Chair: Bernard C. Levy, University of California at Davis (USA)
Zoltan Kato, INRIA
Josiane Zerubia, INRIA
Marc Berthod, INRIA
Wojciech Pieczynski, INT (FRANCE)
This paper deals with the problem of unsupervised Bayesian segmentation of images modeled by Markov Random Fields (MRF). If the model parameters are known then we have various methods to solve the segmentation problem (Simulated Annealing, ICM, etc...). However, when they are not known, the problem becomes more difficult. One has to estimate the hidden label field parameters from the available image only. Our approach consists of a recent iterative method of estimation, called Iterative Conditional Estimation (ICE), applied to a monogrid Markovian image segmentation model. The method has been tested on synthetic and real satellite images.
Cassandra Swain, Vanderbilt University
Tsuhan Chen, AT&T Bell Laboratories (USA)
Foreground and background features are focused (or defocused) differently in an image because corresponding objects are at different depths in the scene. This paper presents a novel approach for segmenting foreground and background in video images based on feature defocus. A modified defocus measurement that distinguishes between high- contrast defocused edges and low-contrast focused edges is presented. Defocus-based segmentation is desirable because defocus techniques are computationally simple. Results indicate that the foreground is easily segmented from moving background. This approach, coupled with motion detection, can segment complex scenes containing both moving background and stationary foreground.
Santhana Krishnamachari, University of Maryland (USA)
Rama Chellappa, University of Maryland (USA)
A multiresolution model for Gauss Markov random fields is presented. Coarser resolution sample fields are obtained by either subsampling or local averaging the sample field at the fine resolution. Although Markovianity is lost under such resolution transformation, coarser resolution non-Markov random fields can be effectively approximated by Markov fields. We use a local conditional distribution invariance approximation to estimate the parameters of the coarser resolution processes from the fine resolution parameters by minimizing the Kullback-Leibler distance between the local conditional distributions. This multiresolution model is used to perform texture segmentation.
F. Salzenstein, Institute National des Telecommunications (FRANCE)
W. Pieczynski, Institute National des Telecommunications (FRANCE)
The aim of our paper is to present a new unsupervised Bayesian image segmentation method using a recent model by Hidden Fuzzy Markov Fields. The main problem of parameter estimation is solved using a recent general method of estimation regarding hidden data, called Iterative conditional Estimation (ICE). This has been successfully applied in classical Hidden Markov Fields based segmentations. The first part of our work involves estimating the parameters defining the Markovian distribution of the fuzzy picture without noise. We then combine this algorithm with the ICE method in order to estimate all the parameters of the noisy picture.
Mary L. Comer, Purdue University (USA)
Edward J. Delp, Purdue University (USA)
In this paper we present a new algorithm for segmentation of noisy or textured images using a multiresolution Bayesian approach. Our algorithm is different from previously proposed multiresolution segmentation techniques in that we use a multiresolution Gaussian autoregressive (AR) model for the pyramid representation of the observed image. Our algorithm also approximates the ``maximization of the posterior margi- nals'' (MPM) estimate of the pixel class labels at each resolution, from coarsest to finest, unlike previously proposed techniques, which have been based on MAP estimation. Experimental results are presented to demonstrate the performance of the new algorithm.
S. Pagnan, Istituto di Automazione Navale
C. Ottonello, D.I.B.E. Universita of Genoa (ITALY)
The paper describes a cumulant-based classifier whose discrimination criterion exploits statistical signal characteristics of higher order than the second one. The performances of the classifier were tested on texture images. In texture classification, the discrimination criterion is usually based on structural characteristics (edge density, co-occurrence matrix) or on statistical parameters (Gaussian-Markov random fields, fractal dimension) of sample textures. As an alternative to statistical approaches, in this paper a third-order cumulant-based criterion is applied and the classifier's performances on images affected by different types of noise are assessed.
Stephen R. Titus, University of Michigan (USA)
Alfred O. Hero III, University of Michigan (USA)
Jeffrey A. Fessler, University of Michigan (USA)
We give estimation error bounds and specify optimal estimators for continuous, closed boundary curves in an NMR image. The boundary is parameterized using periodic B-splines. A Cramer-Rao lower bound on mean square estimate error in the presence of system smoothing and Gaussian noise is derived, and the performance of maximum likelihood and penalized maximum likelihood estimators is compared to this bound. Finally, we comment on the usefulness of estimates of the boundary for providing anatomical side information in the reconstruction of functional tomographic images like those of a PET or SPECT system.
Raynard O. Hinds, Massachusetts Institute of Technology
Thrasyvoulos N. Pappas, AT&T Bell Laboratories (USA)
We present a Bayesian approach for segmenting a sequence of gray-scale images to obtain a binary sketch. We extend a 2-D algorithm to video sequences. The 2-D algorithm is an adaptive thresholding scheme that uses spatial constraints and takes into consideration the local intensity characteristics of the image. We model the segmentation distribution as a 3-D Gibbs Random Field. We add temporal constraints and temporal local intensity adaptation to ensure a smooth transition of the segmentation from frame to frame. For computational efficiency as well as performance we use a multi-resolution approach. We also consider several suboptimal implementations to reduce the delay as well as the amount of computation. We tested the performance of the algorithm on head and shoulders video sequences. The algorithm achieves accurate rendering of the lip and eye movements and preserves the main characteristics of the face, so that it is easily recognizable.
Jae Gark Choi, Korea Advanced Institute of Science & Technology (KOREA)
Si-Woong Lee, Korea Advanced Institute of Science & Technology (KOREA)
Seong-Dae Kim, Korea Advanced Institute of Science & Technology (KOREA)
This paper presents a segmentation and motion estimation method for object-oriented analysis-synthesis coding. A major difficulty in estimating general motion is that it requires a large area of support in order to achieve a good estimation. Unfortunately, when the supporting area is large it is very likely to have multiple moving objects. To solve this problem, we propose a multi-stage segmentation method which is based on optical flow. The basic concept is to group homogeneous subregions with respect to simpler mapping model into large homogeneous regions with respect to more complex mapping model. By applying a hierarchy of mapping parameter model progressively, we can segment the whole changed region into several parabolic patches. Especially person's face in head-and-shoulder images can be described as one object.
Keren O. Perlmutter, Stanford University (USA)
Robert M. Gray, Stanford University (USA)
Richard A. Olshen, Stanford University (USA)
Sharon M. Perlmutter, Stanford University (USA)
A Bayes risk weighted vector quantizer (Bayes VQ) combines compression and low-level classification of images by incorporating a Bayes risk component into the distortion measure used to design the code. The class posterior probabilities required for the Bayes risk computation can be estimated based on a labeled training sequence. We here introduce two new methods for estimating these posteriors. In particular, two types of tree-structured estimators are constructed by applying the classification and regression tree algorithm CART(TM) to eight features of the training sequence. We apply the resulting Bayes VQ systems to aerial photographs where the goal is to compress the images and classify man-made and natural regions. These systems provide classification superior to that of previous work with Bayes VQ while maintaining similar compression performance. The systems also provide moderate to substantial improvement in classification with only a small loss in compression to performance obtained with a modified version of Kohonen's ``learning vector quantizer'' and with an independent design of quantizer and classifier.