Catálogo de publicaciones - libros

In this article we investigate the influence of a Pareto-like noise model on the performance of an artificial neural network used to predict a nonlinear time series. A Pareto-like noise model is, in contrast to a Gaussian noise model, based on a power law distribution which has long tails compared to a Gaussian distribution. This allows for larger fluctuations in the deviation between predicted and observed values of the time series. We define an optimization procedure that minimizes the mean squared error of the predicted time series by maximizing the likelihood function based on the Pareto-like noise model. Numerical results for an artificial time series show that this noise model gives better results than a model based on Gaussian noise demonstrating that by allowing larger fluctuations the parameter space of the likelihood function can be search more efficiently. As a consequence, our results may indicate a more generic characteristics of optimization problems not restricted to problems from time series prediction.

Neural network approaches for time series forecasting, which have the property of simpleness, nonlinearity and effectiveness, have been broadly utilized in many domains. In this paper, an improved training algorithm of back-propagation neural network for time series forecasting by using dynamic learning rate in the training process is proposed. The results of some studied cases demonstrate this algorithm can increase the efficiency of neural network training and the precision of forecasts.

An evolutionary algorithm for parameter estimation of a kernel method for noisy and irregularly sampled time series is presented. We aim to estimate the time delay between time series coming from gravitational lensing in astronomy. The parameters to estimate include the delay, the width of kernels or smoothing, and a regularization parameter. We use mixed types to represent variables within the evolutionary algorithm. The algorithm is tested on several artificial data sets, and also on real astronomical observations. The performance of our method is compared with the most popular methods for time delay estimation. An statistical analysis of results is given, where the results of our approach are more accurate and significant than those of other methods.

In this study, ant colony optimization (ACO) is integrated with the self-organizing map (SOM) for image segmentation. A comparative study with the combination of ACO and Simple Competitive Learning (SCL) is provided. ACO follows a learning mechanism through pheromone updates. In addition, pheromone and heuristic information are normalized and the effects on the results are investigated in this report. Preliminary experimental results indicate that the normalization of the parameters can improve the image segmentation results.

Finding correspondence regions between images is fundamental to recovering three dimensional information from multiple frames of the same scene and content based image retrieval. To be good, correspondence regions should be easily found, richly characterised and have a good trade-off between density and uniqueness. Maximally stable extremal regions (MSER’s) are amongst the best known methods to tackle this problem. Here, we present an implementation of the sieve algorithm that not only generates MSER’s but can also generate stable salient contours (SSC’s) in different ways. The sieve decomposes the image according to local grayscale intensities and produces a tree in nearly () where is the number of pixels. The exact shape of the tree depends on the criteria used to control the merging of extremal regions with less extreme neighbours. We call the resulting data structure a ‘structured image’. Here, a structured image in which MSER’s are embedded is compared with those associated with two types of SSC’s. The correspondence rate generated by each of these methods is compared using the standard evaluation method due to Mikalajczyk and the results show that SSC’s identified using colour and texture moments are generally better than the others.

The Contourlet Transform (CT) can capture the intrinsic geometrical structure of an image. The CT is a redundant transform, and for coding applications this can be a disadvantage. In order to avoid the contourlet redundancy we change the pyramidal stage by a multiscale stage. The new non redundant transform is called: The Wavelet Based Contourlet Transform. We take the advantages offered by the new transform to build a novel feature preserving image coder. The preservation is made both: by a stage of feature definition and extraction and by a proposed modified version of the SPIHT coder. SPIHT modification allows selecting a transform coefficient not only by magnitude, but as pertaining to a feature of interest map. We present tests in order to demonstrate the good performance of the coder. Finally, we compare the results with other existing methods. The coder designed performs well even at very low bit rates.

A central issue in the use of vector quantization (VQ) for speech or image compression is the specification of the codebook. In this paper, the design of an evolutionary codebook based on morphological associative memories (MAM) is presented. The algorithm proposed for codebook generation involves two steps. First, having a set of images, one of the images is chosen to create the initial codebook. The algorithm applied to the image for codebook generation uses the morphological autoassociative memories (MAAM). Second, an evolution process of codebook creation occurs applying the algorithm on new images. This process adds the information codified of the next image to the codebook allowing to recover the images with better quality without affecting the processing speed. The performance of the generated codebook is analyzed in case when MAAM in both and categories are used. The presented algorithm was applied to image set after discrete cosine transformation followed by a quantization process. The proposed algorithm has a high processing speed and provides a notable improvement in signal to noise ratio.

This paper presents a novel shape-based image retrieval model. We focused on the shape feature of objects inside images because there is evidence that natural objects are primarily recognized by their shapes. Using this feature of objects the semantic gap is reduced considerably. Our technique contains an alternative representation of shapes that we have called two segment turning function (2STF). Two segment turning function has a set of invariant features such as invariant to rotation, scaling and translation. Then, based on 2STF, we proposed a complete new strategy to compute a similarity among shapes. This new method was called Star Field (SF). To test the proposed technique, which is made up of a set of new methods mentioned above, a test-bed CBIR system was implemented. The name of this CBIR System is IRONS. IRONS stands for ”Image Retrieval based ON Shape”. Finally, we compared our results with a set of well known methods obtained similar results without the exhaustive search of many of them. This former feature of our proposal is one of the most important contributions of our technique to the visual information retrieval area.

We present an indexing and retrieval system of historic art images based on fuzzy shape similarity. The system is composed of three principal components: object annotation, object shape indexing, and query/retrieval. The object annotation in database images is done manually offline. The object shape indexing and retrieval, however, are done automatically. Annotated object shapes are indexed using an extended curvature scale space (CSS) descriptor suitable for concave and convex shapes. The query/retrieval of pertinent shapes from the database starts with a user drawing query (with a computer mouse or a pen) that is compared to entries in the database using a fuzzy similarity measure. The system is tested on a set of complex color and grey scale images of ancient documents, mosaics, and artifacts from the National Library of Tunisia, the National Archives of Tunisia, and a selection of Tunisian museums. The system’s recall and precision rates were 83% and 60%, respectively.

In electronics mass-production manufacturing, printed circuit board (PCB) inspection is a time consuming task. Manual inspection does not guarantee that PCB defects can be detected. In this paper, a spatial filtering and wavelet-based automatic optical inspection system for detect PCB defects is presented. This approach combines wavelet image compression utility and spatial filtering. Defects are detected by subtracting the approximations of reference image wavelet transform and test image wavelet transform followed by a median filter stage. Finally, defect image is obtained by computing the inverse wavelet transform. Advantages of this approach are also described.