In this paper, a source coding model for learning multiple concept descriptions of data is proposed. Our source coding model is based on the concept of transmitting data over multiple channels, called multiple description (MD) coding. In particular, frame expansions have been used in our MD coding models for pattern classification. Using this model, there are several interesting properties within a class of multiple classifier algorithms that share with our proposed scheme. Generalization of the MD view under an extension of local discriminant basis towards the theory of frames allows the formulation of a generalized class of low-complexity learning algorithms applicable to high-dimensional pattern classification. To evaluate this approach, performance results for automatic target recognition (ATR) are presented for synthetic aperture radar (SAR) images from the MSTAR public release data set. From the experimental results, our approach outperforms state-of-the-art methods such as conditional Gaussian signal model, Adaboost, and ECOC-SVM.

This paper investigates the application of the computational model of Cellular Automata (CA) for pattern classification of real valued data. A special class of CA referred to as Fuzzy CA (FCA) is employed to design the pattern classifier. It is a natural extension of conventional CA, which operates on binary string employing boolean logic as next state function of a cell. By contrast, FCA employs fuzzy logic suitable for modeling real valued functions. A matrix algebraic formulation has been proposed for analysis and synthesis of FCA. An efficient formulation of Genetic Algorithm (GA) is reported for evolution of desired FCA to be employed as a classifier of datasets having attributes expressed as real numbers. Extensive experimental results confirm the scalability of the proposed FCA based classifier to handle large volume of datasets irrespective of the number of classes, tuples, and attributes. Excellent classification accuracy has established the FCA based pattern classifier as an efficient and cost-effective solutions for the classification problem.

Active Queue Management (AQM) can maintain smaller queuing delay and higher throughput by purposefully dropping packets at the intermediate nodes. Most of the existing AQM schemes follow the probability dropping mechanism originated from Random Early Detection (RED). In this paper, we develop a novel packet dropping mechanism for AQM through designing a two-category classifier based on the Fisher Linear Discriminate approach. The simulation results show that the new scheme outperforms other well-known AQM schemes, such as RED, AdaptiveRED, AVQ, PI, REM etc., in the integrated performance. Additionally, our mechanism is simple since it requires few CPU cycles, which makes it suitable for the high-speed routers.

We investigate strategies to improve the utterance verification performance using a 2-class pattern classification approach, including: utilizing N-best candidate scores, modifying segmentation boundaries, applying background and out-of-vocabulary filler models, incorporating contexts, and minimizing verification errors via discriminative training. A connected-digit database recorded in a noisy, moving car with a hands-free microphone mounted on the sun-visor is used to evaluate the verification performance. The equal error rate (EER) of word verification is employed as the sole performance measure. All factors and their effects on the verification performance are presented in detail. The EER is reduced from 29%, using the standard likelihood ratio test, down to 21.4%, when all features are properly integrated.

In face recognition, simple classifiers are frequently used. For a robust system, it is common to construct a multi-class classifier by combining the outputs of several binary classifiers; this is called output coding method. The two basic output coding methods for this purpose are known as OnePerClass (OPC) and PairWise Coupling (PWC). The performance of output coding methods depends on accuracy of base dichotomizers. Support Vector Machine (SVM) is suitable for this purpose. In this paper, we review output coding methods and introduce a new sequential fusion method using SVM as a base classifier based on OPC and PWC according to their properties. In the experiments, we compare our proposed method with others. The experimental results show that our proposed method can improve the performance significantly on the real dataset.

Vector quantization (VQ) is an attractive image compression technique. VQ utilizes the high correlation between neighboring pixels in a block, but disregards the high correlation between the adjacent blocks. Unlike VQ, side-match VQ (SMVQ) exploits codeword information of two encoded adjacent blocks, the upper and left blocks, to encode the current input vector. However, SMVQ is a fixed bit rate compression technique and doesn't make full use of the edge characteristics to predict the input vector. Classified side-match vector quantization (CSMVQ) is an effective image compression technique with low bit rate and relatively high reconstruction quality. It exploits a block classifier to decide which class the input vector belongs to using the variances of neighboring blocks' codewords. As an alternative, this paper proposes three algorithms using gradient values of neighboring blocks' codewords to predict the input block. The first one employs a basic gradient-based classifier that is similar to CSMVQ. To achieve lower bit rates, the second one exploits a refined two-level classifier structure. To reduce the encoding time further, the last one employs a more efficient classifier, in which adaptive class codebooks are defined within a gradient-ordered master codebook according to various prediction results. Experimental results prove the effectiveness of the proposed algorithms.

