Catálogo de publicaciones - libros

This paper presents some experiments in language identification for Spanish and Basque, both official languages in the Basque Country in the North of Spain. We focus on four methods based on phone decoding, some of which make use of phonotactic knowledge. We run also a comparison between the use of a generic and a task-specific phonotactic model. Despite initial poor performances, significant accuracies are achieved when better phonotactic knowledge is used. The use of a task-specific phonotactic model performs slightly better, but it is only useful when using less expensive methods. Finally, we present a temporal evolution of the accuracies. Results show that 5-6 seconds are enough to achieve similar percentage of correctly classified utterances.

Speech recognition systems have proved their usefulness in very different tasks. Nevertheless, the present state-of-the-art of the speech technologies does not make it possible to achieve perfect transcriptions in most of the cases. Owing to this fact, human intervention is necessary to check and correct the results of such systems. We present a novel approach that faces this problem by combining the efficiency of the automatic speech recognition systems with the accuracy of the human transcriptor. The result of this process is a cost-effective perfect transcription of the input signal.

The aim of this work is to integrate segments of words into a category-based Language Model. Two proposals of this kind of models are presented. On the other hand an interpolation of a category-based model with a classical word-based Language Model is studied as well. The models were integrated into an ASR system and evaluated in terms of WER. Experiments on a spontaneous dialogue corpus in Spanish are reported. These experiments show that integrating word segments in a category-based Language Model, a better performance of the model can be achieved.

In this paper, a new approach to the Part-of-Speech (PoS) tagging problem is proposed. The PoS tagging problem can be viewed as a special translation process where the source language is the set of strings being considered and the target language is the sequence of POS tags. In this work, we have used phrase-based machine translation technology to tackle the PoS tagging problem. Experiments on the Penn Treebank WSJ task were carried out and very good results were obtained.

Bilingual documentation has become a common phenomenon in official institutions and private companies. In this scenario, the categorization of bilingual text is a useful tool. In this paper, different approaches will be proposed to tackle this bilingual classification task. On the one hand, three finite-state transducer algorithms from the grammatical inference framework will be presented. On the other hand, a naive combination of smoothed -gram models will be introduced. To evaluate the performance of bilingual classifiers, two categorized bilingual corpora of different complexity were considered. Experiments in a limited-domain task show that all the models obtain similar results. However, results on a more open-domain task denote the supremacy of the naive approach.

Detection of lane markings based on a camera sensor can be a low cost solution to lane departure and curve over speed warning. A number of methods and implementations have been reported in the literature. However, reliable detection is still an issue due to cast shadows, wearied and occluded markings, variable ambient lighting conditions etc. We focus on increasing the reliability of detection in two ways. Firstly, we employ a different image feature other than the commonly used edges: ridges, which we claim is better suited to this problem. Secondly, we have adapted RANSAC, a generic robust estimation method, to fit a parametric model of a pair or lane lines to the image features, based on both ridgeness and ridge orientation. In addition this fitting is performed for the left and right lane lines simultaneously, thus enforcing a consistent result. We have quantitatively assessed it on synthetic but realistic video sequences for which road geometry and vehicle trajectory ground truth are known.

Matrics is a system for recognition of car license plates. It works on standard PC equipment with low-priced capture devices and achieves real-time performance (10 frames per second) with state of the art accuracy: the character error rate is below 1% and the plate error rate is below 3%. The recognition process is divided in two phases: plate localization and plate decoding. The system finds the plate analyzing the connected components of the image after binarization. The decoding algorithm is a Two Level process which uses fast template-based classification techniques in its first stage and optimal segmentation in the second stage. On the whole, the system represents a significant improvement over a previous version which was based on HMM.

The labeling of large quantities of medical video data by clinicians is a tedious and time consuming task. In addition, the labeling process itself is rigid, since it requires the expert’s interaction to classify image contents into a limited number of predetermined categories. This paper describes an architecture to accelerate the labeling step using eye movement tracking data. We report some initial results in training a Support Vector Machine (SVM) to detect cancer polyps in colonoscopy video, and a further analysis of their categories in the feature space using Self Organizing Maps (SOM). Our overall hypothesis is that the clinician’s eye will be drawn to the salient features of the image and that sustained fixations will be associated with those features that are associated with disease states.

We evaluate the appropriateness of applying a functional rather than the typical vectorial approach to a pattern recognition problem. The problem to be resolved was to construct an online system for controlling wood-pulp chip granulometry quality for implementation in a wood-pulp factory. A functional linear model and a functional logistic model were used to classify the hourly empirical distributions of wood-chip thicknesses estimated on the basis of images produced by a 3D laser scanner. The results obtained using these functional techniques were compared to the results of their vectorial counterparts and support vector machines, whose input consisted of several statistics of the hourly empirical distribution. We conclude that the empirical distributions have sufficiently rich functional traits so as to permit the pattern recognition process to benefit from the functional representation.

We present in this paper a method for implementing moment functions in a CMOS retina for object localization, and pattern recognition and classification applications. The method is based on the use of binary patterns and it allows the computation of different moment functions such as geometric and Zernike moments of any orders by an adequate choice of the binary patterns. The advantages of the method over other methods described in the literature are that it is particularly suitable for the design of a programmable retina circuit where moment functions of different orders are obtained by simply loading the correct binary patterns into the memory devices implemented on the circuit. The moment values computed by the method are approximate values, but we have verified that in spite of the errors the approximate values are significant enough to be applied to classical shape localization and shape representation and description applications.