(Return to ICNN97 Homepage) (Return to ICNN'97 Agenda)
ICNN'97 FINAL ABSTRACTS
PR: PATTERN RECOGNITION & IMAGE PROCESSING
ICNN97 Pattern Recognition & Image Processing Session: PR1A Paper Number: 611 Oral
Accelerating back propagation in human face recognition
D.J. Evans, M.H. Ahmad Fadzil and Z. Zainuddin
Keywords: back propagation human face recognition convergence rate
Abstract:
Standard back propagation, as with many gradient based optimisation methods converges slowly as neural network training problems become larger and more complex. This paper describes the employment of two algorithms to accelerate the training procedure in anautomatic human face recognition system. As compared to standard back propagation, the convergence rate is improved by up to 98% with only a minimal increase in the complexity of each iteration.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR1B Paper Number: 374 Oral
Gender classification of human faces using hybrid classifier systems
Srinivas Gutta and Harry Wechsler
Keywords: human face classification gender FERET database
Abstract:
This paper considers a hybrid classification architectures for gender classification of human faces and shows its feasibility using a collection of 2000 face images from the FERET data base (corresponding to 700 male and 300 female subjects). The hybrid approach consists of an ensemble of RBF networks and inductive decision trees (DT). Specifically Cross Validation (CV) experimental results yield an average accuracy rate of 94 % for the hybrid architecture consisting of ensemble of RBF networks ('Model 2') and decision trees ('C4.5'). The benefits of our hybrid architecture, beyond the high accuracy achieved, include (i) robustness via query by consensus provided by the ensembles of RBF networks, and (ii) flexible and adaptive thresholds as opposed to ad hoc and hard thresholds provided by DT.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR1C Paper Number: 670 Oral
Hybrid approaches to frontal view face recognition using the neural network
Kang Sik Yoon, Young Kug Ham and Rae-Hong Park
Keywords: Hybrid approaches frontal view face recognition artificial neural network
Abstract:
Keywords: Face recognition, hidden Markov model (HMM), HMM-NN, NN-HMM
In this paper, for frontal view recognition hybrid approaches using the neural network (NN) and hidden Markov model (HMM) are proposed. In the preprocessing stage, edges of a face are detected using the conventional locally adaptive threshold (LAT) scheme and facial features are extracted based on generic knowledge of facial components. In constructing a database with normalized features, we employ HMM parameters of each person computed by the forward-backward algorithm. Computer simulation shows that the proposed HMM-NN algorithm yields higher recognition rate compared with several conventional face recognition algorithms.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR1D Paper Number: 69 Oral
Automatic facial feature extraction by applying genetic algorithms
Chun-Hung Lin and Ja-Ling Wu
Keywords: Feature extraction Genetic Algorithms Estimation region growing
Abstract:
An automatic facial feature extraction algorithm is presented in this paper. The algorithm is composed of two main stages: the face region estimation stage and the feature extraction stage. In the face region estimation stage, a second-chance region growing method is adopted to estimate the face region of a target image. In the feature extraction stage, genetic search algorithms are applied to extract the facial feature points within the face region. It is shown by simulation results that the proposed algorithm can automatically and exactly extract facial features with limited computational complexity.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR1E Paper Number: 239 Oral
A principal component based BDNN for face recognition
L. J. Shen and H. C. Fu
Keywords: principal component BDNN face recognition
In this paper we propose a high performance two-stage hybrid structure for face recognition. The first stage is an eigenface based recognizer, which serves as a candidate faces selector. As our experience, the Top 1 recognition rate is only 65%, however the Top 10 hit rate can be up to 98.15%. The Top 10 candidate faces are similar to each other, thus these faces are called simial faces. Since the projections of the similar faces are too close in the eigenspace, it's very hard to distinguish a target face from similar face set. Thus, we propose the "Horizontal Average Gray Scale (HAGS)" as a new type of feature for the second stage recognizer. A paired-Bayesian-decision neural network (pBDNN) is used for the second stage recognizer, which identifies the target from the similar faces.
