Image analysis – Image segmentation
Patent
1998-03-23
2000-06-27
Rogers, Scott
Image analysis
Image segmentation
382199, 382203, 382204, 382260, G06K 934, G06K 946, G06T 500, G06T 760
Patent
active
060816171
DESCRIPTION:
BRIEF SUMMARY
FIELD OF INVENTION
This invention concerns methods and apparatus for processing data, particularly but not exclusively images, in order to separate the signal into two or more components having distinct characteristics, for example to separate different sized objects in an image or to remove a noise component from a signal. The invention is of particular application in the fields of image analysis, image compression, pattern recognition and the like.
BACKGROUND TO THE INVENTION
A common image processing operation is to segment an image into regions having distinct characteristics, such as the different objects in an image. For example, in an image compression scheme the image may be segmented into relatively large regions which can be described in terms of their boundaries and additional characteristics such as texture. This enables the image to be transmitted more efficiently. In pattern recognition and image analysis, an image may be decomposed into the major objects in the scene, and then information describing these objects, such as their edges, is passed to subsequent processing.
In order to separate an image into objects of different scale (size) it is common to perform a multiscale decomposition using some form of spatial filtering (for an extensive review see Lindeberg, 1994 Scale-space theory in computer vision"). In this approach, filters are used to produce images that have been filtered to a range of scales (so objects of smaller scale than the filter are attenuated or removed). The advantages of these multiscale representations crystallised in the 1980's when the concept of `scale-space` emerged (Witkin, 1983, "Scale space filtering" 8th Int Joint Conf Artificial Intelligence).
The key point about scale-space representations is that objects in a scene are usually represented by local intensity extrema (maxima or minima) bounded by edges. A large object can be located either by finding the corresponding extremum (zero crossings of the first derivative) or the corresponding edges (zero crossings of the second derivative). However, this is only reliable if the image is first smoothed with a filter in order to remove fine detail and noise. It is not sufficient to apply any filtering (convolution) operator for it is very desirable that the smoothing does not introduce any new extrema and that the intensity of fine scale features is attenuated monotonically as the filtering scale increases. This property is known as scale-space causality, and is fundamental to an effective scale-space decomposition for use in image analysis etc. Furthermore, most filtering operators do not achieve scale space causality.
Initially, attention was focused on one dimensional signals and it was shown that the only linear (convolution) filters that exhibit this scale space property are Gaussian (diffusion equation related) convolution filters (Yuille and Poggio, 1987 "Scaling Theorems for Zero Crossings", IEEE Trans on Pattern Analysis and Machine Intelligence 9:15-25). Gaussian convolution, implemented as a set of increasing scale difference of Gaussians, Laplacian of Gaussians etc, had already been proposed as a way to decompose a two-dimensional image into a number of scale related "spatial" channels (Marr 1982 Vision" W. H. Freeman and Co). These methods were used to segment an image into regions and, by higher level classification stages, to analyze the image content. However, they only approximated the scale-space causality property for multidimensional signals.
Later it was realised that Gaussian convolution in one dimension simply represents a particular solution to the heat or diffusion equation (Koenderink, 1984, "The structure of Images", Biological Cybemetics 50:363:370). Solving these equations using a 2-D formulation leads to a method of performing a scale-space decomposition in 2-D that conforms to the scale-space causality property. This approach is known as `diffusion based` imaging. However, there are problems with the diffusion based approach. The two most apparent ones are that: 1) At large scales
REFERENCES:
patent: 4183013 (1980-01-01), Agrawala et al.
patent: 4991224 (1991-02-01), Takahashi et al.
patent: 5048096 (1991-09-01), Beato
patent: 5235434 (1993-08-01), Wober
patent: 5424783 (1995-06-01), Wong
patent: 5577131 (1996-11-01), Oddou
patent: 5751852 (1998-05-01), Marimont et al.
patent: 5867593 (1999-02-01), Fukuda et al.
patent: 5995668 (1999-11-01), Corset et al.
J, Serra and P. Salembier, "Connected Operators and Pyramids", SPIE Proceedings vol. 2030, Jul. 12-13, 1993, pp. 65-76.
J. Crespo and R. Schafer, "Image Partition Using An Iterative Multi-Resolution Smoothing Algorithm", IEEE Int'l Conference, Mar. 23-26, 1992, pp. 561-564.
Bangham James Andrew
Bragg Nigel Lawrence
Young Robert William
Cambridge Consultants Limited
Rogers Scott
LandOfFree
Data processing method and apparatus does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Data processing method and apparatus, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Data processing method and apparatus will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1790755