Catálogo de publicaciones - libros

In this paper we present a model based approach for super-resolving an image from a sequence of zoomed observations. From a set of images taken at different camera zooms, we super-resolve the least zoomed image at the resolution of the most zoomed one. Novelty of our approach is that decimation matrix is estimated from the given observations themselves. We model the most zoomed image as an autoregressive (AR) model, learn the parameters and use in regularization to super-resolve the least zoomed image. The AR model is computationally less intensive as compare to Markov Random Field (MRF) model hence the approach can be employed in real-time applications. Experimental results on real images with integer zoom settings are shown. We also show how the learning of AR parameters in subblocks using Panchromatic (PAN) image gives better results for the multiresolution fusion process in remote sensing applications.

The extraction of reliable features is a key issue for autonomous underwater vehicle navigation. Imaging sonars can produce acoustic images of the surroundings of the vehicle. Despite of the noise, the phantoms and reflections, we believe that they are a good source for features since they can work in turbid water where other sensors like vision fail. Moreover, they can cover wide areas incrementing the number of features visible within a scan. This work presents an algorithm to extract linear features from underwater structured environments including as major contributions a novel sonar model sensor and an adapted implementation of the Hough transform.

This paper presents a cascaded system for recognition of the circular road-signs. The system consists of two compound detectors-classifiers. Each operates on the Gaussian scale-space and does template matching in the log-polar domain. The first module is responsible for detection of the potential sign areas at the coarsest level of the pyramid. The second one, in turn, refines the already found places at the finest level. Thanks to this composition, as well as to the efficient matching in the log-polar domain, the system is very robust in terms of recognition of the signs with different scales and rotations, as well as under partial occlusions, poor illumination conditions, and noise.

For a learning algorithm, especially a linear algorithm, it can usually be extended to its kernel version endowed with the power of extracting non-linear features. In this paper, we explore two key questions in the kernelization of an algorithm. The first is the existence of the kernel version of an algorithm. We propose a new method to determine whether an algorithm can be kernelized. It has the advantage that it is not limited by the specific form of the algorithm and shows an insight view of kernelization. The second question is how to kernelize an algorithm. We prove a kind of equivalence between two kernelization processes. Related details are also discussed.

In this paper we propose a probabilistic observation model for stereo vision systems which avoids explicit data association between observations and the map by marginalizing the observation likelihood over all the possible associations. We define observations as sets of landmarks composed of their 3D locations, assumed to be normally distributed, and their SIFT descriptors. Our model has been integrated into a particle filter to test its performance in map building and global localization, as illustrated by experiments with a real robot.

The -nearest-neighbor rule is a well known pattern recognition technique with very good results in a great variety of real classification tasks. Based on the neighborhood concept, several classification rules have been proposed to reduce the error rate of the -nearest-neighbor rule (or its time requirements). In this work, two new geometrical neighborhoods are defined and the classification rules derived from them are used in several real data classification tasks. Also, some voting ensembles of classifiers based on these new rules have been tested and compared.

In this paper, we introduce a new framework for classification of short duration voltage reductions in the area of Power Quality Monitoring using Multiway Principal Component Analysis (MPCA). Firstly, we recast the sags occurred in High Voltage (HV) and Medium Voltage (MV) lines in a format which is suitable for MPCA. Then, MPCA technique is employed for building statistical models for classification of sags originated in HV and MV networks and recorded in the same substation. Projecting sags registered in different substations to MPCA models of other substations has been also explored to deduce similarities and dissimilarities between different substations according to the sags registered in them.

We describe the recognition system we used for the “On-Line Tamil Handwritten Character Recognition Competition”, hosted by the “International Workshop on Frontiers in Handwriting Recognition”. The system is based on continuous density and characterized by feature extraction based on time and frequency-domain features. In that contest we have obtained the third best score in character classification accuracy.

This paper introduces a new type of feature in text categorization. Based on an interesting linguistic observation, Loose N-gram feature, defined as co-occurring words within limited range, is quite different from traditional features, such as words, phrases or n-grams. Not only retaining useful context information, this kind of feature also has considerable classification ability. The features generated by our algorithm have acceptable statistical characteristics, thus can effectively avoid the sparseness problem. Experiment results show that the Loose N-gram feature is helpful and promising in statistical text categorization systems, especially for the categorization tasks which rely on more semantic information. Our new type of feature could also be helpful in Information Retrieval research.

A constrained variational deconvolution approach for multi-channel images is presented. Constraints are enforced through a reparametrisation which allows a differential geometric reinterpretation. This view point is used to show that the deconvolution problem can be formulated as a standard gradient descent problem with an underlying metric that depends on the imposed constraints. Examples are given for bound constrained colour image deblurring, and for diffusion tensor magnetic resonance imaging with positive definiteness constraint. Numerical results illustrate the effectiveness of the methods.