Catálogo de publicaciones - libros

We present a method for inferring and enforcing geometrical constraints on an approximate 3D model for building reconstruction applications. Compared to previous works, this approach requires no user intervention for constraints definition, is inherently compliant with the detected constraints and handles over-constrained configurations. An iterative minimization is applied to search for the model that minimizes the distance to the initial approximate model. Results on real data show the gain obtained by this algorithm.

When building shape models, it is first necessary to filter out the similarity transformations from the original configurations. This is normally done using Procrustes analysis, that is minimising the sum of squared distances between the corresponding landmarks under similarity transformations. In this article we propose to align shapes using the Minimum Description Length (MDL) criterion. Previously MDL has been used to locate correspondences. We show that the Procrustes alignment with respect to rotation is not optimal.

The MDL based algorithm is compared with Procrustes on a number of data sets. It is concluded that there is improvement in generalisation when using Minimum Description Length. With a synthetic example it is shown that the Procrustes alignment can fail significantly where the proposed method does not.

The Description Length is minimised using Gauss-Newton. In order to do this the derivative of the description length with respect to rotation is derived.

In this paper a fully automated segmentation system for the femur in the knee in Magnetic Resonance Images and the brain in Single Photon Emission Computed Tomography images is presented. To do this several data sets were first segmented manually. The resulting structures were represented by unorganised point clouds. With level set methods surfaces were fitted to these point clouds. The iterated closest point algorithm was then applied to establish correspondences between the different surfaces. Both surfaces and correspondences were used to build a three dimensional statistical shape model. The resulting model is then used to automatically segment structures in subsequent data sets through three dimensional Active Shape Models. The result of the segmentation is promising, but the quality of the segmentation is dependent on the initial guess.

We present an efficient algorithm for fitting three dimensional (3-D) Active Appearance Models (AAMs). We do so, by introducing a 3-D extension of a recently proposed method that is based on the inverse compositional image alignment algorithm. We demonstrate its applicability for the segmentation of the left ventricle in short axis cardiac MRI. We perform experiments to evaluate the speed and segmentation accuracy of our algorithm on a total of 1473 cardiac MR images acquired from 11 patients. The fitting is around 60 times faster than standard Gauss-Newton optimization, with a segmentation accuracy that is as good as, and often better than Gauss-Newton.

Recently new nuclear medical diagnostic imaging methods have been introduced by which it is possible to diagnose Parkinson’s disease, PD, at an early stage. The binding of a specific receptor ligand [123I]-FP-CIT in specific structures of the human brain is visualized by way of single photon emission computerized tomography, SPECT. The interpretation of the SPECT data can be accessed by visual analysis and manual quantification methods. We have developed a computer aided automatic decision support system in attempt to facilitate the diagnosis of Parkinson’s disease from the acquired SPECT images. A rigid model was used for the segmentation of the basal ganglia of the human brain. The aim of the study was to develop an automated method for quantification and classification of the images.

The study comprises SPECT scans 89 patients, who underwent a [123I]-FP-CIT examination because of suspected Parkinson’s disease. An experienced nuclear medicine physician interpreted the images and diagnosed 65 of the patients as most likely suffering from Parkinson’s disease.

The method included the following steps; (i) segmentation of basal ganglia by fitting a rigid 3D model to the image, (ii) extraction of 17 features based on image intensity distribution in the basal ganglia and a reference based on image intensity distribution outside the basal ganglia, (iii) classification using Support Vector Machine (SVM).

The classification based on the automated method showed a true acceptance of 96.9% and a true rejection of 91.6%. The classification based on a manual quantification method gave a true acceptance of 98.5% and a true rejection of 100%. The method proposed here is fully automatic and it makes use of the full 3D data set in contrast to a method that is widely used at hospitals today which only uses a few 2D image slices.

We present a new method for producing locally non intersecting polygon meshes of naturally branching structures. The generated polygon mesh follows the object’s underlying structure as close as possible, while still producing polygon meshes that can be visualized efficiently on commonly available graphic acceleration hardware. A priori knowledge of vascular branching systems is used to derive the polygon mesh generation method. Visualization of the internal liver vessel structures and naturally looking tree stems generated by Lindenmayer-systems is used as examples. The method produce visually convincing polygon meshes that might be used in clinical applications in the future.

In screening X-ray mammography two different views are captured and analysed of both breasts. In the four X-ray images some special signs are looked for. First the individual images are analysed independently, but good result can be achieved only if joint analysis of the images is also done. This paper proposes a simple procedure for the joint analysis of the breast’s two views. The procedure is based upon the experiences of radiologists: masses and calcifications should emerge on both views, so if no matching is found, the given object is a false positive hit. First a reference system is evolved to find corresponding regions on the two images. Calcification clusters obtained in individual images are matched in “2.5 D” provided by the reference system. Masses detected in individual images are further examined with texture segmentation. The proposed approach can significantly reduce the number of false positive hits both in calcification and in mass detection.

For databases of facial images, where each subject is depicted in only one or a few images, the query precision of interactive retrieval suffers from the problem of extremely small class sizes. A potential way to address this problem is to employ automatic even though imperfect classification on the images according to some high level concepts. In this paper we point out that significant improvement in terms of the occurrence of the first subject hit is feasible only when the classifiers are of sufficient accuracy. In this work Support Vector Machines (SVMs) are incorporated in order to obtain high accuracy for classifying the imbalanced data. We also propose an automatic method to choose the penalty factor of training error and the width parameter of the radial basis function used in training the SVM classifiers. More significant improvement in the speed of retrieval is feasible with small classes than with larger ones. The results of our experiments suggest that the first subject hit can be obtained two to five times faster for semantic classes such as “black persons” or “eyeglass-wearing persons”.

We propose the use of eye-movements as a biometric. A case study investigating potentials of eye-movement data for biometric purposes was conducted. Twelve participants’ eye-movements were measured during still and moving objects viewing. The measured data includes pupil sizes and their dynamics, gaze velocities and distances of infrared reflections of the eyes. For still object viewing of 1 second duration, identification rate of 60 % can be obtained by using dynamics of pupil diameters. We suggest an integration of the eye-movement-based identification into general video-based biometric systems.

This paper investigates the problem of object surface reflectance modeling, which is sometimes referred to as , for photorealistic rendering and effective multimedia applications. A number of methods have been developed for estimating object surface reflectance properties in order to render real objects under arbitrary illumination conditions. However, it is still difficult to densely estimate surface reflectance properties faithfully for complex objects with interreflections. This paper describes a new method for densely estimating the non-uniform surface reflectance properties of real objects constructed of convex and concave surfaces. Specifically, we use registered range and surface color texture images obtained by a laser rangefinder. Then, we determine the positions of light sources in order to capture color images to be used in discriminating diffuse and specular reflection components of surface reflection. The proposed method can densely estimate the reflectance parameters of objects with diffuse and specular interreflections based on an inverse global illumination rendering. Experiments are conducted in order to demonstrate the usefulness and the advantage of the proposed methods through comparative study.