Catálogo de publicaciones - libros

It is well known that linear filters are not powerful enough for many low-level image processing tasks. But it is also very difficult to design robust non-linear filters that respond exclusively to features of interest and that are at the same time equivariant with respect to translation and rotation. This paper proposes a new class of rotation-equivariant non-linear filters that is based on the principle of group integration. These filters become efficiently computable by an iterative scheme based on repeated differentiation of products and summations of the intermediate results. Our experiments show that the proposed filter detects pollen porates with only half as many errors than alternative approaches, when high localization accuracy is required.

In this paper, the properties of a novel color image filtering technique capable of impulse noise removal and edge enhancement are analyzed. The new filtering design is a generalization of the well known . The proposed filtering class is minimizing the cumulated dissimilarity measure of a group of pixels from the filtering window. The described filter is computationally efficient, easy to implement and very effective in suppressing impulsive noise, while preserving image details and strongly enhancing its edges.

Deblurring is important in many visual systems. This paper presents a novel approach for nonstationary blurred image reconstruction with ringing reduction in a variational Bayesian learning and regularization framework. Our approach makes effective use of the image statistical prior and image local spatial conditions through the whole learning scheme. A nature image statistics based marginal prior distribution is used not only for blur kernel estimation but also for image reconstruction. For an ill-posed blur estimation problem, variational Bayesian ensemble learning can achieve a tractable posterior using an image statistic prior which is translation and scale-invariant. During the deblurring, nonstationary blurry images have stronger ringing effects. We thus propose an iterative reweighted regularization function based on the use of an image statistical prior and image local spatial conditions for perceptual image deblurring.

We propose an unsupervised “local learning” algorithm for learning a metric in the input space. Geometrically, for a given query point, the algorithm finds the minimum volume ellipsoid (MVE) covering its neighborhood which characterizes the correlations and variances of its neighborhood variables. Algebraically, the algorithm maximizes the determinant of the local covariance matrix which amounts to a convex optimization problem. The final matrix parameterizes a Mahalanobis metric yielding the MVE metric (MVEM). The proposed metric was tested in a supervised learning task and showed promising and competitive results when compared with state of the art metrics in the literature.

A natural (human) eye can easily detect large visual patterns or objects emerging from spatially distributed discrete entities. This aspect of pattern analysis has been barely addressed in literature. We propose a biologically inspired approach derived from the concept of visual attention to associate together the distributed pieces of macro level patterns. In contrast to the usual approach practiced by the existing models of visual attention, this paper introduces a short-term excitation on the features and locations related to the current focus of attention in parallel to the spatial inhibition of return. This causes the attention system to fixate on analogous units in the scene that may formulate a meaningful global pattern. It is evident from the results of experiments that the outcome of this process can help in widening the scope of intelligent machine vision.

Glaucoma is one of the most common causes of blindness and it is becoming even more important considering the ageing society. Because healing of died retinal nerve fibers is not possible early detection and prevention is essential. Robust, automated mass-screening will help to extend the symptom-free life of affected patients. We devised a novel, automated, appearance based glaucoma classification system that does not depend on segmentation based measurements. Our purely data-driven approach is applicable in large-scale screening examinations. It applies a standard pattern recognition pipeline with a 2-stage classification step. Several types of image-based features were analyzed and are combined to capture glaucomatous structures. Certain disease independent variations such as illumination inhomogeneities, size differences, and vessel structures are eliminated in the preprocessing phase. The “vessel-free” images and intermediate results of the methods are novel representations of the data for the physicians that may provide new insight into and help to better understand glaucoma. Our system achieves 86 % success rate on a data set containing a mixture of 200 real images of healthy and glaucomatous eyes. The performance of the system is comparable to human medical experts in detecting glaucomatous retina fundus images.

Most face recognition systems are based on some form of batch learning. Online face recognition is not only more practical, it is also much more biologically plausible. Typical batch learners aim at minimizing both training error and (a measure of) hypothesis complexity. We show that the same minimization can be done incrementally as long as some form of ”scaffolding” is applied throughout the learning process. Scaffolding means: make the system learn from samples that are neither too easy nor too difficult at each step. We note that such learning behavior is also biologically plausible. Experiments using large sequences of facial images support the theoretical claims. The proposed method compares well with other, numerical calculus-based online learners.

Ever since the first fishermen ventured into the sea, tides have been the subject of intense human observation. As a result, computational models and ‘tide predicting machines’, mechanical computers for predicting tides have been developed over 100 years ago. In this work we propose a statistical model for short-term prediction of sea levels at high tide in the tide influenced part of the Weser at Vegesack. The predictions made are based on water level measurements taken at different locations downriver and in the German Bight. The system has been integrated tightly into the decision making process at the Bremen Dike Association on the Right Bank of the Weser.

In this paper we propose a generative statistical approach for the three dimensional (3D) extraction of the branching structure of unfoliaged deciduous trees from urban image sequences. The trees are generatively modeled in 3D by means of L-systems. A statistical approach, namely Markov Chain Monte Carlo – MCMC is employed together with cross correlation for extraction. Thereby we overcome the complexity and uncertainty of extracting and matching branches in several images due to weak contrast, background clutter, and particularly the varying order of branches when projected into different images. First results show the potential of the approach.

Cascades of classifiers constitute an important architecture for fast object detection. While boosting of simple (weak) classifiers provides an established framework, the design of similar architectures with more powerful (strong) classifiers has become the subject of current research. In this paper, we focus on greedy strategies recently proposed in the literature that allow to learn sparse Support Vector Machines (SVMs) without the need to train full SVMs beforehand. We show (i) that asymmetric data sets that are typical for object detection scenarios can be successfully handled, and (ii) that the complementary training of two sparse SVMs leads to sequential two-stage classifiers that slightly outperform a full SVM, but only need about 10% kernel evaluations for classifying a pattern.