Catálogo de publicaciones - libros

In this paper we propose an extension to the standard Markov Random Field (MRF) model in order to handle layers. Our extension, which we call a Factorial MRF (FMRF), is analogous to the extension from Hidden Markov Models (HMM’s) to Factorial HMM’s. We present an efficient EM-based algorithm for inference on Factorial MRF’s. Our algorithm makes use of the fact that layers are a priori independent, and that layers only interact through the observable image. The algorithm iterates between wide inference, i.e., inference within each layer for the entire set of pixels, and deep inference, i.e., inference through the layers for each single pixel. The efficiency of our method is partly due to the use of graph cuts for binary segmentation, which is part of the wide inference step. We show experimental results for both real and synthetic images.

We first present an improvement of the for motion segmentation by newly introducing the . We point out that this improvement does not always fare well due to the it introduces. In order to judge which solution to adopt if different segmentations are obtained, we test two measures using real images: the standard test, and the geometric model selection criteria.

Deformable models are by their formulation able to solve surface extraction problem from noisy volumetric images. This is since they use image independent information, in form of internal energy or internal forces, in addition to image data to achieve the goal. However, it is not a simple task to deform initially given surface meshes to a good representation of the target surface in the presence of noise. Several methods to do this have been proposed and in this study a few recent ones are compared. Basically, we supply an image and an arbitrary but reasonable initialization and examine how well the target surface is captured with different methods for controlling the deformation of the mesh. Experiments with synthetic images as well as medical images are performed and results are reported and discussed. With synthetic images, the quality of results is measured also quantitatively. No optimal method was found, but the properties of different methods in distinct situations were highlighted.

In this paper, we propose a general Eulerian framework for region-based active contours named DREAMS. We introduce a general criterion including both region-based and boundary-based terms where the information on a region is named “descriptor”. The originality of this work is twofold. Firstly we propose to use shape optimization principles to compute the evolution equation of the active contour that will make it evolve as fast as possible towards a minimum of the criterion. Secondly, we take into account the variation of the descriptors during the propagation of the curve. Indeed, a descriptor is generally globally attached to the region and thus “region-dependent”. This case arises for example if the mean or the variance of a region are chosen as descriptors. We show that the dependence of the descriptors with the region induces additional terms in the evolution equation of the active contour that have never been previously computed. DREAMS gives an easy way to take such a dependence into account and to compute the resulting additional terms. Experimental results point out the importance of the additional terms to reach a true minimum of the criterion and so to obtain accurate results. The covariance matrix determinant appears to be a very relevant tool for homogeneous color regions segmentation. As an example, it has been successfully applied to face detection in real video sequences.

In video sequence processing, shadow remains a major source of error for object segmentation. Traditional methods of shadow removal are mainly based on colour difference thresholding between the background and current images. The application of colour filters on MPEG or MJPEG images, however, is often erroneous as the chrominance information is significantly reduced due to compression. In addition, as the colour attributes of shadows and objects arc often very similar, discrete thresholding cannot always provide reliable results. This paper presents a novel approach for adaptive shadow removal by incorporating four different filters in a neuro-fuzzy framework. The neuro-fuzzy classifier has the ability of real-time self-adaptation and training, and its performance has been quantitatively assessed with both indoor and outdoor video sequences.

Natural scenes consist of a wide variety of stochastic patterns. While many patterns are represented well by statistical models in two dimensional regions as most image segmentation work assume, some other patterns are fundamentally one dimensional and thus cause major problems in segmentation. We call the former and the latter . In this paper, we propose a stochastic algorithm for parsing an image into a number of region and curve processes. The paper makes the following contributions to the literature. Firstly, it presents a generative model for curve processes in the form of Hidden Markov Model (HMM). The hidden layer is a Markov chain with each element being an image base selected from an over-complete basis, such as Difference of Gaussians (DOG) or Difference of Offset Gaussians (DOOG) at various scales and orientations. The rope model accounts for the geometric smoothness and photometric coherence of the curve processes. Secondly, it integrates both 2D region models, such as textures, splines etc with 1D curve models under the Bayes framework. Because both region and curve models are generative, they compete to explain input images in a layered representation. Thirdly, it achieves global optimization by effective Markov chain Monte Carlo methods in the sense of maximizing a posterior probability. The Markov chain consists of reversible jumps and diffusions driven by bottom up information. The algorithm is applied to real images with satisfactory results. We verify the results through random synthesis and compare them against segmentations with region processes only.

