Pattern Recognition in Soft Computing Paradigm

The following sections are included:

• Series Editor's Preface

• Volume Editor's Preface

• Contents

In this contribution, from three decades of research in various disciplines, algorithms for dimensionality reduction, serving for feature extraction and multivariate data projection are briefly reviewed and compared in an unifying approach. The applicability of the reviewed dimensionality reduction methods for data compression or discriminance analysis in classification systems is assessed. Techniques for ensuing interactive data visualization, data navigation and visual exploratory data analysis are presented. Advanced mapping algorithms and hierarchical concepts are introduced that in conjunction with interactive visualization techniques explicitly support the remarkable human perceptive and associative capabilities for visual exploratory data analysis. The benefit of the visualization approach for systematic recognition system design is pointed out.

The Self-Organizing Map (SOM) is one of the most popular neural network methods. It is a powerful tool in visualization and analysis of high-dimensional data in various engineering applications. The SOM maps the data on a two-dimensional grid which may be used as a base for various kinds of visual approaches for clustering, correlation and novelty detection. In this chapter, we present novel methods that enhance the SOM based visualization in correlation hunting and novelty detection. These methods are applied to two industrial case studies: analysis of hot rolling of steel and continuous pulp process. A research software for fast development of SOM based tools is briefly described.

Radiative transfer algorithms in combination with empirical formulae have been the most popular approach to the analysis of ocean primary productivity from remotely sensed images of the Earth. These methods fully rely on the limited amounts of ground truth data available and assumptions regarding how sensor, Earth's surface and atmospheric properties influence the radiation captured in different ranges of the electromagnetic spectrum. As these assumptions are restraining, multi-spectral and fusion techniques based on the application of unsupervised neural networks can contribute to the improvement in ocean colour studies and enable analysis of complex water types. This chapter presents the application of a hierarchy of self-organizing feature maps to clustering and differentiation of oceanic waters. The practical studies are performed on imagery captured all over the Pacific Ocean by the Ocean Colour and Temperature Scanner on board the Japanese satellite ADEOS.

Feature selection is very important for the success of any automated pattern recognition system. Removal of redundant features improves the efficiency of a classifier as well as cut down the cost of feature extraction. Recently artificial neural networks (ANN) are becoming popular for solving pattern classification problems. But the choice of proper architecture and model from various alternatives is still an open problem for research. In this work a new architecture, a modified version of multilayer feed forward neural network, proposed earlier for pattern classification has been used as a tool for feature selection. A proper algorithm for feature subset selection has been proposed in which the features are evaluated according to their contribution to classification rate of the net for the unknown samples. Simulation of the proposed algorithm has been done with two types of data sets and the results seem to be promising.

We present a soft computing approach to character recognition from printed documents. For feature extraction, we define a number of fuzzy sets on the Hough transform of character pattern pixels and synthesize additional fuzzy sets by t-norms. The height of each t-norm constitutes a feature element and a set of 'n' such feature elements form an n-dimensional feature vector for the character. A 3n-dimensional vector is then generated from the n-dimensional feature vector by defining three linguistic fuzzy sets, namely, weak, moderate and strong. These 3n-dimensional vectors form a Multilayer Perceptron (MLP) input for training by back propagation. The MLP outputs represent fuzzy sets denoting the belongingness of an input pattern to a number of fuzzy pattern classes. The feature set is chosen by optimizing a Feature Quality Index (FQI) using genetic algorithm. A two-state Markov chain is used to model degraded document images for simulation tests. The system can recognize characters with an accuracy of 98%.

Clustering is primarily used to uncover the true underlying structure of a given data set. Most algorithms for clustering often depend on initial guesses of the cluster centers and assumptions made as to the number of subgroups presents in the data. In this paper, we propose a method for fuzzy clustering without initial guesses on cluster number in the data set. Our method assumes that clusters will have the normal distribution. Our method can automatically estimate the cluster number and form the clusters according to the number. In it, Genetic Algorithms (GAs) with two chromosomic coding techniques are evaluated. Graph structured coding can derive high fitness value. Linear structured can save the number of generation.

This chapter deals with classification problems, such as diagnosis, in which classes are defined by categorical forms. Recently, soft computing techniques including learning algorithm in neural networks and fuzzy expert system are applied to the classification problems. Kanagawa et al. present an associative classification method using the cellular neural networks. This chapter proposes a new classification method using the associatron as an associative memory machine. We alter the associatron so as to have three stable leveled outputs, following in the steps of a method given by Kanagawa et al. Examples of diagnosis of liver disease and the problem of iris classification are demonstrated.

