Browse Prior Art Database

Method for model-based denoising of images and image sequences

IP.com Disclosure Number: IPCOM000010727D
Publication Date: 2003-Jan-15
Document File: 6 page(s) / 175K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method for model-based denoising of images and image sequences. Benefits include improved functionality and improved reliability.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 21% of the total text.

Method for model-based denoising of images and image sequences

Disclosed is a method for model-based denoising of images and image sequences. Benefits include improved functionality and improved reliability.

Background

        � � � � � Image denoising provides the user with an enhanced, less-noisy image that shows faint structures more clearly than in the raw image sequence. For single images, denoising is limited to the spatial domain and various techniques exist to remove noise by spatially averaging image pixels in local neighborhoods. However, standard isotropic smoothing calculates the brightness at individual image points (pixels) as the weighted average across a local neighborhood surrounding the pixels. Although this technique reduces the noise variance, it also causes brightness patterns to be blurred and smeared out. In particular, in low signal-to-noise (S/N) applications, such as focused ion beam (FIB) images of silicon structures, the faint brightness patterns of interest are lost.

        � � � � � Blurring of brightness structures can be reduced by using anisotropic diffusion filtering for directional, structure-preserving denoising. The underlying idea is to replace isotropic smoothing filter kernels by directional sensitive kernels that are adapted to local image structure. These techniques average pixels predominantly along iso-brightness contours and avoid averaging across edge features in the image. Several algorithmic approaches to this problem exist.

        � � � � � For example, two noisy images (see Figure 1, a and c) can be denoised using robust, anisotropic smoothing (see Figure 1, b and d). Let f(x,y) be a noisy image over the spatial coordinates x and y. The goal of image denoising is to find a reconstructed (denoised) image g(x,y) over the same spatial domain. Given the value of f and a set of criteria, g can be calculated by minimizing the following objective function, E(g,f), with respect to g:

E(g,f) = l1� ES(g,f) + l 2� ED(g,f) + l 3� ET(g,gt-1)

        � � � � � The value ES(g,f) = ρ(g-gn, σ1) is a spatial smoothness constraint across neighboring pixels. The value gn1ED(g,f) = ρ(f-g, σ2) is a data similarity constraint. The value ET(g,gt-1) = ρ(g-gt-1, σ1) is a temporal coherency constraint, and ρ(.) denotes a robust error norm.

        � � � � � While this and similar techniques yield significant improvement of signal-to-noise ratio in images, these techniques are not exploiting model-based information provided by the application domain. For this reason conventional techniques for anisotropic diffusion filtering still suffer from the following problems:

•        � � � � To preserve edge features, spatial smoothness constraints enforce conservation of image brightness across regions of large brightness variation. Because these edge regions must be identified from the image data without prior knowledge, noisy edges remain unmodified. This technique yields jagged and fuzzy looking edges in low S/N image data (see Figure 1, d). Furthermore, spurio...