Catálogo de publicaciones - libros

In this paper an FPGA-based implementation of a sequential discrete time cellular neural network (DT-CNN) with 3×3 templates is described. The architecture is based on a single pipelined cell which is employed to emulate a CNN with larger number of neurons. This solution diminishes the use of hardware resources on the FPGA and allows the cell to process real time input data in a sequential mode. Highly efficient FPGA implementation has been achieved by manual design based on low level instantiation and placement of hardware primitives. The Intellectual Property Core offers an appropriate tradeoff between area and speed. Our architecture has been developed to assist designers implementing discrete CNN models with performance equivalent to hundreds or millions of neurons on low cost FPGA-based systems.

The continuous advances in VLSI technologies, computer architecture and software development tools make it difficult to find the adequate implementation platform of an ANN for a given application. This paper describes HANNA, a software tool designed to automate the generation of hardware prototypes of MLP and MLP-like neural networks over FPGA devices. Coupled with traditional Matlab/Simulink environments the generated model can be synthesized, downloaded to the FPGA and co-simulated with the software version to trade off area, speed and precision requirements. The tool and our design methodology are validated through two examples.

One of the typical applications of neural networks is based on their ability to generate fitting surfaces. However, for certain problems, error specifications are very restrictive, and so, the performance of these networks must be improved. This is the case of analog CMOS circuits, where models created must provide an accuracy which some times is difficult to achieve using classical techniques. In this paper we describe a modelling method for such circuits based on the combination of classical neural networks and electromagnetic techniques. This method improves the precision of the fitting surface generated by the neural network and keeps the training time within acceptable limits.

In this paper, we propose a wavelet neural network suitable for reducing nonlinear distortion introduced by a traveling wave tube amplifier (TWTA) over multicarrier systems. Parameters of the proposed network are identified using an hybrid training algorithm, which adapts the linear output parameters using the least square algorithm and the nonlinear parameters of the hidden nodes are trained using the gradient descent algorithm. Computer simulation results confirm that the proposed wavelet network achieves a bit error rate performance very close to the ideal case of linear amplification.

Currently, the accuracy of speech processing systems is stro- ngly affected by the acoustic noise. This is a serious obstacle to meet the demands of modern applications and therefore these systems often needs a noise reduction algorithm working in combination with a precise voice activity detector (VAD). This paper presents a new voice activity detector (VAD) for improving speech detection robustness in noisy environments and the performance of speech recognition systems. The algorithm defines an optimum likelihood ratio test (LRT) involving Multiple and correlated Observations (MO). The so defined decision rule reports significant improvements in speech/non-speech discrimination accuracy over existing VAD methods with optimal performance when just a single observation is processed. The algorithm has an inherent delay in MO scenario that, for several applications including robust speech recognition, does not represent a serious implementation obstacle. An analysis of the methodology for a pair-wise observation dependence shows the improved robustness of the proposed approach by means of a clear reduction of the classification error as the number of observations is increased. The proposed strategy is also compared to different VAD methods including the G.729, AMR and AFE standards, as well as recently reported algorithms showing a sustained advantage in speech/non-speech detection accuracy and speech recognition performance.

This paper analyzes the behaviour of closed-loop controlled systems. Starting from the measured data, the aim is to establish a classification of the system operation states. Digital Signal Processing is used to represent temporal signal with spatial patterns. A neuro-fuzzy scheme (dFasArt) is proposed to classify these patterns, in an on-line way, characterizing the state of controller performance. A real scale plant has been used to carry out several experiments with good results.

PET is a nuclear medical examination which constructs a three-dimensional image of metabolism inside the body; in this article in particular, images are taken from the brain. The high complexity inherent to the interpretation of the brain images makes that any help is important to the specialists in order to accurate the diagnostic. In order to reach reliable and good images, a normalization process is suggested in this paper, consisting of centring the brain in the three-dimensional image, scaling it according to a template brain and, finally, rotating the brain according to the inclination of the template. For not reducing the quality of the information the application works with PET image format and radioactivity measures instead of translate to an ordinary colour image.

This paper describes an automatic method for segmenting single and multiple neoplastic hepatic lesions in computed-tomography (CT) images. The structure of the liver is first segmented using the approximate contour model. Then, the appropriate histogram transformations are performed to enhance neoplastic focal lesions in CT images. To segment neoplastic lesions, images are processed using binary morphological filtration operators with the application of a parameterized mean defining the distribution of gray-levels of pixels in the image. Then, the edges of neoplastic lesions situated inside the liver contour are localized. To assess the suitability of the suggested method, experiments have been carried out for two types of tumors: hemangiomas and hepatomas. The experiments were conducted on 60 cases of various patients. Thirty CT images showed single and multiple focal hepatic neoplastic lesions, and the remaining 30 images contained no disease symptoms. Experimental results confirmed that the method is a useful tool supporting image diagnosis of the normal and abnormal liver. The proposed algorithm is 78.3% accurate.

A system that performs the detection and tracking of a face in real-time in real video sequences is presented in this paper. The face is detected in a complex environment by a model of human colour skin. Very good results are obtained, since the colour segmentation removes almost all the complex background and it is realized to a very high-speed, making the system very robust. On the other hand, fast and stable real-time tracking is then achieved via biometric feature extraction of face using connected components labelling. Tracking does not require a precise initial fit of the model. Therefore, the system is initialised automatically using a very simple 2D face detector based on target ellipsoidal shape. Results are presented showing a significant improvement in detection rates when the whole sequence is used instead of a single image of the face. Experiments in tracking are reported.

The goal of the See ColOr project is to achieve a non-invasive mobility aid for blind users that will use the auditory pathway to represent in real-time frontal image scenes. More particularly, we have developed a prototype which transforms HSL coloured pixels into spatialized classical instrument sounds lasting for 300 ms. Hue is sonified by the timbre of a musical instrument, saturation is one of four possible notes, and luminosity is represented by bass when luminosity is rather dark and singing voice when it is relatively bright. Our first experiments are devoted to static images on the computer screen. Six participants with their eyes covered by a dark tissue were trained to associate colours with musical instruments and then asked to determine on several pictures, objects with specific shapes and colours. In order to simplify the protocol of experiments, we used a tactile tablet, which took the place of the camera. Overall, experiment participants found that colour was helpful for the interpretation of image scenes.