Cyberia: CBIR

Content Based Image Retireval

Content-based image retrieval (CBIR) systems use features that can be extracted from the image files themselves, for use in searching a collection of images, rather than relying on manual indexing or text descriptions by humans.

We extend this by using some simple image transformations to create several additional representations of the images in the collection, some or all of which may be used in an information retrieval task on the collection We refer to this as multi-channel, as opposed to conventional single channel, CBIR.

We have implemented an example system employing an existing basic CBIR feature-extraction technology, using three color and four texture features, from images drawn from 34 categories of the COREL image collection. From these we created four channels, the original images, their negatives, their grayscale images, and the negative of the grayscale. We have found that the combination (using a simple similarity-based merge) of the results of all four channels is a material improvement (22% greater) on the precision of the original single channel case.

Compare the top forty results from a single channel with the top ten results of each of four channels.

Valid Query IDs are: 290xx, 700xx, 750xx, 770xx, 840xx, 1040xx, 1050xx, 1070xx, 1080xx, 1090xx, 1100xx, 1120xx, 1340xx, 1350xx, 1680xx, 2680xx, 2890xx, 3260xx, 3510xx, 3540xx, 3910xx, 4150xx, 4470xx, 4580xx, 4990xx, 5210xx, 5350xx, 5530xx, 5810xx, 5910xx, 6440xx, 6550xx, 6610xx, 6840xx for xx=00-99

Cyberia | People | Papers | Posters | Presentations | Testbed | Contact