Friday, 26 Feb 1999.
Blobworld: Region-based image retrieval
Chad Carson, U.C. Berkeley
Image database users want to find objects in images, but most current systems retrieve images based only on low-level features such as global color and texture histograms. I will describe Blobworld, a new approach to image retrieval that approaches object-level queries.
Blobworld is based on finding coherent image regions which roughly correspond to objects. the image is segmented into regions by fitting a mixture of Gaussians to the pixel distribution in a joint color-texture-position feature space. Each region ("blobl") is then associated with color and texture descriptors. Querying is based on the user specifying attributes of one or two regions of interest, rather than a description of the entire image.
Experiments indicate that queries for distinctive objects have much higher precision using Blobworld than using global image features. Blobworld querying is also more intuitive than global-feature querying because it allows the user to interact with the internal representation of the image; this helps the user formulate effective queries and understand their results.
I will briefly discuss approaches to indexing Blobworld feautes in order to allow fast access to large collections of images.
This is joint work with Serge Belongie, Jitendra Malik, Megan Thomas, Joe Hellerstein, Ray Larson, Ginger Ogle, and Joyce Gross.