## Learning to Parse Images (2000)

### Cached

### Download Links

- [www.gatsby.ucl.ac.uk]
- [www.cs.berkeley.edu]
- [www.gatsby.ucl.ac.uk]
- DBLP

### Other Repositories/Bibliography

Citations: | 2 - 0 self |

### BibTeX

@TECHREPORT{Teh00learningto,

author = {Yee Whye Teh},

title = {Learning to Parse Images},

institution = {},

year = {2000}

}

### OpenURL

### Abstract

Segmentation and recognition are important subtasks of image interpretation. The general approach to statistical image interpretation tries to solve each separately. This introduces various problems as segmentation and recognition are interrelated. These problems can be avoided by solving segmentation and recognition simultaneously. This is achieved by viewing segmentation and recognition as subtasks of finding the correct parse tree of the image, and viewing image interpretation as a search in the space of parse trees. Credibility networks are an instantiation of this idea. They are graphical models which describe a probability distribution over all possible parse trees with the leaves corresponding to pixels. The parameters of a credibility network can be learned and inference can be achieved using meanfield approximations. During inference, the results of segmentation and recognition are iteratively improved upon. Simulations showed that credibility networks can perform interesting toy problems, for example hand-written digit classication and segmentation.