Learning low-level vision-英文文献.pdfVIP

Proceedings of the IEEE Workshop on Statistical and Computational Theories of Vision, Fort Col lins, CO, June, 1999. (Published on Web). Learning low-level vision William T. Freeman 1 and Egon C. Pasztor 1 A bstract We show a learning-based metho d for low-level vision problems. We set-up a Markov network of patches of the image and the underlying scene. A factorization approximation allows us to easily learn the parameters of the Markov network from synthetic exam- ples of image/scene pairs, and to eciently propagate image information. Monte Carlo simulations justify this approximation. We apply this to the \sup er-resolution problem (estimating high frequency details from a low-resolution image), showing go o d results. For the motion estimation problem, we show resolution of the ap erture problem and lling-in arising from application of the same probabilistic machinery. This is also distributed as MERL TR99-12, /rep orts/TR99- 12/index.html . For related technical rep orts, please see /rep orts/TR99-08/index.html and /rep orts/TR99-05/index.html . 1. MERL, a Mitsubishi Electric Res. Lab. 201 Broadway Cambridge, MA 02139 freeman, pasztor@ . c Copyright Mitsubishi Electric Information Technology Center America, 1999, all rights reserved. 1 1 Intro duction We seek a machinery for learning low-level vision problems. These problems could include: motion analysis, inferring shap e and alb edo from a photograph, or estimating colors. For these problems, given image data, we want to estimate an underlying scene. The scene quantities to b e estimated could b e pro jected ob ject velo cities,