GSRA Position F09-09
Bridging the Semantic Gap in Image Retrieval
Faculty Name: Michael Farmer
Department: CSEP
Campus Address: 214 Murchie Science Building
Email: farmerme@umflint.edu
Phone: 8107623423
Project Description: Image query is the task of using a computer to automatically find images of particular interest to users. Popular tools such as Google Image have provided a means of allowing users to search large databases of images for the perfect image to suit their needs based on the userâ€TMs selection of keywords. Unfortunately, the Google image search engine and other similar image search engines are not actually searching the contents of the images they have stored, but rather they are searching a collection of pre-recorded keywords that are associated with these images. This approach is clearly limited by the human labor associated with analyzing images and recording keywords. A better approach would be to develop a search engine that could search images directly, similarly to how humans would search an image for contents.
Unfortunately, this is an extremely difficult problem. The ability to automatically search for objects of interest based on their shape without prior annotation, has not been successful for arbitrary images, particularly when the backgrounds may contain clutter such as objects and regions that are not of interest. This type of general image search is called shape-based image retrieval and it is widely regarded as one of the hardest problems in general image retrieval. Shape-based retrieval is particularly difficult due to the difficulty of the segmentation problem, where the object of interest is identified and removed from the background. What is required is a mechanism for bridging the semantic gap by integrating segmentation and classification, thereby allowing the segmentation algorithm to use the knowledge of the object of interest to guide the segmentation. The goal of this project is to extend past research in integrated segmentation and classification to build a shape-based image retrieval system.
I have developed an approach for bridging the gap called the Wrapper Method of image segmentation. It has been demonstrated effective for extracting complex shapes from images for specific applications. This project would extend this research to allow the wrapper to work on a broader class of images, with the ultimate goal being robust general image query.
Semesters Desired: Fall 2009; Winter 2010
GSRA Position Description: The GSRA would be responsible for implementing algorithms in MATLAB to fill in the various missing components of this image query environment. The GSRA would be required to read recommended journal articles and implement particular algorithms defined in these articles. Additionally, the GSRA would be responsible for collecting training images to "teach" the system about various categories of objects for which the system would be able to search. The work is very programming intensive and there are also certain aspects of the work which are highly mathematical.
Specific Day/Time Requirements: No specific days or times.
Special Requirements: Must be proficient in a high-level programming language such as C, C++, Java, or MATLAB. Must also be knowledgeable in the analysis of computer algorithms. Image Processing knowledge a plus.
Graduate Students in These Programs May Apply: Position has candidate in mind