example) and semantic concepts mediated by an image to yield more specific and improved internet search results. They recently presented their findings at the First International Conference on Semantic and Digital Media Technologies (SAMT) 2006 in Athens, Greece, in December and their paper, titled "Image Retrieval Using Multimodal Keywords" won the SAMT 2006 Conference Best Paper Award. The Image, Video and Multimedia Systems Laboratory of the National Technical University of Athens hosted the event.

"In the past, researchers have attempted to use region-based features that represented images at object-level, or using relevance feedback (RF) techniques that incorporate a user's perception of the images," Agrawal said.

A good example of this would be if a person performed a search on the term dog. Typically, a series of photos would appear, as well as text and other information pertaining to dogs. This search uses global features, which, according to Argawal, do not adequately capture all important properties of an image. When users engage in an internet search, the system tries to include all information pertaining to the search word, thus creating an extensive compilation of information to examine. As a result, users must study all strings of retrieved information to find the image or content they require for their specific needs.

The paper that Agrawal and his colleagues presented at this international conference proposes the creation of a system that uses both visual keywords derived from low-level MPEG-7 color features and textural annotations to make searches more efficient and specific for users. MPEG-7 is a standard for describing multimedia content data that supports interpretations of semantics determination, which is accessible by computer code. Currently, internet sites represent images using single sets of global features, which broaden one's searches beyond the scope of what an individual needs. Agrawal and his colleagues overcome this problem by dividing an image into several sub images called tiles using a template of a given size. The team basically considers each image as a document and each template region as a visual keyword. This means that each image is represented by multiple template regions and each tile is represented using the MPEG-7 scalable color, color structure and color layout descriptors. According to the group's paper, "these descriptors have been proven to be very efficient in multimedia content-based search and retrieval." Unfortunately, these descriptors also present users with large numbers of distinct tiles.

To reduce this large number of image tiles, Agrawal and his team cluster and treat each tile in the same cluster as if they were the same. Their approach uses the term image matrix and terms consisting of textural and visual keywords. Each visual keyword is a tile representing a specific cluster of similar tiles.

During their experiments to test their hypothesis, the group used a modified latent semantic indexing (LSI) algorithm to help retrieve relevant images. The results of their experiments show that if used together, visual keywords and textural annotations can greatly improve the quality of retrieval results, thereby helping users find what they need through their internet searches more quickly and efficiently.

The group expects to continue their research into this subject and present their findings at conferences in 2007. To learn more about this ongoing research project, contact Rajeev Agrawal at (810) 762-7905.

Written by Gary J. Erwin
(810) 762-9538
gerwin@kettering.edu