Supporting by the proposed feature, a pDBNN could make an accurate classification between any two similar faces. In order to demonstrate the proposed hybrid system, we conducted some experiments on an in house database, which contains 675 images taken from 135 people. The training data for the pBDNN were small orientation (-22.5, 0, 22.5 degrees), and the large orientation (-45 and 45 degrees) images were for testing. Our experimental results show that the hybrid recognition structure improvs the recognition rate for 17% more than the eigenface system alone (65%) without any rejection, and 26% more with 31% of rejection.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR2A Paper Number: 296 Oral
Two-stage neural network for volume segmentation of medical images
Mohamed N. Ahmed and Aly A. Farag
Keywords: neural network volume segmentation medical images
Abstract:
In this paper, we present a new system to segment and label CT/MRI Brain slices using feature extraction and unsupervised clustering. In this technique, each voxel is assigned a feature pattern consisting of a scaled family of differential geometrical invariant features. The invariant feature pattern is then assigned to a specific region using a two-stage neural network system. The first stage is a self-organizing principal components analysis (SOPCA) network that is used to project the feature vector onto its leading principal axes found by using principal components analysis. This step provides an effective basis for feature extraction. The second stage consists of a self-organizing feature map (SOFM) which will automatically cluster the input vector into different regions. The optimum number of regions (clusters) is obtained by a model fitting approach. Finally, a 3D connected component labeling algorithm is applied to ensure region connectivity. Implementation and performance of this technique are presented. Compared to other approaches, the new system is more accurate in extracting 3D anatomical structures of the brain, and can be adapted to real-time imaging scenarios.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR2B Paper Number: 236 Oral
An RBF based classifier for detection of microcalcifications in mammograms with outlier rejection capability
A. Hojjatoleslami, L. Sardo and J. Kittler
Keywords: RBF classifier outlier mammograms
Abstract:
The results of a study carried out on a large database of mammographic images using an RBF network for density estimation are presented. The classifier is built via the Bayes rule from an estimate of the class conditional probability density functions. The aim is the detection of microcalcifications. Though the recognition rate must be high, a minimum number of false alarms should also be attained. The results obtained using a MLP neural network, K-NN and Gaussian classifiers are also presented for comparison. The ROC curve for image identification demonstrates a superior performance for the RBF classifier where less than 15% of normal images were misclassified for 100% abnormal images identification. A simple outlier detection mechanism has also been examined, which has shown to be useful in flagging data acquisition errors or ambiguous cases also requiring medical attention.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR2C Paper Number: 598 Oral
Tissue characterisation with NMR spectroscopy: current state and future prospects for the application of neural networks analysis
Paulo J.G. Lisboa, Neil M. Branston, Wael El-Deredy and Alfredo Vellido
Keywords: Tumour detection NMR spectroscopy artificial neural network
Abstract:
Nuclear Magnetic Resonance (NMR) Spectroscopy has considerable potential for non-invasive characterisation of tissue biochemistry and the diagnosis of tissue abnormalities, ranging from focal lesions in the brain, to tumours in any area of the body to assessing effect of HIV damage. However, the realisation of the full clinical potential NMR spectroscopy will depend on extracting information from the spectra directly and specifically related to the biochemistry of different tissue types under various normal and pathological circumstances. This paper reviews the progress made in the application of neural network analysis to the automatic characterisation of NMR data, raising some key issues and providing a perspective of the future of this technology.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR2D Paper Number: 667 Oral
Recognition of handwritten digits using structural information
Sven Behnke, Marcus Pfister and Raul Rojas
Keywords: handwritten digit recognition structural information feature extraction
Abstract:
This article presents an off-line method for recognizing handwritten digits. Structural information and quantitative features are extracted from images of isolated numerals to be classified by a hybrid multi-stage recognition system. Feature extraction starts with the raw pixel-image and derives more structured representations like line-drawings and attributed structural graphs. Classification is done in two steps: a) the structural graph is matched to prototypes, b) for each prototype there is a neural classifier which has been trained to distinguish digits represented by the same graph-structure. The performance of the described system is evaluated on two large databases (provided by SIEMENS AG and NIST) and is compared to other systems. Finally, the combination of the described system and a TDNN classifier is discussed. The experimental results indicate that there is an advantage in using structural information to enhance an unstructured neural classifier.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR2E Paper Number: 541 Oral
A neuro-fuzzy approach to recognize arabic handwritten characters
Adel M. Alimi
Keywords: neuro-fuzzy arabic handwritten characters cursive handwriting fuzzy membership
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PR2F Paper Number: 575 Oral
An empirical evaluation of bagging and boosting for artificial neural networks
David W. Opitz and Richard Maclin
Keywords: bagging boosting artificial neural networks
Abstract:
Bagging (Breiman 1996a) and Boosting (Freund & Schapire 1996) are two relatively new but popular methods for producing classifier ensembles.
