Abstracts

Skeleton is a region-based shape feature that is extracted from binary image data. A very illustrative definition of the skeleton is given using the prairie-fire analogy: the object boundary is set on fire and the skeleton is formed by the loci where the fire fronts meet and quench each other.

The lecture is to give a brief description of the three general skeletonozation methods including distance-based, Voronoi, and thinning approaches.

Some applications (mostly in 3D) are also presented.

Computerized tomography was originally a method of diagnostic radiology that was used to obtain the density distribution within the human body based on x-ray projection samples. From a mathematical point of view it seeks to determine (perhaps only approximately) an unknown function f(x, y, z) defined over the three-dimensional (3D) Euclidean space E3 from weighted integrals over subspaces of E3 (called projections). Since the values of f can vary over a wide range, a huge number of projections are needed to ensure an accurate reconstruction of it.

There are other situations where we want to reconstruct an object but it is not possible to examine the object itself as only its projections are available. For example, in industrial applications like non-destructive testing or reverse engineering we normally cannot investigate the interior of an industrial object itself. However, its x-ray projections can be measured so one can make assumptions about the shape of the object and the material(s) it is made of. Other examples arise from the field of electron microscopy where the task is to identify biological macromolecules composed of ice, protein, and RNA or to determine crystalline structures. In most of these applications there is a restriction that only a small number of projections can be taken, thus methods of computerized tomography usually cannot approximate the function f that well. However, there is a hope that f can be determined from just a small number of projections too, since we have some prior knowledge here: the range of f is discrete and consists of only a small number of possible values. This leads us to the field of discrete tomography where it is assumed that only a handful of projections are available, but the prior information about the possible values of f can be used.

A more restricted situation is when f can only take the values 0 or 1. This kind of tomography is called binary tomography, which also has a variety of applications. For example, in electron microscopy 0 and 1 can represent the absence and presence of a certain atom in the crystalline structure, respectively. Similarly, in angiography the values 0 and 1 can describe the absence or presence of a contrast agent in heart chambers or in segments of blood vessels.

Although tomographical problems can also be investigated in higher dimensions, in fact almost all the 3D applications use two-dimensional (2D) slices of the object being studied, so reducing a 3D tomography problem to a 2D one. After the reconstruction the 2D slices are integrated together to produce a 3D object. In this talk we give a brief overview of the basic definitions and problems of 2D binary tomography.

No abstract available.

No abstract available.

No abstract available.

In this lecture we present the components of a thermal video processing application designated to provide support for rescue scenarios. The videos are recorded with state-of-the-art portable thermal cameras, and video processing tasks are considered to extract useful information from them. As we focus on fire scenarios, successful hotspot- and human detection/tracking are desired goals. The formal video content description is provided in terms of an XML Schema, which provides the possibility to populate metadata to the knowledge base (ontology) of the system.

There is an increasing number of applications that require accurate aligning of one image with another taken from different viewpoints, by different imaging devices, or at different times.

Automatic image registration methods are the easiest to use. Besides this, registration is often a "building block" of a bigger project, where the more automatic these "blocks" are the better.

We can distinguish between two approaches. In the first case, geometrical information contents (e.g., corner points, contours, surfaces) are extracted from the images and these are used for registration. In the second case, the image intensity values are used directly utilizing so called voxel similarity measures.

After a brief introduction to image registration, these approaches will be discussed.

In current clinical practice X-ray fluoroscopy is widely used for visualization of blood vessels, bones and instruments during minimally invasive procedures. Although fluoroscopy provides high-resolution images during the intervention in real time, these images contains only a 2D projection, have very low soft-tissue contrast and the patient is impacted by high radiation exposure. Combining these images with 3D (CT, MR, etc.) images results in a more complete representation of the patient anatomy. This makes it easier and safer to perform complex procedures, in shorter time, with better outcomes.

In the first part of the presentation an overview of current interventional X-ray systems will be given, including details about the most important medical procedures and the applied image processing techniques. In the second part image registration methods for 2D fluoroscopy images and various 3D images will be discussed.

No abstract available.

The advantages of representing objects in images as fuzzy spatial sets are numerous and have lead to increased interest for fuzzy approaches in image analysis. Fuzziness is an intrinsic property of images. It is additionally introduced in digital image processing by discretization, and as a natural outcome of most imaging devices. The fuzzy membership of a point reflects the level to which that point fulfils certain criteria to belong to the object; the membership of a pixel is often determined to be proportional to the part of the pixel area covered by the observed object. Preservation of fuzziness implies preservation of important information about objects and images. Therefore, fuzziness should, in general, be kept and utilized as long as possible when analyzing the image data.

