The  Weizmann  Institute  of  Science
                 Faculty of Mathematics and Computer Science


                         Vision and Robotics Seminar

                                 Yacov Hel-Or
                        Department of Computer Science
                     The Interdisciplinary Center  (IDC)


                                will speak on


               A Discriminative Framework for Wavelet Denoising

Abstract:
In recent years, "Transform based de-noising" has received a great deal of
attention  due to its simplicity and excellent performance.  According to this
approach, a noisy image is de-noised in three stages:

1. The noisy image is first transformed using a predefined transform basis
(usually some type of Wavelet transform).

2. The transform coefficients are modified using a set of lookup-tables (LUT).

3. The modified coefficients are transformed back to the image domain, yielding
an estimate of the original (noise-free) image.

The main challenge using this approach is designing a set of LUTs such that the
restoration results are optimal. Over the last decade or so numerous studies
dealing with this problem were introduced. All of these studies followed a
generative approach in which the LUTs are calculated based on models that are
postulated as statistical prior for natural images.

In this work a new and practical scheme for transform based de-noising will be
introduced.  The suggested approach takes a discriminative viewpoint in which a
set of LUTs are calculated based on a class of image examples with known noise
components. The LUTs are designed to optimally discriminate between the noise
and the data components in this example class.  The obtained LUTs are
demonstrated to be different than the classical soft/hard thresholding in the
over-complete transform cases, while denoising results exhibit state-of-the-art
performance.

Since the developed approach does not require modeling of image priors nor of
the degradation process, it can be implemented seamlessly for other image
restoration problem, such as: image deblurring, JPEG artifact removal and
more.

Joint work with Doron Shaked, HP Labs - Haifa, Israel.





                     The lecture will take place in the
                               Faculty Room 141
                        on Thursday, January 24, 2008
                                12:00 - 13:00