Francisco Javier López Aligué

SEGMENTATION OF BOVINE LIVESTOCK IMAGES USING GA AND ASM IN A TWO-STEP APPROACH

In this work a system based on genetic algorithms (GA) and active shape models (ASM) is presented that, in a two-step approach, is able to accurately find a particular object within the image. The aim of the first stage is to approximately locate the object in order to serve as an initialization for the second. Following a method similar to that used by other authors, a model of the desired shape (cows in lateral position) is constructed using PDM, and later the search within the image is carried out based on instances of that model, and using genetic algorithm techniques, for which several objective functions are suggested. In the second stage we attempt to make a fine adjustment of our instance to the object in the image, starting from the result of the previous process. Consequently, we propose two different techniques compared afterwards: deformable contours (specifically active shape models) and a restricted genetic search. Finally, the system is tested over a database of 309 animal images, and results are presented.

A Comparative Study of Two Neural Models for Cloud Screening of Iberian Peninsula Meteosat Images

In this work we make a comparative study of the results obtained in the automatic interpretation of the Iberian Peninsula Meteosat images by means of neural networks techniques, in particular, multi-layer perceptrons and self organizing maps. The interpretation of these images implies their segmentation in the classes SEA (S), LAND (L), LOW CLOUDS (CL), MIDDLE CLOUDS (CM), HIGH CLOUDS (CH) and CLOUDS WITH VERTICAL GROWTH (CV).

NeuSim: A Modular Neural Networks Simulator for Beowulf Clusters

We present a neural net works simulator that we have called NeuSim, designed to work in Beowulf clusters. We have been using this simulator during the last years over several architectures and soon we want to distribute it under GPL license. In this work we offer a detailed description of the simulator, as well as the performance results obtained with a Beowulf cluster of 24 nodes.

Generic Neural Network Model and Simulation Toolkit

This work presents a generic higher-order model of neuron behaviour, connection scheme and learning rule, suited for high-speed parallel processing. In contrast to the construction of a real application, it would be more operational to expend effort in parameterizing the problem-solving architecture, offering a testbed as a useful simulation tool for experimenting with a variety of network designs within the said generic model. We include some initial simulation results applied to image processing and pattern recognition tasks.

A CNN Model for Grey Scale Image Processing

A modification of the CNN model is proposed in this work. An iterated-map is defined instead of the original differential equation while a sigmoid function is taken as the cell output. Modifications in the structure and values of the synaptic scheme allow the use of the model in different tasks. The networks behaviour is mainly determined by its feedforward term, while feedback simply adds optimization to the networks output The networks stability is discussed. Edge detection and contrast enhancement are performed, stressing the fact that they are only different aspects of the same network property: the enhancement of brightness gradients. Some examples are presented in which grey scale images are processed, revealing the capabilities of the model both in detecting edges and in enhancing contrasts.

A High Order Neural Model

The interaction between afferent nerves had always been regarded as a phenomenon which is produced outside the neuron. This work presents an extension of the classic concept of interaction between inputs, including the possibility of higher-order effects at the level of neuronal activity function. The mathematics is formulated with a view to its algorithmic simulation being implemented in a multiprocessor system by means of an adequate programming language running as a multi-elemental processor parallel computer. The system that is finally presented is a higher-ordered neural network with nonsupervised learning implemented in a multilayer structure.

Associative Memory for Early Detection of Breast Cancer

We present a new associative neural network design especially indicated for the early detection of malignant lesions in breast cancer screening. It is a BAM in which we have made some changes to the functioning of its neurons, and for which we have developed an automatic selection algorithm for the prototypes used to calculate the thresholds of the neurons conforming the input layer. The result is a structure that, while considerably reduced, is highly effective in identifying the images that indicate the presence of malignant tumours in screening for breast cancer. We endowed the network with a special pre-processing stage for the treatment of this kind of radiographic image. This pre-processing yields a more detailed analysis of possible signs of tumours.

A Neural Associative Pattern Classifier

In this work, we study the behaviour of the Bidirectional Associative Memory (BAM) in terms of the supporting neural structure, with a view to its possible improvements as a useful Pattern Classifier by means of class associations from unknown inputs, once mentioned classes have been previously defined by one or even more prototypes. The best results have been obtained by suitably choosing the training pattern pairs, the thresholds, and the activation functions of the networks neurones, by means of certain proposed methods described in the paper. In order to put forward the advantages of these proposed methods, the classifier has been applied on an especially popular hand-written character set as the well-known NIST#19 character database, and with one of the UCIs data bases. In all cases, the method led to a marked improvement in the performance achievable by a BAM, with a 0% error rate.

Synthesis of Adaptive Memories with Neural Networks

The general formulation of bidirectional associative memories presents certain difficulties when the associations of pairs of patterns do not suppose a local energy minimun. To avoid these problems, the present paper describes an adaptive scheme which al lows the correlation matrix to be modified so as to reach the energy minimun while at the same time identifying the input patterns. The strategy described here allows the adaptation of the matrix to be performed for each external input, so that it can henceforth be described as a supervised type of training scheme. A consequence is its synthesis by means of neural networks with both the BAM and the adaptive mechanism itself integrated in distinct layers, allowing either of them to be changed without altering the other.

A Structured Development for the Implementation of a Multilayer Neural Network (Cram Project)

The main target of the CRAM (Computer Research on Associative Memory) Project is the practical implementation of a multilayer virtual Neural Network. The hardware used is a multiprocessor configuration IEEE 1014 Standard (VME Bus) with the 68020 CPU and the 68882 Floating Point Coprocessor running under the VersaDos Operating System, well suited for developing and debugging real-time programs.