An ensemble consists a set of independently trained classifiers (such as neural networks or decision trees) whose predictions are combined when classifying instances. Previous research suggests that an ensemble as a whole is often more accurate than any of the single classifiers in the ensemble. In this paper we evaluate Bagging and Boosting as methods for creating an ensemble of neural networks. We also include results from Quinlan's (1996) decision tree evaluation of these methods. Our results indicate that the ensemble methods can indeed produce very accurate classifiers for some datasets, but that these gains may depend on aspects of the dataset. In particular, we find that Bagging is probably appropriate for most problems, but when properly applied, Boosting may produce even larger gains in accuracy.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR3A Paper Number: 457 Oral
Evolutionary radial basis neural networks
Nicolaos B. Karayiannis and Weiqun Mi
Keywords: Evolutionary radial basis neural networks hybrid learning pattern classification
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PR3B Paper Number: 655 Oral
Determination of the number of components in Gaussian mixtures based on agglomerative clustering
Swarup Medasani and Raghu Krishnapuram
Keywords: Gaussian mixtures agglomerative clustering expectation maximization number of components
Abstract:
Modeling data sets by mixtures is a common technique in many pattern recognition applications. The Expectation Maximization (EM) algorithm for mixture decomposition suffers from the disadvantage that the number of components in the mxiture needs to be specified. In this paper, we propose a new objective function, the minimum of which gives the number of components automatically. The proposed method, known as the Agglomerative Gaussian Mixture Decomposition algorithm, is then used to determine the number of hidden nodes in a radial basis function network. We present results on real data sets which indicate that the proposed method is not sensitive to initialization and gives better classification rates.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR3C Paper Number: 272 Oral
Soft-competitive-growing classifier with unsupervised fine-tuning
Jose L. Alba, Laura Docio, and Simon Ruibal
Keywords: EM algorithm Paramter estimation Classification
Abstract:
In this paper a method for growing a Gaussian-mixture-based network is developed. The constructive technique is based on an EM algorithm to estimate the parameters and the number of nodes is iteratively increased by means of discriminant placement. The growth control is imposed by an information theoretic criterion that prevents the network from becoming extremely complex and loosing generalization capabilities. After the growing phase is finished, another EM algorithm is used with labeled and unlabeled data in order to fine-tune network parameters. This solution improves the test-performance for the applications where labeled data is insufficient and the classes are not highly overlaped. We report results on some artificially generated examples and on terrain classification over a Landsat-TM image.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR3D Paper Number: 77 Oral
Parallel Stochastic grammar induction
Stefan C. Kremer
Keywords: stochastic grammar induction recurrent netorks
Abstract:
This paper examines the problem of sto\-chastic grammar induction and gives a formal analysis of observed limitations of a classical algorithm. It then describes a parallel approach to the problem which avoids these limitations. Finally a proof is presented which shows that a popular training algorithm already in use for recurrent connectionist networks implements the new approach.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR3E Paper Number: 61 Oral
Combining statistical pattern recognition approach with neural networks for recognition of large-set categories
Yoshimasa Kimura, Toru Wakahara, and Kazumi Odaka
Keywords: Statistical Pattern recognition Classification character recognition
Abstract:
We present a two-stage hierarchical system consisting of a statistical pattern recognition (SPR) module and artificial neural network (ANN) to recognize a large number of categories including similar category sets. In the first stage, the SPR module performs classification. If the first candidate does not belong to a pre-determined similar category set, the first candidate is accepted as the final result; other wise, the first candidate is sent to the ANN module. In the second stage, ANN performs classification for similar categories to select correct a candidate from the pre-determined candidate set designated by the first candidate. The new scheme offers improved system performance by sharing tasks between SPR and ANN according to the degree of classification difficulty and forming specialized ANNs for each similar category. The system achieves higher performance for the recognition of 3,201 handprinted characters than a traditional system constructed with just the SPR module.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR3F Paper Number: 354 Oral
Possibilistic fuzzy classification using neural networks
Hisao Ishibuchi and Manabu Nii
Keywords: pattern classification fuzzy classification possibility analysis reject option
Abstract:
In this paper, we examine the performance of a possibilistic fuzzy classification method where the possibility area of each class is identified by the learning of a multilayer feedforward neural network. Our method does not always assign an input pattern to a single class, because the possibility areas of different classes may overlap one another in the pattern space. First, we illustrate our possibilistic fuzzy classification method. Next we examine its performance by computer simulations on real-world test problems. Then we discuss the relation between our method and the reject option. Finally we extend our method to the case where a rejection penalty is explicitly given in classification problems.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR4A Paper Number: 111 Oral
Unsupervised Hierarchical fingerprint matching
A. Murat Ozbayoglu and Cihan H. Dagli
Keywords: Hierarchical classification model identification fingerprint matching unsupervised classification
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PR4B Paper Number: 312 Oral
Structural adaptation in neural networks with application to land mine detection
Sassan Sheedvash and Mahmood R. Azimi-Sadjadi
Keywords: Structural adaptation land mine detection orthogonal projection
Abstract:
