Catálogo de publicaciones - libros

In this paper we present a novel single-frame image zooming technique based on so-called “self-examples”. Our method combines the ideas of fractal-based image zooming, example-based zooming, and nonlocal-means image denoising in a consistent and improved framework. In Bayesian terms, this example-based zooming technique targets the MMSE estimate by learning the posterior directly from examples taken from the image itself at a different scale, similar to fractal-based techniques. The examples are weighted according to a scheme introduced by Buades to perform nonlocal-means image denoising. Finally, various computational issues are addressed and some results of this image zooming method applied to natural images are presented.

Unlike conventional raster-scan based connected-component labeling algorithms which detect the connectivity of object pixels by processing pixels in an image one by one, this paper presents an efficient run-based two-scan labeling algorithm: the run data obtained during the scan are recorded in a queue, and are used for detecting the connectivity later. Moreover, unlike conventional label-equivalence-based algorithms which resolve label equivalences between provisional labels that are assigned during the first scan, our algorithm resolve label equivalences between the representative labels of equivalent provisional label sets. In our algorithm, at any time, all provisional labels that are assigned to a connected component are combined in a set, and the smallest label is used as the representative label. The corresponding relation of a provisional label to its representative label is recorded in a table. Whenever different connected components are found to be connected, all provisional label sets concerned with these connected components are merged together, and the smallest provisional label is taken as the representative label. When the first scan is finished, all provisional labels that were assigned to each connected component in the given image will have a unique representative label. During the second scan, we need only to replace each provisional label with its representative label. Experimental results on various types of images demonstrate that our algorithm is the fastest of all conventional labeling algorithms.

In this paper we pursue the goal of constructing a biologically meaningful coordinate system that (i) carries information about the spatial distribution of gene activity regions and (ii) captures its effect on the internal structure and shape of an organism during its growth. Geometrically, this “Darcyan” coordinate system is curvilinear and comprised of an interior grid surrounded by a closed curve, namely the boundary of the organism. We explore two computational methods of constructing it, one based on potential theory (Poisson equation) and the other based on level set methods, with particular emphasis placed on the latter. We propose a novel algorithm that uses image processing tools for the extraction of the boundary, from which is produced the interior Darcyan coordinate system by means of level set evolution. Examples show the ability of the proposed algorithm to handle the complex geometry of the initial boundary such as significant oscillations, corners and cusps.

The Hilbert Transform (HT) and the analytic signal (AS) are widely used in their one-dimensional version for various applications. However, in the bi-dimensional (2D) case as occur for images, the definition of the 2D-HT is not unique and several approaches to it have been developed, having as one of the main goals to obtain a meaningful 2D-AS or analytic image, which can be used for various practical applications. In this work, one particular approach to the 2D-HT is introduced that allowed the calculation of analytic images which satisfy the basic properties that these functions have in the 1D case, and that produces a 2D spectrum equal to zero in one quadrant. The methods for calculation of the discrete version of the 2D-HT and the associated AS are presented and analyzed, as well as two applications, for edge detection and for envelope detection in a 2D AM modulated radial chirp.

This paper presents an approach based on graph cuts initially used for motion segmentation that is being applied to the non-rigid registration problem. The main contribution of our method is the formulation of landmarks in the graph cut minimization framework. In the graph cut method, we add a penalty cost based on landmarks to the data energy. In the presence of a landmark, we adjust the weights to cut strategic links. Our formulation also allows the spread of a landmark influence to its neighborhood. We first show with synthetic images that minimization with graph cuts can indeed be used for non-rigid registration and show how landmarks can guide the minimization process towards a customized solution. We later use this method with real images and show how landmarks can successfully guide the registration of a coronary angiogram.

A new method for determination of the best hypothesis in Hough transform (HT)-based methods to detect ellipses is described. The method uses a new distinctiveness measurement to rank the hypotheses and determine a suitable candidate. The method is aimed at improving HT robustness to image noise and quantization effects. Experiments on images that demonstrate the method’s applicability are also presented.

This paper presents a closed edge detection method based on a level lines selection approach. The proposed method is based on an unsupervised probabilistic scheme using an method. A level line is considered meaningful if its contrast and length is unlikely to be due to chance. Besides being unsupervised, this method exploits a tree structure. The first step of the proposed approach is to reduce the meaningful level lines set using this hierarchical structure. Compared with a previous method using the same principle, our method achieve a 67% reduction rate of irrelevant levels lines. The second step of the proposed approach illustrates the high flexibility of using closed edge boundaries such as levels lines. Using a rather simple curvature analysis, the proposed method detects anatomical structures boundaries from CT scan images.

Scientific image processing involves a variety of problems including image modeling, reconstruction, and synthesis. In this paper we develop a constrained sampling approach for porous media synthesis and reconstruction in order to generate artificial samples of porous media. Our approach is different from current porous media reconstruction methods in which the Gibbs probability distribution is maximized by simulated annealing. We show that the artificial images generated by those methods do not contain the variability that typical samples of random fields are required to have.

This paper proposes a high-accuracy rotation estimation algorithm using 1D Phase-Only Correlation (POC). In general, the rotation angle between two images is estimated as follows: (i) convert the image rotation into the image shift by polar mappings of the amplitude spectra of images, and (ii) estimate the translational displacement between the polar mappings to obtain the rotation angle. The problem of rotation estimation between two images is replaced to 1D displacement estimation between pairs of horizontal lines at the same vertical position in two polar mappings. The proposed algorithm employs 1D POC instead of 2D matching for estimating a rotation angle. The use of 1D POC to estimate the rotation angle makes it possible to reduce the computational cost significantly without sacrificing the estimation accuracy.

A method is proposed to enhance contour primitives of multi-part objects in complex images. It consists in first extracting circular arcs and straight-line segment primitives from the image edge map. The produced binary constant-curvature primitive map includes primitives of interest on the object contour or silhouette, as well as two types of distractors: internal texture segments and external background segments. Each obtained primitive is enhanced using an exhaustive evaluation of a number of pairwise grouping criteria. Finally, an iterative relaxation procedure adjusts the weight of each primitive according to the weights of its best-matching primitives. A subjective ground-truth binary map may be used to assess the degree to which the final weighted map corresponds to a selective enhancement of contour primitives.