Image segmentation has been, and still is, a relevant research area in Computer Vision, and hundreds of segmentation algorithms have been proposed in the last 30 years. However, it is well known that elemental segmentation techniques based on boundary or region information often fail to produce accurate segmentation results. Hence, in the last few years, there has been a tendency towards algorithms which take advantage of the complementary nature of such information. This paper reviews different segmentation proposals which integrate edge and region information and highlights 7 different strategies and methods to fuse such information. In contrast with other surveys which only describe and compare qualitatively different approaches, this survey deals with a real quantitative comparison. In this sense, key methods have been programmed and their accuracy analyzed and compared using synthetic and real images. A discussion justified with experimental results is given and the code is available on Internet.

We present a novel approach for motion grouping from two frames, that recovers the dense velocity field, motion boundaries and regions, based on a 4-D Tensor Voting computational framework. Given two sparse sets of point tokens, we encode the image position and potential velocity for each token into a 4-D tensor. The voting process then enforces the motion smoothness while preserving motion discontinuities, thus selecting the correct velocity for each input point, as the most salient token. By performing an additional dense voting step we infer velocities at every pixel location, motion boundaries and regions. Using a 4-D space for this Tensor Voting approach is essential, since it allows for a spatial separation of the points according to both their velocities and image coordinates. Unlike other methods that optimize a specific objective function, our approach does not involve initialization or search in a parametric space, and therefore does not suffer from local optima or poor convergence problems. We demonstrate our method with synthetic and real images, by analyzing several difficult cases — opaque and transparent motion, rigid and non-rigid motion, curves and surfaces in motion.

Deformable models like snakes are a classical tool for image segmentation. Highly deformable models extend them with the ability to handle dynamic topological changes, and therefore to extract arbitrary complex shapes. However, the resolution of these models largely depends on the resolution of the image. As a consequence, their time and memory complexity increases at least as fast as the size of input data. In this paper we extend an existing highly deformable model, so that it is able to locally adapt its resolution with respect to its position. With this property, a significant precision is achieved in the interesting parts of the image, while a coarse resolution is maintained elsewhere. The general idea is to replace the Euclidean metric of the image space by a deformed non-Euclidean metric, which geometrically expands areas of interest. With this approach, we obtain a new model that follows the robust framework of classical deformable models, while offering a significant independence from both the size of input data and the geometric complexity of image components.

A novel deformable template is presented which detects and localizes shapes in grayscale images. The template is formulated as a Bayesian graphical model of a two-dimensional shape contour, and it is matched to the image using a variant of the belief propagation (BP) algorithm used for inference on graphical models. The algorithm can localize a target shape contour in a cluttered image and can accommodate arbitrary global translation and rotation of the target as well as significant shape deformations, without requiring the template to be initialized in any special way (e.g. near the target).

The use of BP removes a serious restriction imposed in related earlier work, in which the matching is performed by dynamic programming and thus requires the graphical model to be tree-shaped (i.e. without loops). Although BP is not guaranteed to converge when applied to inference on non-tree-shaped graphs, we find empirically that it does converge even for deformable template models with one or more loops. To speed up the BP algorithm, we augment it by a pruning procedure and a novel technique, inspired by the 20 Questions (divide-and-conquer) search strategy, called ”focused message updating.” These modifications boost the speed of convergence by over an order of magnitude, resulting in an algorithm that detects and localizes shapes in grayscale images in as little as several seconds on an 850 MHz AMD processor.