Understanding the meaning behind visual data is increasingly important as the quantity of digital images in circulation explodes, and as computing in general and the Internet in specific shifts quickly towards an increasingly visual presentation of data. However, the remarkable amount of variance inside categories (e.g. different kinds of chairs) combined with the occurrence of similarity between categories (e.g. similar breeds of cats and dogs) makes this problem incredibly difficult to solve. In particular, the <i>semantic segmentation</i> of images into contiguous regions of similar interpretation combines the difficulties of object recognition and image segmentation to result in a problem of great complexity, yet great reward. This thesis proposes a novel solution to the problem of semantic segmentation, and explores its application to image search and retrieval. Our primary contribution is a new image information processing tool: the <i>semantic texton forest</i>. We use semantic texton forests to perform (i) semantic segmentation of images and (ii) image categorization, achieving state-of-the-art results for both on two challenging datasets. We then apply this to the problem of image search and retrieval, resulting in the Palette Search System. With Palette Search, the user is able to search for the first time using <i>Query by Semantic Composition</i>, in which he communicates both what he wants in the result image and where he wants it.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:605649 |
| Date | January 2008 |
| Creators | Johnson, M. A. |
| Publisher | University of Cambridge |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
Page generated in 0.0022 seconds