NSF Research Experiences for Undergraduates (REU)
Summer 2007 - Computer Vision and Image Processing
Department of Computer Science, Utah State University

  Name:

  Samuel Barrett

  Major:

  Computer Science

  School:

  Stevens Institute of Technology

  Home Town:

  Johnson City, Tennessee

About Me

I just finished my junior year at Stevens Institute of Technology, so I only have one more year left. I'm majoring in computer science and minoring in both math and music technology. I play both the flute and oboe as a soloist, in small ensembles, and in my school's concert band. I like to play ultimate frisbee and racquetball. I plan on going to graduate school to get a PhD in computer science and eventually do research in artificial intelligence.

Project

This summer, I performed research in the area of content based image retrieval. Most current image retrieval systems like Google image search require users to label images, while content based image retrieval systems attempt to learn about what the image represents directly from the image itself. Learning takes place through relevance feedback where users label returned images as either relevant or irrelevant to the search image. Learning can be broken into two time frames: short term and long term. Short term learning takes place over a single query where images are returned several times and the user's feedback is used to learn which low level features (like colors or textures) are important for the query. Long term learning takes place over many queries and the system stores which images are related to each other.

My work added onto existing techniques in both time frames. For short term learning, I implemented a new system that uses a blocking scheme to segment an image into several images and then uses a method for propogating the user's labels to other images developed by Wu and Yap. For long term learning, I built on Han et. al.'s work. They tracked which images were related to each other by using users' past feedback. However, tracking the image to image relationships is inefficient. Therefore, I created the idea of a semantic concept that would be similar to a label a user might provide. Then, the system found how many semantic concpets were necessary for the database and found if each image was related to each semantic concept.

Results

I found that my short term learning technique of combining an image blocking scheme with label propagation and fuzzy SVM learning was more effective than past techniques. Also, my long term learning technique was far more efficient than Han et. al.'s memory learning scheme, while still producing good results. Finally, both of my proposed learning schemes were more resillient to user mislabelling image than past technique. For more in depth results, please look at my final presentation or final paper

Feel free to write me at: