Catálogo de publicaciones - libros

In this work a novel texture model particularly suited for unsupervised image segmentation is proposed. Any texture is represented at region level by means of a finite-state hierarchical model resulting from the superposition of several Markov chains, each associated with a different spatial direction. Corresponding to such a modeling, an optimization scheme, referred to as Texture Fragmentation and Reconstruction (TFR) algorithm, has been introduced.

The TFR addresses the model estimation problem in two sequential layers: the former “fragmentation” step allows to find the terminal states of the model, while the latter “reconstruction” step is aimed at estimating the relationships among the states which provide the optimal hierarchical structure to associate with the model. The latter step is based on a probabilistic measure, i.e, the region gain, which accounts for both the region scale and the inter-region interaction.

The proposed segmentation algorithm was tested on a segmentation benchmark and applied to high resolution remote-sensing forest images as well.

ECOC is a diffused and successful technique to implement a multiclass classification system by decomposing the original problem in several two-class problems. In this paper we propose ECOC systems with a reject option carried out through two different schemes. The first one estimates the reliability of the output of the ECOC system and does not require any change in its structure. The second scheme, instead, estimates the reliability of the internal dichotomizers and implies a slight modification in the decoding stage. A final investigation is done on the sequential combination of both methods.

Since texture is scale dependent, multi-scale techniques are quite useful for texture classification. Scale-space theory introduces multi-scale differential operators. In this paper, the N-jet of derivatives up to the second order at different scales is calculated for the textures in Brodatz album to generate the textures in multiple scales. After some preprocessing and feature extraction using principal component analysis (PCA), instead of combining features obtained from different scales/derivatives to construct a combined feature space, the features are fed into a for classification. The learning curves are used to evaluate the performance of the proposed texture classification system. The results show that this new approach can significantly improve the performance of the classification especially for small training set size. Further, comparison between combined feature space and combined classifiers shows the superiority of the latter in terms of performance and computation complexity.

The computation of non-negative tensor factorization may become very time-consuming when large datasets are used. This study shows how to accelerate NTF using multiresolution approach. The large dataset is preprocessed with an integer wavelet transform and NTF results from the low resolution dataset are utilized in the higher resolution dataset. The experiments show that the multiresolution based speed-up for NTF computation varies in general from 2 to 10 depending on the dataset size and on the number of required basis functions.

One of the basic problems in discrete tomography is the reconstruction of discrete sets from few projections. Assuming that the set to be reconstructed fulfils some geometrical properties is a commonly used technique to reduce the number of possibly many different solutions of the same reconstruction problem. Since the reconstruction from two projections in the class of so-called -convex sets is NP-hard this class is suitable to test the efficiency of newly developed reconstruction algorithms. However, until now no method was known to generate sets of this class from uniform random distribution and thus only ad hoc comparison of several reconstruction techniques was possible. In this paper we first describe a method to generate some special -convex discrete sets from uniform random distribution. Moreover, we show that the developed generation technique can easily be adapted to other classes of discrete sets, even for the whole class of -convexes. Several statistics are also presented which are of great importance in the analysis of algorithms for reconstructing -convex sets.

The study of natural image statistics considers the statistical properties of large collections of images from natural scenes, and has applications in image processing, computer vision, and visual computational neuroscience. In the past, a major focus in the field of natural image statistics have been the statistics of outputs of linear filters. Recently, attention has been turning to nonlinear models. The contribution of this paper is the empirical analysis of the statistical properties of a central nonlinear property of natural scenes: the local log-contrast. To this end, we have studied both second-order and higher-order statistics of local log-contrast. Second-order statistics can be observed from the average amplitude spectrum. To examine higher-order statistics, we applied a higher-order-statistics-based model called independent component analysis to images of local log-contrast. Our results on second-order statistics show that the local log-contrast has a power-law-like average amplitude spectrum, similarly as the original luminance data. As for the higher-order statistics, we show that they can be utilized to learn intuitively meaningful spatial local-contrast patterns, such as contrast edges and bars. In addition to shedding light on the fundamental statistical properties of natural images, our results have important consequences for the analysis and design of multilayer statistical models of natural image data. In particular, our results show that in the case of local log-contrast, oriented and localized second-layer linear operators can be learned from the higher-order statistics of the nonlinearly mapped output of the first layer.

One of the most difficult steps of intramedullary nailing of femoral shaft fractures is distal locking - the insertion of distal transverse interlocking screws, for which it is necessary to know the position and orientation of the distal locking holes of the intramedullary nail. This paper presents a novel parameter decomposition approach for solving this problem using single calibrated X-ray image. The problem is formulated as a model-based optimal fitting process, where the to-be-optimized parameters are decomposed into two sets: (a) the angle between the nail axis and its projection on the imaging plane, and (b) the translation and rotation of the geometrical models of the distal locking holes around the nail axis. By using a hybrid optimization technique coupling an evolutionary strategy and a local search algorithm to find the optimal values of the latter set of parameters for any given value of the former one, we reduce the multiple-dimensional model-based optimal fitting problem to a one-dimensional search along a finite interval. We report the  −  experimental results, which demonstrate that the accuracy of our approach is adequate for successful distal locking of intramedullary nails.

The estimation of a patch position in an image is a long established but still relevant topic with many applications, e.g. pose estimation and tracking in image sequences. In most systems the position estimate needs to be fused with other estimates, and hence, covariance information is required to weight the different estimates in the right way. In this paper we address the issue with covariance estimation in the case of sum of absolute difference (SAD) block matching. First, we derive the theory for covariance estimation in the case of SAD matching. Second, we evaluate the suggested method in a virtual 3D patch tracking scenario in order to verify the performance in real-world scenarios.

The paper presents an approach to multimodal image registration. The method is developed for aligning infrared (IR) and visual (RGB) images of facades. It is based on mapping clouds of points extracted by a corner detector applied to both images. The experiments show that corners are suitable features for our application. In the alignment process a number of transformation hypotheses is generated and evaluated. The evaluation is performed by measuring similarity between the RGB corners and the transformed corners from IR image. Directed partial Hausdorff distance is used as a robust similarity measure. The implemented system has been tested on various IR-RGB pairs of images of buildings. The results show that the method can be used for image registration, but also expose some typical problems.

This paper proposes a fast 3D reconstruction approach for efficiently generating watertight 3D models from multiple short baseline views. Our method is based on the combination of a GPU-based plane-sweep approach, to compute individual dense depth maps and a subsequent robust volumetric depth map integration technique. Basically, the dense depth map values are transformed to a volumetric grid, which are further embedded in a graph structure. The edge weights of the graph are derived from the dense depth map values and if available, from sparse 3D information. The final optimized surface is obtained as a min-cut/max-flow solution of the weighted graph. We demonstrate the robustness and accuracy of our proposed approach on several real world data sets.