Many of us have tried photo search tools like Google Images, only to be frustrated with the outcome. For example, try conducting a search for “tiger” photos. You might be surprised how many times golfer, Tiger Woods, shows up instead of the animal.
Professor Shih-Fu Chang from Columbia University is hoping to change all of that by using pixels from photos to collect information rather than the names the pictures are randomly tagged with.
“There are billions of images on a database like Google,” said Chang. “Similar images tend to have similar semantics, so we hope to use that information to have better results.”
Chang visited Boston University’s Electrical and Computer Engineering Department last week as part of the Distinguished Lecture Series, which brings groundbreaking engineers to the university. He explained his research to faculty, students, staff, and other members of Boston’s engineering community on September 29.
Chang, who is chairman of electrical engineering and director of the digital video and multimedia lab at Columbia, has made significant contributions in multimedia search, media forensics, mobile media adaptation, and international standards.
During his talk, Chang discussed many possibilities to improve image searches. The best solution is still up for debate, but one of the answers may be image retrieval with relevance feedback, a method that involves marking the best and worst image results. Over time, computers could potentially recognize the pixels that more commonly make up the good images.
Chang also discussed another method, visual search via brain state decoding, which can look at how a brain responds to like images. The downside to this method, he said, was that people can only look at pictures for a few minutes at a time before they get tired, so this kind of data will need to be augmented by automatic methods like those using machine vision.
Whatever the answer, Chang said that the demand to solve this problem is great – and the solution is likely to come from electrical and computer engineers.
Chang’s talk was the first in the three-part Fall 2010 Distinguished Lecture Series. The next talk features Professor Mark A. Horowitz of Stanford University who will speak on the topic, “Encapsulating Designer Knowledge: Improving Digital and Mixed Signal Design.” Hear him on October 20, 2010, at 4 p.m. in PHO 211.
-Rachel Harrington (email@example.com)