We propose an adaptive complex-amplitude texture classifier that takes into consideration height as well as reflection statistics of interferometric synthetic aperture radar (SAR) images. The classifier utilizes the phase information to segment the images. The system consists of a two-stage preprocessor and a complex-valued SOFM. The preprocessor extracts a complex-valued feature vectors corresponding to height and reflectance statistics of blocks in the image. The following SOFM generates a set of templates (references) adaptively and classifies a block into one of the classes represented by the templates. Experiment demonstrates that the system segments an interferometric SAR image successfully into a lake, a mountain, and so on. The performance is better than that of a conventional system dealing only with the amplitude information.

This paper describes a method for training a pattern classifier that will perform well after it has been adapted to changes in input conditions. Considering the adaptation methods which are based on the transformation of classifier parameters, we formulate the problem of optimizing classifiers, and propose a method for training them. In the proposed training method, the classifier is trained while the adaptation is being carried out. The objective function for the training is given based on the recognition performance obtained by the adapted classifier. The utility of the proposed training method is demonstrated by experiments in a five-class Japanese vowel pattern recognition task with speaker adaptation.

In this paper, we present a class of combinatorial-logical classifiers called test feature classifiers. These are polynomial functions that can be used as pattern classifiers of binary-valued feature vectors. The method is based on so-called tests, sets of features, which are sufficient to distinguish patterns from different classes of training samples. Based on the concept of test we propose a new distance-based test feature classifiers. To test the performance of the classifiers, we apply them to a well-known phoneme database and to a textual region location problem where we propose a new effective textual region searching system that can locate textual regions in a complex background. Experimental results show that the proposed classifiers yield a high recognition rate than conventional ones, have a high ability of generalization, and suggest that they can be used in a variety of pattern recognition applications.

In order to investigate the dynamic behavior of quantized interconnection neural networks on neuro-chips, we have designed and fabricated hardware neural networks according to design rule of a 1.2 µm CMOS technology. To this end, we have developed programmable synaptic weights for the interconnection with three values of 1 and 0. We have tested the chip and verified the dynamic behavior of the networks in a circuit level. As a result of our study, we can provide the most straightforward application of networks for a dynamic pattern classifier. The proposed network is advantageous in that it does not need extra exemplar to classify shifted or reversed patterns.

Feedback of class memberships is incorporated into multimodal pattern classifiers and their unsupervised learning algorithm is presented. Classification decision at low levels is revised by the feedback information which also enables the reconstruction of patterns at low levels. The effects of the feedback are examined for the McGurk effect by using a simple model.

A major topic in the field of network and telecommunications is doing business on the World Wide Web (WWW), which is called Electronic Commerce (EC). Another major topic is blending Artificial Intelligence (AL) techniques with the WWW. In the Ex-W-Pert Project we have proposed an agent model for EC components that blends the traditional expert systems' reasoning engine with a multi-layer knowledge base, communication and documentation engines. In this project, EC is viewed as a society of software agents, such as customer, search, catalog, manufacturer, dealer, delivery and banker agents, interacting and negotiating with each other. Each agent has a knowledge-base and a reasoning engine, a communication engine and a documentation engine. The knowledge-base is organized in three layers: skill layer, rule layer and knowledge layer (S-R-K layers). In this project, for each EC agent, we identify the class of problems to be solved and build the knowledge base gradually for each layer. We believe that using this multi-layer knowledge base system will speed up the reasoning and ultimately reduce the operation costs.

Many image-processing techniques are based on texture features or gradation features of the image. However, Landsat images are complex; they also include physical features of reflection radiation and heat radiation from land cover. In this paper, we describe a method of constructing a super-resolution image of Band 6 of the Landsat TM sensor, oriented to analysis of an agricultural area, by combining information (texture features, gradation features, physical features) from other bands. In this method, a knowledge-based hierarchical classifier is first used to identify land cover in each pixel and then the least-squares approach is applied to estimate the mean temperature of each type of land cover. By reassigning the mean temperature to each pixel, a finer spatial resolution is obtained in Band 6. Computational results show the efficiency of this method.

