Image analysis – Pattern recognition – Classification
Reexamination Certificate
1999-09-13
2003-10-07
Dastouri, Mehrdad (Department: 2623)
Image analysis
Pattern recognition
Classification
C382S195000, C382S240000, C382S257000, C382S308000
Reexamination Certificate
active
06631212
ABSTRACT:
FIELD OF THE INVENTION
The present invention generally relates to photo segmentation and object extraction using texture classification performed with a novel combination of multiresolution simultaneous autoregressive texture segmentation and a wavelet classification.
BACKGROUND OF THE INVENTION
Texture is a fundamental characteristic of natural images that, in addition to color, plays an important role in human visual perception and in turn provides information for image understanding and scene interpretation. A number of texture features have been developed to represent texture characteristics for texture classification and segmentation, including multiresolution simultaneous autoregressive models (MSAR) (Mao et al. “Texture Classification and Segmentation Using Multiresolution Simultaneous Autoregressive Models,”
Pattern Recognition
, Vol. 25, No. 2, pp. 173-188, 1992), Markov random field (MRF) models (Panjwani et al. “Markov Random Field Models for Unsupervised Segmentation of Texture Color Images,”
IEEE Trans. Pattern Anal. Mach. Intel
., Vol. 17, pp. 939-954, October 1995), Gabor filters (Weldon et al. “Integrated Approach to Texture Segmentation Using Multiple Gabor Filters,”
In Proc. IEEE Int. Conf. Image Process
., 1997), and wavelet coefficients (Unser, “Texture Classification and Segmentation Using Wavelet Frames,”
IEEE Trans. Image Process
., Vol. 4, pp. 1549-1560, November 1995; and Porter et al. “A Robust Automatic Clustering Scheme for Image Segmentation Using Wavelets,”
IEEE Trans. Image Process
., Vol. 5, pp. 662-665, April 1996), all of which articles are incorporated herein by reference.
PROBLEM TO BE SOLVED BY THE INVENTION
While each of these texture features has certain unique advantages, one common problem in using them for image segmentation is that noise in the extracted features may result in misclassification, that takes the form of holes and other fragments. Another interesting, yet challenging, problem in texture segmentation is what may be called the boundary effect. It usually appears as inaccurate segmentation of boundaries or superfluous narrow regions between two textures. It is conjectured that such a boundary effect is caused by misclassification when the trajectory of the feature vectors makes a transition through feature space (Weldon et al. supra). To make matters worse, the misclassification may be interpreted as a third texture, depending on the nature of the features present in a particular image.
Through the invention, a two-stage segmentation approach is utilized to identify and reclassify potential problem regions. Below, the inventive initial segmentation method is described, which is an improved way of normalizing MSAR features.
SUMMARY OF THE INVENTION
It is, therefore, an object of the present invention to provide a structure and method for photo segmentation and object extraction. As would be known by one ordinarily skilled in the art given this disclosure, the inventive two-stage scheme can be generalized beyond MSAR and wavelet features. The criterion for selecting two sets of features could be based on the requirements that the first set of features should have high signal-to-noise ratio in the interior of homogeneous textured regions, and the second set of features should have high spatial resolution.
A single set of features is unlikely to have both properties due to the fact that texture is an area-oriented characteristic and texture features are computed within a local neighborhood. On one hand, features based on larger local neighborhood windows capture longer correlation and therefore are more robust, but otherwise suffer from poor spatial resolution near boundaries between textures. On the other hand, features based on smaller local neighborhood windows tend to have better spatial resolution but may not be as robust to noise and subtle textural differences.
An embodiment of the invention is a method of segmenting textures in an image which includes a first process of identifying first features in the image (the first process has a first spatial support), preparing an uncertainty map of the first features (the uncertainty map includes first confidence pixels and second confidence pixels, and the first confidence is higher than the second confidence), a second process of identifying second features in the image (the second process having a second spatial support smaller than the first spatial support), preparing a classifier based on the first confidence pixels and the second features, and reclassifying the second confidence pixels based on the classifier to segment the textures.
The first spatial support has a signal to noise ratio higher and a resolution lower than that of the second spatial support. The inventive method further includes weighting the first features based on a ratio of between-class variance to within-class variance of the image and edge-filtering variance within the image to remove feature inhomogeneity associated with texture boundaries. The first process of identifying the first features includes a multiresolution simultaneous autoregressive (MSAR) modeling operation. The second process of identifying the second features includes a wavelet modeling operation. After the MSAR modeling operation, MSAR coefficients are clustered using unsupervised segmentation. The inventive method further includes, after the clustering of the MSAR coefficients, a morphological erosion operation.
Another embodiment of the invention is the segmenting of textures in an image which includes computing multiresolution simultaneous autoregressive (MSAR) features in the image, preparing an uncertainty map of the MSAR features (the uncertainty map includes first confidence pixels and second confidence pixels, and the first confidence is higher than the second confidence), computing wavelet features in the image, preparing a classifier based on the first confidence pixels and the wavelet features, and reclassifying the second confidence pixels based on the classifier to segment the textures.
Another embodiment of the invention is the segmenting of textures in an image which includes computing multiresolution simultaneous autoregressive (MSAR) features in the image, preparing a confidence map of the MSAR features, computing wavelet features in the image, producing a refined confidence map based on the wavelet features, and segmenting the textures based on the refined confidence map.
A further embodiment of the invention is a computer system for segmenting textures in an image which includes a first processor identifying first features in the image (the first processor has a first spatial support), a mapping unit preparing an uncertainty map of the first features (the uncertainty map includes first confidence pixels and second confidence pixels, and the first confidence is higher than the second confidence), a second processor identifying second features in the image (the second processor has a second spatial support smaller than the first spatial support), a classification unit forming a classifier based on the first confidence pixels and the second features, and a segmentor reclassifying the second confidence pixels based on the classifier.
The inventive two-stage segmentation scheme is designed to play to the strength of two sufficiently different sets of texture features while overcoming their drawbacks. Combined with an efficient way of identifying “uncertain” regions where misclassification is likely to occur, and a self-supervised training mechanism, the invention offers an alternative superior to postprocessing by morphological operations alone.
REFERENCES:
patent: 5815198 (1998-09-01), Vachtsevanos et al.
patent: 6105015 (2000-08-01), Nguyen et al.
Zhang et al, Wavelet-based Multiresolution Stochastic Image Models, Nov. 1995, IEEE Proceedings ISBN: 0-8186-7190-4, pp. 479-484.*
Bouman et al, Multiple Resolution Segmentation of Textured Images, Feb. 1992, IEEE transactions ISBN: 0162-8828, vol. 13, No. 2, pp. 99-113.*
“Texture Classification and Segmentation Using Multiresolution Simultaneous Autoregressive Models
Luo Jiebo
Savakis Andreas E.
Dastouri Mehrdad
Wood David M.
LandOfFree
Twostage scheme for texture segmentation based on clustering... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Twostage scheme for texture segmentation based on clustering..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Twostage scheme for texture segmentation based on clustering... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3153658