This paper presents a new approach for structural adaptation in multi-layer neural networks in general and the application of the proposed method to land mine target detection and classification problem. The new algorithm uses time and order update formulations of the orthogonal projection theorem to derive a recursive weight updating procedure and architectural variation of the network during the training process. The proposed approach provides optimal network structure in the sense that the mean-squared error is minimized for the newly created topology. This algorithm is used in conjunction with a data representation scheme to perform land mine target detection and classification. The simulation results on targets with different compositions indicated superior detection and classification performance when compared to the conventional methods.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR4C Paper Number: 291 Oral
A visual multi-expert neural classifier
Chester Ornes and Jack Sklansky
Keywords: multi-expert neural classifier automatic pattern classifier dimensionality reduction
Abstract:
We describe a high-performance neural network that, in addition to classifying a query, allows the user to visualize the relationship between the query and the data in the training set. We show in applications, including medical diagnosis and image segmentation, that our classifier achieves low error rates while providing a visual explanation for classifier decisions. We demonstrate the properties of the classifier using synthetic data, and compare the performance of the visual neural classifier to Kohonen's self-organizing map.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR4D Paper Number: 562 Oral
Indexing for object recognition using large neural networks
Mark R. Stevens, Charles W. Anderson and J. Ross Beveridge
Keywords: Efficient indexing object recognition template matching
Abstract:
Template matching is an effective means of locating vehicles in outdoor scenes, but it tends to be a computationally expensive. To reduce processing time, we use large neural networks to predict, or index, a small subset of templates that are likely to match each window in an image. Results on actual LADAR range images show that limiting the templates to those selected by the neural networks reduces the computation time by a factor of 5 without sacrificing the accuracy of the results.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR4E Paper Number: 626 Oral
ARTMAP-FD: Familiarity discrimination applied to radar target recognition
Gail A. Carpenter, Mark A. Rubin and William W. Streilein
Keywords: ARTMAP-FD Familiarity discrimination radar target recognition
Abstract:
ARTMAP-FD extends fuzzy ARTMAP to perform familiarity discrimination. That is, the network learns to abstain from meaningless guesses on patterns not belonging to a class represented in the training set. ARTMAP-FD can also be applied in conjunction with sequential evidence accumulation. Its performance is illustrated here on simulated radar range profile data.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR4F Paper Number: 622 Oral
Multiscale image factorization
John L. Johnson, Mary Lou Padgett and William A. Friday
Keywords: Multiscale image factorization hierarchical image decomposition noise
Abstract:
A new hierarchical image decomposition is described which resolves an image into a set of image product factors. The set is ordered in scale from coarse to fine image detail. When the factored set is multiplied together it reproduces the original image. By selecting factors, coarse scene elements such as shadows, and fine scene factors such as noise, can be isolated. The scale of detail is controlled by the linking strength of a pulse coupled neural network, on which the system is based.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR5A Paper Number: 24 Oral
Soft nearest neighbor classification
Yoram Baram
Keywords: Soft classification relative separator domain of indecision
Abstract:
It is shown how soft classification, which allows for the creation of indecision domains near given separation surfaces between two classes, applies to the nearest neighbor method, and how the optimal size of the indecision domain can be found from the training data. The performance of the soft nearest neighbor classifier is compared to that of the conventional classifier using stock trading data.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR5B Paper Number: 357 Oral
A scalable method for classifier knowledge reuse
K. Bollacker and J. Ghosh
Keywords: classifier learning maximum likelihood
Abstract:
Just as a person's life-long experience helps him/her in novel tasks, it would be useful to leverage the knowledge in previously trained classifiers in learning future classification tasks that may be related. We present a maximum posterior probability method for classifier knowledge reuse that is novel in its scalability with the quantity of classifiers reused and in its ability to incorporate different classifier architectures. Also, we describe a mutual information based relevance criterion to identify previously trained classifiers that may help in the current task. Results from application of this method and criterion to public domain data sets demonstrate their usefulness in improving classifier performance, speeding up learning, and assisting in problem decomposition.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR5C Paper Number: 461 Oral
some further studies on detection number of clusters
Yiu-Ming Cheung and Lei Xu
Keywords: number of clusters unsupervised learning clustering
Abstract:
To determine number of clusters in the unsupervised learning, a criterion based on the Bayesian-Kullback YING-YANG Machine Learning Scheme has recently been proposed (Xu, 1995), and has been proved by Theorem 1 of the paper (Xu, 1996b). As the condition presented in this theorem is hard to be directly checked in practice, in this paper we give one equivalent condition, which is easier to be examined for a given specific problem. Furthermore, since the implementation of the criterion requests a large quantity of computing costs to estimate parameters for each candidate cluster number k, here we also propose a heuristic algorithm for those data from mixture populations with the same priori probabilities and equal- and-isotropic variance, which can make the number of k as small as possible by selecting reasonable k's instead of one by one to test. Preliminary experiments have shown that our proposed algorithm can save computing costs considerably.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR5D Paper Number: 325 Oral
