Image analysis – Image segmentation – Region labeling
Patent
1995-10-12
1998-06-30
Boudreau, Leo
Image analysis
Image segmentation
Region labeling
382171, 382228, G06K 934
Patent
active
057745811
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
The invention relates to a device for segmenting a discrete assembly of data corresponding to sites (m,n), with m being between 1 and M inclusive and n being between 1 and N inclusive, based on digital input signals corresponding to said data. This invention is particularly applicable in the field of image segmentation for realising, for example videophone-type transmissions or for treating images supplied by satellites or images obtained in the medical field or for composite pictures, and thus relates to all sorts of images of the conventional type or images taken from a distance (between the pick-up device and the object of the scene which is most proximate), as well as images for distributing grey levels, and texture images or images combining several types of structures.
Every image segmentation process is used for classifying, i.e. regrouping the pixels by assigning a common label to them per region obtained from the segmentation. Such an operation consists of constructing a model from observable data, which model facilitates subsequent treatments. However, rather than constructing a deterministic model, it is preferred to search and then evaluate probabilities relating to situations which are reflected by these observable data with a likelihood which is the mark of the degree of confidence in this particular interpretation. In setting up such a non-deterministic model, it cannot be excluded that in the searched classification each pixel is influenced by the class of relationship with its neighbours. Based on this hypothesis, it has been contemplated to consider the assembly of pixels of an image as a Markov field (i.e. as the generalization of a Markov chain).
The article "Segmentation of textured images using a multiresolution approach" by C. Bouman and B. Liu at the "1988 International Conference on Acoustics, Speech and Signal Processing" (held in New York USA from 11 to 14 Apr. 1988) and published in the "Reports" of this Conference, Volume 2, pp. 1124-1127, presents a method of segmenting images based on an approximated estimation of the maximum value a posteriori of a random Markov field, on the basis of the hypothesis that an a priori distribution of labels assigned to pixels is given by a Gibbs distribution. This method proceeds through segmentations which are increasingly more precise and the process of segmentation with a growing image resolution is interrupted when each pixel can be assigned to a class.
SUMMARY OF THE INVENTION
It is an object of the invention to provide a device for segmenting data and notably digital signals corresponding to images, with which a segmentation method is carried out which also makes use of a model by way of a Markov field but is based on a different approach and leads to an optimum result with very few iterations.
To this end, the invention relates to a segmentation device which is characterized in that it comprises, in series: computing the average values from T classes and classification by searching that average value which is most approximate to the computed signal; said labels on the one hand and said digital input signals on the other hand and comprising: for determining conditional probabilities p(q/a,b,c), comprising in series: identity or non-identity of the levels corresponding to each of the fifteen possible configurations within a bidimensional square model (q,a,b,c) of a Markov chain, q being referred to in said segmentation assembly as the label to be determined at the "infant" site, and a, b, c being referred to as the labels of the "parent" sites; model thus chosen; p(q/abc)=exp L(q/abc)/Z(abc), in which Z is a normalization factor and in which L(q/abc) is a factor propagating the influence from the parents to the infants, defined by the expressions (3) to (14): ##EQU1## (2) also at the output of the initial segmentation sub-assembly and in parallel with the first path, a second path for determining conditional probabilities p.sub.q (x.sub.mn), comprising in series: of said T classes, probabilities, a sub-asse
REFERENCES:
patent: 4847786 (1989-07-01), Wang et al.
patent: 5239596 (1993-08-01), Mahoney
"Segmentation of Textured Images Using a Multiresolution Approach" by C. Bouman and B. Liu, 1988 International Conference on Acoustics, Speech and Signal Processing, Published in Reports of this Conference, vol. 2, pp. 1124-1127.
Devijver Pierre Alain
Fassnacht Carola
Boudreau Leo
Gathman Laurie E.
Tran Phuoc
U.S. Philips Corporation
LandOfFree
Device for segmenting a discrete assembly of data does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Device for segmenting a discrete assembly of data, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Device for segmenting a discrete assembly of data will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-1868762