Poggio and Edelman have shown that for each object there exists a smooth mapping from an arbitrary view to its standard view and that the mapping can be learnt from a sparse data set. They have demonstrated that such a mapping function can be well approximated by Gaussian GRBF(Genaralized Radial Basis Function) network. In this paper, we have applied this network to the recognition of three kinds of hand shape changes, such as grasp, stroke and flap. Also during learning stage, we have introduced a structural learning algorithm to GRBF network in order to explore the small-sized and essential network structure for the recognition. Finally, we have designed a system to capture motion of the hand with this GRBF by using data glove.

Neural computing has emerged as a practical and powerful tool for "nonlinear" multivariate statistical analysis. In this paper, nonlinear discriminant analysis using a neural network is considered and applied to a medical diagnosis problem. The probabilistic interpretation of the network output is discussed in classification problems. The principle of the likelihood in network models is employed based on a probabilistic approach regarding the connection weights of the network as unknown parameters, and the maximum likelihood estimators of outputs are also introduced. Additionally, statistical techniques are formulated in terms of the principle of the likelihood of network models. The statistical tools for the inference illustrated here include i) deviance to evaluate the goodness-of-fit of a network model, ii) selection of the best fit model among several competing network models, and iii) likelihood-ratio chi-square statistics for pruning of neural network parameters and selection of a "best" subset of predictor variables.

In this chapter, we show that the measurement of all attribute values is not always necessary when new patterns are classified by trained neural networks. We assume that a neural network with n input units has already been trained for an n-dimensional pattern classification problem. Our goal is to reduce the average (or expected) number of attributes that are to be measured for classifying new patterns by the trained neural network. Our task is different from the feature selection because the number of measured attributes for classifying each new pattern is not the same. That is, some patterns can be classified by a few attributes while other patterns require the measurement of all the n attributes. Our interval-arithmetic-based approach, in which each unmeasured attribute is represented by an interval that includes its possible values, tries to reduce the number of attributes to be measured with no deterioration of the classification performance.

We propose a method for fuzzy rule generation directly from numerical data for designing classifiers. First a fuzzy partition is imposed on the domain of each feature, which results in a set of fuzzy values for each feature. Then a descriptor-pattern table is constructed using the training data and the fuzzy feature values. Rules are now discovered from the descriptor-pattern table. The rule generation process finds the distinct descriptors to discover simple rules and if required generates further rules using conjunction of common descriptors or conjunction of common descriptors and negation of distinct descriptors. A rule minimization process is then initiated to retain a small set of rules adequate to learn the training data. We suggest three possible schemes for generation of the initial fuzzy partitioning of the feature space and a genetic algorithm based tuning scheme is used to refine the rule base. Finally, the proposed scheme is tested on some real data. Unlike, most of the classifiers, the proposed method can detect ambiguous data and declare them to be unclassified - this is a distinct advantage.

A framework is presented, which allows for the automated generation of texture filters by exploiting the 2D-Lookup algorithm and its optimization by evolutionary algorithms. To use the framework, the user has to give an original image, containing the structural property-of-interest (e.g. a surface fault), and a binary image (goal image), wherein each position of the structural property-of-interest is labeled with the foreground color. Doing so, the framework becomes capable of evolving the configuration of the 2D-Lookup algorithm towards a texture filter for the structural property-of-interest. Genetic programming (GP) is used as the evolutionary algorithm. For this GP approach, a filter generator derives two operations based on formal superoperators from the tree, which represents an individual of the evolving population. The specification of the 2D-Lookup matrix is performed by a relaxation technique. The approach will be demonstrated on texture fault examples.

Texture image segmentation is the first essential and important step of low level vision. The extraction of texture features is a most fundamental problem to texture image segmentation. Many methods for extracting texture features of the image have been proposed, such as statistical features, co-occurrence features, two-dimensional AR features, and fractal based features etc. In this paper, a new method for extracting texture features of the image is proposed. In this method, the gray scale image is first decomposed into a series of binary images by variable thresholds, and then topological features of all of these binary images are computed. Using these topological features as texture features, we apply a pyramid linking with band-pass filter neural networks to segment the texture image into some homogeneous areas. Several experiments on synthetic texture images have been carried out to verify the efficacy of the new method.

These days, development of computer device leads to computer user explosion. On the other hand, conventional technology cannot handle user's diversity, because the computer cannot treat each user's subjectivity. It is important to materialize techniques which can answer user's subjective requests flexibly for the coming new media epoch. This research focuses on studying how subjective impressions are grounded in physical perceptions, especially in visual perception. In this paper, we propose a model for subjectivity and physical perceptions, and give an example in an image database system.

ART MUSEUM is our prototype system which can treat each user's subjective interpretation for advanced image database. It allows users to retrieve images from a database of impressionist paintings (e.g. Monet's works) by using adjective words. The system learns the correlation between the feature of sample paintings and user's descriptions, and provides interactive and progressive learning mechanisms. We show the interesting experimental results on ART MUSEUM (Multimedia Database with Sense of Color and Construction upon the Matter of ART) system.

The following sections are included:

• About the Authors

• Keyword Index