The intention is to, first, give a brief introduction to the fuzzy set theory, and then, to discuss some of its applications to image analysis. Fuzzy segmentation methods and fuzzy shape analysis techniques are in focus. Several ways to incorporate fuzzy methods in segmentation process, and also ways to represent objects in images as fuzzy sets, are discussed. Methods to generalize well-known concepts, like perimeter and area of a set, or distance between elements of a set, from crisp to fuzzy sets, are explained. It is shown that, by using fuzzy approaches, an improved precision of some shape description can be achieved. In addition, some defuzzification approaches, that are used to reduce fuzzy sets to their crisp representatives, are presented, and their properties discussed.

The surgical repair of fractured bones is often a difficult task, and the fixation of these bones has to be planned very carefully. This is why trauma surgeons may use a Computer Aided Surgery (CAS) system to improve surgical accuracy. With the introduction of mechanical analysis, a complete biomechanical laboratory is created and the surgeon is able to predict the stability of the surgical plan before the intervention.

The lecture gives an overview of a CAS system. Basic steps of the procedure are covered, including segmentation, surface extraction, geometrical model generation, surgical planning, and finite element mesh generation.

No abstract available.

Structure and objects in video images are often not known. If we search for camera registration or focused areas of these situations, some a priori knowledge was needed.

Now I present two methods for finding structeres in video images without any preliminary object definition: Co-motion statistics and estimation of relative focus map.

First, a new motion-based method is presented for automatic registration of images in multi-camera systems, to permit synthesis of wide-baseline composite views. Unlike existing static-image and motion based methods, our approach does not need any a priori information about the scene, the appearance of objects in the scene, or their motion. We introduce an entropy-based preselection of motion histories and an iterative Bayesian assignment of corresponding image areas. Finally, correlated point-histories and data-set optimization lead to the matching of the different views.

Another application of co-motion methods is finding the vanishing-point position for planar reflected images or shadows, or the horizontal vanishing line, making use of motion statistics derived from a video sequence. I also present a new automatic solution for finding focused areas based on localized blind deconvolution. This approach makes possible to determine the relative blur of image areas without a priori knowledge about the images or the shooting conditions. The method uses deconvolution based feature extraction and a new residual error based classification for the discrimination of image areas. We also show the method´s usability for applications in content based indexing and video feature extraction.

In many branches, as biomedicine, astronomy, robotic and so on, images are a fundamental tool of investigation. However, sometimes the bad quality of the available images does not permit to use them immediately. Thus, it is necessary to eliminate the noise and effects of the blur from the observed images. The problem of restoring images deals with estimating the original images, by starting from the observed image and the characteristic of the blur and the noise. The image restoration problem is ill-posed in the sense of Hadamard, that is in some case, the solution nether exists, nor is unique, nor can be stable in presence of noise. Thus, regularization techniques are useful in order to transform this problem in a well-posed one. By these techniques the solution is defined as the minimum of a suitable energy function. The discontinuities of the ideal image are unknown and need to be estimated. In order to minimize the resulting non-convex energy function either deterministic or probabilistic techniques can be used.

Regions in real images are often homogeneous, neighboring pixels usually have similar properties (intensity, color, texture, ...). Markov Random Fields (MRF) are often used to capture such contextual constraints in a probabilistic framework. MRFs are well studied with a strong theoretical background hence providing a tool for rigorous and concise image modeling. Furthermore, they allow Markov Chain Monte Carlo (MCMC) sampling of the (hidden) underlying structure which greatly simplifies inference and parameter estimation. In this talk, we will give a short yet complete introduction to MRF image modelization by explaining how to construct a minimalistic model, how to estimate model parameters, and then how to infer the most likely segmentation of an image.

The face detection and facial gesture recognition has become a very popular area of research in digital image processing. This speak overviews the theoretical background of this area. Afterwards we show some applications in a multi-modal human-computer interaction system developed in our Lab.

No abstract available.

The Hough transform is a technique to isolate features of a particular shape within an image. Since it requires that the desired features be specified in some parametric form, the classical Hough transform is most commonly used for the detection of regular curves such as lines, circles, ellipses, etc. The Hough transform finds many industrial applications, as most manufactured parts contain feature boundaries which can be described by regular curves. However, when it comes to natural objects, lines are rarely straight and circles seldom perfect circles. Combining the Hough transform with a distance transform enables easy and robust detection of non-perfect shapes.

The intention with the lecture is to give a brief description of the Hough transform for detecting lines and circles in images. How a distance transform can be combined with a weighted Hough transform for a more tolerant shape detection, will also be presented. The technique will be illustrated on a project to segment and measure seedlings, to facilitate quantitative evaluation of seed vitality.