Feature extraction from wavelet coefficients for pattern recognition tasks
Stefan Pittner and Sagar V. Kamarthi
Keywords: Feature extraction wavelet coefficients pattern recognition
Abstract:
This paper deals with the assessment of the value of process parameters from the wavelet coefficients of a measured process signal. Since a direct assessment from all wavelet coefficients will often turn out to be tedious or leads to inaccurate results, a preprocessing routine that computes robust features directly correlated to the process parameters is highly desirable. In this paper, a new efficient feature extraction method based on the fast wavelet transform is presented. This method divides the matrix of computed wavelet coefficients into clusters equal to rowvectors. The important frequency ranges have a larger number of clusters than the less important frequency ranges. The features of a process signal are provided by the euclidean norms of each such vector. The effectiveness of this new method has been verified on a flank wear estimation problem in turning processes.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR5E Paper Number: 411 Oral
Fuzzy image enhancement and associative feature matching in radiotherapy
G. Krell, H. R. Tizhoosh, T. Lilienblum, C. J. Moore and B. Michaelis
Keywords: Fuzzy image enhancement associative feature matching radiotherapy
Abstract:
In radiotherapy, the images of the Electronic Portal Imaging Device serve to verification of shape and location of the therapy beam with respect to the patient's anatomy during radiation. The desired patient-beam relation is given by a simulator image that is captured for treatment planning. Because of the acquisition physics the unprocessed Electronic Portal Images are poor in quality compared to the simulator images. The conventional EPI allows only a rough verification of patient position relative to bony structures. This paper presents an approach that combines a Fuzzy Image Enhancement technique with a linear Associative Memory. The main idea is the inclusion of additional knowledge for the restoration of the Electronic Portal Image to allow a more reliable feature extraction. The Fuzzy Image Enhancement uses expert knowledge of the physician. The Associative Memory matches corresponding structures in electronic portal and simulator image. Firstly, the images are enhanced by the Fuzzy Image Enhancement. For the training of the Associative Memory, the enhanced and modified simulator image is used for learning data generation. The associative memory is then recalled by the warped and enhanced Electronic Portal Image. In this way, an image with a much higher quality than from conventional solutions is obtained.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR5F Paper Number: 594 Oral
A fault tolerant chinese bank check recognition system based on SOM neural networks
Wang Song, Ma Feng and Xia Shaowei
Keywords: chinese bank check recognition system, SOM neural networks, Learning Vector Quantization
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PR6A Paper Number: 203 Oral
JPEG post processing image enhancement
Alex Lopez, Angel Torres and Shawn Hunt
Keywords: Image Enhancement Image filtering JPEG image compression
Abstract:
A nonlinear filter based on the feedforward Neural Network topology is presented. This study was undertaken to investigate the usefulness of 'smart' filters in image post processing. The filter has shown to be useful in recovering high frequencies, such as those lost during the JPEG compression-decompression process. The filtered images have a higher signal to noise ratio, and a higher perceived image quality. Simulation studies showing examples of the high frequency recovery, and the statistical properties of the filter are given.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR6B Paper Number: 446 Oral
Improved dynamic bit allocation in image coding using a self-organizing map with learning vector quantization
Joao Souza Neto, Sebastiao do Nascimento and Francisco Assis de O. Nascimento
Keywords: dynamic bit allocation image coding self-organizing map learning vector quantization
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PR6C Paper Number: 472 Oral
A neural implementation of interpolation with a family of kernels
Frank M. Candocia and Jose C. Principe
Keywords: interpolation kernels SOFM
Abstract:
A paradigm for interpolating images based on a family of kernels is presented. Each kernel is "tuned" to specific image characteristics and contains the information responsible for the local creation of missing detail. This interpolation process (1) exploits the correlation that exists in the local structure of images via a self-organizing feature map (SOFM) and (2) establishes an optimal set of linear associative memories (LAM's) from the homologous neighborhoods of a set of low and high resolution image counterparts. Each LAM creates members of the family of interpolation kernels. We compare the performance of this technique with the commonly used bilinear and spline interpolation methods and demonstrate its ability to generalize well.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR6D Paper Number: 222 Oral
Automatic extraction of drainage networks from digital terrain elevation data: a local network approach
A. Fern, M. T. Musavi, and J. Miranda
Keywords: drainage networks digital terrain elevation local network
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PR6E Paper Number: 385 Oral
Low bit rate image compression with orthogonal projection pursuit neural networks
S. R. Safavian, Hamid R. Rabiee, M. Fardanesh, and R. L. Kashyap
Keywords: Low bit rate image compression orthogonal projection pursuit projection pursuit neural networks
Abstract:
A new multiresolution algorithm for image compression based on projection pursuit neural networks is presented. High quality low bit-rate image compression is achieved first by segmenting an image into regions of different sizes based on perceptual variation in each region and then constructing a distinct code for each block by using the orthogonal projection pursuit neural networks. This algorithm allows one to adaptively construct a better approximation for each block by optimally selecting the basis functions from a universal set. The convergence is guaranteed by orthogonalizing the selected bases at each iteration. The coefficients of the approximations are obtained by back-projection with convex combinations. Our experimental results shows that at rates below 0.5 bits/pixel, this algorithm shows excellent performance both in terms of PSNR and subjective image quality.
