Catálogo de publicaciones - libros

An industrial rock classification system is constructed and studied. The local texture information in many image patches is extracted and classified. The decisions made at the local level are fused to form the high-level decision on the image/rock as a whole. The main difficulties of this application lay in significant variability and inhomogeneity of local textures caused by uneven rock surfaces and intrusions. Therefore, an emphasis is paid to the derivation of informative representation of local texture and to robust classification algorithms. The study focuses on the co-occurrence representation of texture comparing the two frequently used strategies, namely the approach based on Haralick features and methods utilizing directly the co-occurrence likelihoods. Apart of maximum-likelihood (ML) classifiers also an alternative method is studied considering the likelihoods to prototypes as feature of a new space. Unlike the ML methods, a classifier built in this space may leverage all training examples. It is experimentally illustrated, that in the rock classification setup the methods directly using the co-occurrence estimates outperform the feature-based techniques.

Co-occurrence matrices are proved to be useful tool for the purpose of texture recognition. However, they are sensitive to the change of the illumination conditions. There are standard preprocessing approaches to this problem. However, they are lacking certain qualities. We studied the tangent kernel SVM approach as an alternative way of building illumination-robust texture classifier. Testing on the standard texture data has shown promising results.

This paper presents a new method for speckle noise reduction in medical ultrasound images. It is based on the statistical description of the envelope of ultrasound signal by the virtue of the Nakagami-m distribution. Parameter of this distribution is used to adjust an adaptive filter.

The theory of neighborhood sequences is applicable in many image-processing algorithms. The theory is well developed for the square grid. Recently there are some results for the hexagonal grid as well. In this paper, we are considering all the three regular grids in the plane. We show that there are some very essential differences occurring. On the triangular plane the distance has metric properties. The distances on the square and the hexagonal case may not meet the triangular inequality. There are non-symmetric distances on the hexagonal case. In addition, contrary to the other two grids,the distance can depend on the order of the initial elements of the neighborhood sequence.Moreover in the hexagonal grid it is possible that circles with different radii are the same (using different neighborhood sequences). On the square grid the circles with the same radius are in a well ordered set, but in the hexagonal case there can be non-comparable circles.

A multitemporal fast adaptive recursive restoration method based on the underlying spatial probabilistic image model is presented. The method assumes linear degradation model with the unknown possibly non-homogeneous point-spread function and additive noise. Pixels in the vicinity of image steep discontinuities are left unrestored to minimize restoration blurring effect. The method is applied for astronomical sunspot image restoration, where for every ideal undegraded unobservable image several degraded observed images are available.

In limited angle tomography, the projection views over a complete angular range of 180° are not available for image reconstruction. The missing part of the projection or sinogram data need to be extrapolated numerically, if standard image reconstruction methods are applied. A novel stackgram domain can be regarded as an intermediate form of the sinogram and image domains, in which the signals along the sinusoidal trajectories of a sinogram can be processed independently. In this paper, we compare extrapolation of incomplete sinogram data in the sinogram and stackgram domains along the angular directions. The extrapolated signals are assumed to be band–limited, other assumptions about the data are not made. In this study, we employed simulated numerical data with different ranges of the limited projection views. According to our experiments, extrapolation of the incomplete data in the stackgram domain provides quantitatively better results as compared to extrapolation in the sinogram domain. In addition, tangential degradation in the reconstructed images can not be observed in the case of stackgram extrapolation, in contrast to angular sinogram extrapolation.

A classification of centres of maximal balls (CMBs) in ℤ derived from generalizations of the chessboard and city block metrics to 3D, a weighted metric, and the Euclidean metric is presented. Using these metrics, the set of CMBs (the medial axis) can be extracted. One difficulty with skeletonization in 3D is that of guaranteeing reversibility. A reversible skeleton generally consists of both surfaces and curves. Previous attempts to construct connected skeletons including the CMBs uses conditions based on local neighbourhood configurations. However, a local neighbourhood might be too small and, most important, does not allow a consistent definition for surface- and curve-parts of the skeleton. The classification of the CMBs presented in this paper will be a tool for defining which parts of a 3D skeleton are surfaces and curves.

Algorithms for volume evaluation of 3D objects on freehand ultrasound images are considered. The position sensor used provides image spatial position and orientation data. The algorithms are based on Watanabe formula for volume computation and use cubic spline interpolation. They allow object volume evaluation on initial image sequence without reconstruction of 3D cube avoiding inevitable data loss at this pre-processing stage. The algorithm accuracy was tested on simulated and real objects.

Tuning a complete image processing chain (IPC) remains a tricky step. Until now researchers focused on the evaluation of single algorithms, based on a small number of test images and ad hoc tuning independent of input data. In this paper we explain how, by combining statistical modeling with design of experiments, numerical optimization and neural learning, it is possible to elaborate a powerful and adaptive IPC. To succeed, it is necessary to build a large image database, to describe input images and finally to evaluate the IPC output. By testing this approach on an IPC dedicated to road obstacle detection, we demonstrate that this experimental methodology and software architecture ensure a steady efficiency. The reason is simple: the IPC is globally optimized, from a large number of real images (180 out of a sequence of 30 000) and with adaptive processing of input data.

Filter networks, i.e. decomposition of a filter set into a layered structure of sparse subfilters has been proven successful for e.g. efficient convolution using finite impulse response filters. The efficiency is due to the significantly reduced number of multiplications and additions per data sample that is required. The computational gain is dependent on the choice of network structure and the graph representation compactly incorporates the network structure in the design objectives. Consequently the graph representation forms a framework for searching the optimal network structure. It also removes the requirement of a layered structure, at the cost of a less compact representation.