The millimeter wave (MMW) radar has good compromise characteristics of both microwave radar and optical sensors. It has better angular and range resolving abilities than microwave radar, and a longer penetrating range than optical sensors. We used the MMW radar to detect targets located in the sea and among sea ice clutter based on fractals, wavelets, and neural networks. The wavelets were used as feature extractors to decompose the MMW radar images and to extract the feature vectors from approximation signals at different resolution levels. Unsupervised neural classifiers with parallel computational architecture were used to classify sea ice, sea water and targets based on the competitive learning algorithm. The fractal dimensions could provide a quantitative description of the roughness of the radar image. Using these techniques, we can detect targets quickly and clearly discriminate between sea ice, sea water, and targets.

A multiple classifier system can be a powerful solution for robust pattern recognition. It is expected that the appropriate combination of multiple classifiers may reduce errors, provide robustness, and achieve higher performance. In this paper, high performance Chinese syllable recognition is presented using combinations of multiple classifiers. Chinese syllable recognition is divided into base syllable recognition (disregarding the tones) and recognition of 4 tones. For base syllable recognition, we used a combination of two multisegment vector quantization (MSVQ) classifiers based on different features (instantaneous and transitional features of speech). For tone recognition, vector quantization (VQ) classifier was first used, and was comparable to multilayer perceptron (MLP) classifier. To get robust or better performance, a combination of distortion-based classifier (VQ) and discriminant-based classifier (MLP) is proposed. The evaluations have been carried out using standard syllable database CRDB in China, and experimental results have shown that combination of multiple classifiers with different features or different methodologies can improve recognition performance. Recognition accuracy for base syllable, tone, and tonal syllable is 96.79%, 99.82% and 96.24% respectively. Since these results were evaluated on a standard database, they can be used as a benchmark that allows direct comparison against other approaches.

A algorithm for designing a pattern classifier, which uses MDL criterion and a binary data structure, is proposed. The algorithm gives a partitioning of the range of the multi-dimensional attribute and gives an estimated probability model for this partitioning. The volume of bins in this partitioning is upper bounded by ο((log N/N)K/(K+2)) almost surely, where N is the length of training sequence and K is the dimension of the attribute. The convergence rates of the code length and the divergence of the estimated model are asymptotically upper bounded by ο((log N/N)2/(K+2)). The classification error is asymptotically upper bounded by ο((log N/N)1/(K+2)). Simulation results for 1-dimensional and 2-dimensional attribute cases show that the algorithm is practically efficient.

A learning algorithm is presented for nearest neighbor pattern classifiers for the cases where mixed supervised and unsupervised training data are given. The classification rule includes rejection of outlier patterns and fuzzy classification. This partially supervised learning problem is formulated as a multiobjective program which reduces to purely super-vised case when all training data are supervised or to the other extreme of fully unsupervised one when all data are unsupervised. The learning, i. e. the solution process of this program is performed with a gradient method for searching a saddle point of the Lagrange function of the program.

The main problem in statistical pattern recognition is to design a classifier. Many researchers point out that a finite number of training samples causes the practical difficulties and constraints in designing a classifier. However, very little is known about the performance of a classifier in small training sample size situations. In this paper, we compare the classification performance of the well-known classifiers (k-NN, Parzen, Fisher's linear, Quadratic, Modified quadratic, Euclidean distance classifiers) when the number of training samples is small.

A method for the recognition of Arabic printed scripts entered from an image scanner is presented. The method uses the Hough transformation (HT) to extract features, Dynamic programming (DP) matching technique, and a topological classifier to recognize the characters. A process of characters recognition is further divided into four parts: preprocessing, segmentation of a word into characters, features extraction, and characters identification. The preprocessing consists of the following steps: smoothing to remove noise, baseline drift correction by using HT, and lines separation by making an horizontal projection profile. After preprocessing, Arabic printed words are segmented into characters by analysing the vertical and the horizontal projection profiles using a threshold. The character or stroke obtained from the segmentation process is normalized in size, then thinned to provide it skeleton from which features are extracted. As in the procedure of straight lines detection, a threshold is applied to every cell and those cells whose count is greater than the threshold are selected. The coordinates (R, θ) of the selected cells are the extracted features. Next, characters are classified in two steps: In the first one, the character main body is classified using DP matching technique, and features selected in the HT space. In the second one, simple topological features extracted from the geometry of the stress marks are used by the topological classifier to completely recognize the characters. The topological features used to classify each type of the stress mark are the width, the height, and the number of black pixels of the stress marks. Knowing both the main group of the character body and the type of the stress mark (if any), the character is completely identified.