_____
ICNN97 Pattern Recognition & Image Processing Session: PR6F Paper Number: 47 Oral
A Neural network incorporating adaptive gabor filters for image texture classification
Keisuke Kameyama, Kenzou Mori, and Yukio Kosugi
Keywords: Gabor filters texture classification back propagation training
Abstract:
A novel neural network architecture for image texture classification is introduced. The proposed Kernel Modifying Neural Network (KM Net) which incorporates a convolution filter kernel array and a classifier in one, enables an automated texture feature extraction in the multichannel texture classification through simultaneous modification of the kernels and the connection weights by a backpropagation-based training rule. The first layer units working as the convolution kernels are constrained to be an array of Gabor filters, which achieves a most efficient texture feature localization.
The following layers work as a classifier of the extracted texture feature vectors. The capability of the KM Net and its training rule is verified with basic texture classification problems on synthetic and fabric texture images, and also with a biological tissue classification problem in an ultrasonic echo scan image.
_____
ICNN97 Learning & Memory Session: PRP2 Paper Number: 591 Poster
Resolutionable cellular neural networks
Mamoru Tanaka, Kenya Jin'no, Jun'ichi Miyata, Masaaki Imaizumi, Toshiaki Shingu and Hiroshi Inoue
Keywords: Resolutionable cellular neural networks spatio-temporal dynamics RGB color
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 2 Poster
Face recognition system based on neural networks and fuzzy logic
Ahmad Fadzil M. H. and Lim Cheah Choon
Keywords: Face recognition neural networks fuzzy logic
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 163 Poster
Competitive neural networks as adaptive algorithms for non-stationary clustering: experimental results on the color quantization of image sequences
A. I. Gonzalez and M. Grana
Keywords: Competitive neural networks non-stationary clustering color quantization
Abstract:
In this paper we consider the application of several architectures of Competitive Neural Networks to the adaptive computation of cluster representatives (codevectors) over non-stationary data. Adaptive computation shifts the emphasis from robust global optimization to fast local optimization from good initial conditions. The paradigm of non-stationary Clustering is represented by the problem of Color Quantization of image sequences. Experimental results applying the diverse architectures to the adaptive computation of color representatives for the Color Quantization of an image sequence are given and discussed.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 194 Poster
Evolutionary artificial neural networks for competitive learning
A.D. Brown and H. C. Card
Keywords: Evolutionary artificial neural networks competitive learning genetic algorithm
Abstract:
We present experiments which show that a genetic algorithm (GA) can effectively search for a set of local feature detectors, which can be used by higher neural network layers to perform an image classification task. Three different methods of encoding hidden unit weights into the GA are presented, including one which coevolves all the feature detectors in a single chromosome, and two which promote the cooperation of feature detectors by encoding them in their own chromosome. The fitness function measures the classification percentage and confidence of the networks. The three algorithms are all capable of finding a set of feature detectors which allow for 100 percent classification performance, but a novel variant of the cooperative method produces the most consistent, highest confidence classifiers.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 255 Poster
Identifying contact formations from force signals: A comparison of fuzzy and neural network classifiers
Marjorie Skubic, Shawn P. Castrianni, and Richard A. Volz
Keywords: Fuzzy logic Classification Membership function generation
Abstract:
In this paper, we present and compare two methods of identifying single-ended contact formations from force sensor patterns. Instead of using geometric models of the workpieces, both methods use force sensor signals only. In the first method, fuzzy logic is used to model the patterns in the force signals. Membership functions are generated automatically from training data and then used by the fuzzy classifier. In the second method, a neural network architecture is used to learn the mapping from force signals to contact formation class. Experimental results are presented for both the fuzzy and neural network classifiers, and the results are compared. In some cases, the fuzzy classifier has better performance, and in other cases, the neural net classifier is better. The results are discussed, and, finally, a training modification is presented which dramatically improves the performance of the inadequate neural net classifiers.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 125 Poster
Genetic algorithms for clustering, feature selection and classification
Lin Yu Tseng and Shiueng Bien Yang
Keywords: Genetic Algorithms Clustering Feature selection Classification
Abstract:
In solving clustering problem, traditional methods, for example, the K-means algorithm and its variants, usually ask the user to provide the number of clusters. Unfortunately, the number of clusters in general is unknown to the user. Therefore, the clustering becomes a tedious trial-and-error work and the clustering result is often not very promising especially when the number of clusters is large. In this paper, we propose a genetic algorithm for the clustering problem. This algorithm can automatically cluster the data according to their similarities and automatically find the proper number of clusters. We also apply the genetic algorithm to the classification problem and obtain good results. Another genetic algorithm is proposed for the feature selection problem. This algorithm can not only search for a good set of features but also find the weight of each feature such that the application of these features associated with their weights to the classification problem will achieve a good classification rate. Experimental results are given to illustrate the effectiveness of these genetic algorithms.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 176 Poster
A knowledge-based approach for fault detection and isolation in analog circuits
Mohamed A. El-Gamal
Keywords: Fault detection rule based connectionist NN Fault isolation
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 198 Poster
Image retrieval System capable of learning the user's sensibility using neural networks
Yoshiteru Yageyama and Hideo Saito
Keywords: Image retreival Learning
Abstract:
With the advent of the multimedia era, the needs to get the image that an user wants from a lot of images is going to be more concerned about. In this paper, we propose an interactive image retrieval system which employs back-propagation neural networks using the words that represents the user's sensibility, in order to deal with the user's ambiguous queries. When an user inputs the words , this system sets the synapse of the network which represents both the user and the word and displays candidate images according to the output values of the neural network. The user evaluates the similarity to the image that he wants to get until the system displays the optimal images, and the system produces the set of teach signals according to the user's evaluation. After training the network, the system displays new candidate images. The inputs of the neural network are image features which has one-to-one correspondence with images in the databases. We implemented this system on Sun SPARC station, and could see that the system could improve the candidate images each time an user evaluate them.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 335 Poster
Adapting the 2-class recursive deterministic perceptron neural network to m-classes
Mohamed Tajine, David Elizondo, Emile Fiesler and Jerzy Korczak
Keywords: 2-class recursive deterministic perceptron ne linear separability 2-valued RDP
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 671 Poster
Rough fuzzy set theoretic approach to evaluate the importance of input features in classification
Manish Sarkar and B. Yegnanarayana
Keywords: Rough fuzzy sets input features classification
Abstract:
Artificial neural networks are currently employed to capture the reasoning process involved in bidding a hand in a Contract Bridge game. Input hand in a Bridge game is conveniently represented as a series of 52 one/zero, where presence or absence of a card is denoted by 1 or 0. In this input representation all the cards, that are present in a hand, receive equal importance. Since the class discriminatory property of all the cards are not same to classify an input hand, the representation of each input pattern however should be biased based on the importance of each card. This necessitates a way to measure the importance of each card, i.e. feature, individually. While measuring importance of a particular feature, influence of the other features, player's experience etc. are not possible to taken into account; hence measurement is in fact based on an incomplete knowledge. The notion of rough set can be effectively exploited to determine the importance of each feature from this incomplete knowledge. Moreover, the classification task involved in bidding is inherently fuzzy. Hence, in this paper a rough-fuzzy set based measure is proposed to evaluate the importance of each feature. The efficacy of the proposed scheme is demonstrated by some experimental results.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 464 Poster
An experimental comparative study on several soft and hard-cut EM algorithms for mixture of experts
Wing-kai Lam, Fai Yung and Lei Xu
Keywords: hard-cut EM algorithms soft-cut EM algorithms EM algorithm mixture of experts
Abstract:
Mixture of Expert (ME) (Jacobs, Jordan and Nowlan, 1991) and EM EM algorithms are very popular in supervised learning. Recently, an alternative ME model (Xu, Jordan and Hinton, 1995) and a number of Hard-cut EM algorithms for both original and alternative ME (Xu, 1996) are proposed by one of the present authors. In this paper, we ry to conduct a systematic experimental comparison on the two models through their implementation in soft and hard-cut EM algorithms. The comparison is based on the aspects of (1) the number of converged experiments with satisfactory results, (2) the classification correctness, (3) the training and testing error and, (4) time required. Experimental results obtained illustrate that the soft and hard-cut EM algorithms for the alternative ME have the highest percentage of convergence and classification correctness, much smaller training and testing error when compared with those algorithms for the original ME. Moreover, it requires much fewer number of iteration for the alternative ME to converged than that for the original ME.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 242 Poster
Texture segmentation using gaussian markov random fields and LEGION
Erdogan Cesmeli and DeLiang Wang
Keywords: Texture segmentation gaussian markov random fields LEGION
Abstract:
An image segmentation method is proposed for texture analysis. The method is composed of two main parts. The first part determines a novel set of texture features based on Gaussian Markov Random Field (GMRF). Unlike other GMRF-based methods, our method is not limited by a fixed set of texture types. The second part is LEGION (Locally Excitatory Globally Inhibitory Oscillator Networks) which is a 2D array of neural oscillators. The coupling strengths between neighboring oscillators are calculated based on texture feature differences. When LEGION is simulated, the oscillators corresponding to the same texture tend to oscillate in synchrony, whereas different texture regions tend to attain different phases. Results demonstrating the success of our method on real texture images are provided.
Keywords: image segmentation, texture, markov random field, artificial neural networks, LEGION
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 50 Poster
Theoretical and experimental analyses of restoring degraded images based on contenues hopfield networks
Lei Wang, Feihu Qi and Yulong Mo
Keywords: image restoration Hopfield Neural Networks
Abstract:
This paper proposes a modified full parallel self-feedback continuous Hopfield neural network model to restore degraded images. Theoretical analyses show that this model is able to ensure its energy converging to the global minimum more precisely , therefore good restored images are obtained. The result of this model on restoring uniform velocity linear motion-blurred images is compared with J.K.Paik's method. Experimental results indicate that the SNR(signal-to-noise ratio) of the images restored from this model are improved obviously and the visual quality of them are quite good.
Keywords: continuous Hopfield neural network, image restoration, uniform velocity linear motion
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 421 Poster
A comparative study of layered neural networks on misclassification in pattern recognition
Kenichi Takahashi
Keywords: layered neural networks misclassification pattern recognition
Abstract:
When an unknown pattern is the input to a layered neural network, the neural network may classify erroneously the unknown pattern into one of training classes, depending on the similarity between an input pattern and training patterns. In this paper, four neural network models for reducing misclassification are considered, and their performance is compared with that of the basic layered network model.
Each of these four models has more output units for increasing redundancy than the basic model has, while the number of units in the input layer and the number of units in the hidden layer for the five models are kept constant. In computer simulations, random patterns and mosaic face images are used to examine the performance of the five models. It is shown through the computer simulations that two models are effective in reducing misclassification.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 554 Poster
Blind deconvolution by self-organization
Wen-Pin tai, Ruei-Sung Lin and Cheng-Yuan Liou
Keywords: Blind deconvolution self-organization blurred images
Abstract:
In this work, we devise a self-organizing network to solve both the unknown system and unknown input in blind deconvolution of blurred images. We utilize a criterion function which has a similar form as the Kullback-Leibler cross information formula to adapt the network's weights to approach the unknown system function.
This adaptation gradually reduces the criterion value which is a distance measure between the system output and the output of the adapted system with a reconstructed input signal. The weight matrices of the neurons in the network will be shifted versions of the system function and will be aligned in the network according to their shifts during convergence. This is because that the convolution operation copes with this network scheme and the hidden topology of the shifted system functions can be aligned similarly in a 2D plane.
Keywords: self-organizing network, blind deconvolution, blurred image, Kullback-Leibler information criterion.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 302 Poster
A neural network classifier for conflicting information environments
Pu Sun and Kenneth Marko
Keywords: neural network classifier conflicting information environments non-convergence phenomena
Abstract:
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 581 Poster
Non-stationary texture segmentation using an AM-FM model
M.S. Pattichis, C. Christodoulou, C.S. Pattichis and A.C. Bovik
Keywords: Non-stationary texture segmentation AM-FM model Learning vector quantization
Abstract:
We present a novel method for segmenting non-stationary textures. Our approach uses a multi-dimensional AM-FM representation for the texture, and provides the FM features to an SOFM-LVQ neural network system that performs the segmentation. For the segmentation, we use the eigenvalues of the instantaneous frequency gradient tensor, and show how these eigenvalues capture the non-stationary structure of a texture. For a woodgrain image, the segmentation results are shown to capture the essential non-stationary nature of the grain.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 486 Poster
Invariant pattern recognition of 2D images using neural networks and frequency-domain representation
Fernando Cesar C. De Castro, Jose Nelson Amaral and Paulo Roberto G. Franco
Keywords: Invariant pattern recognition 2D images frequency-domain representation
Abstract:
Frequency domain representation of two dimensional gray-level images is used to develop a pattern recognition method that is invariant to rotation, translation and scaling. Frequency domain representation is a natural feature detector that allows the use of only few directions of highest energy as training data for a set of Artificial Neural Networks (ANNs). We developed a new algorithm that uses the spectral information stored in these ANNs to compare a given image with a known pattern, determining the relative translation between them and yielding a measure of their similarity. The representation and method we adopted has the advantage of leaving only the rotation of the object as a free parameter to be determined by the algorithm. We minimize the spectral resolution noise using Spectral Directional Filtering. Our experimental results indicate that the proposed method has excelent discriminating power.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 418 Poster
A committee of networks classifier with multi-resolution feature extraction for automatic target recognition
Lin-Cheng Wang, Sandor Der, and Nasser M. Nasrabadi
Keywords: Neural Net Classifier Automatic Target recognition FLIR imagery stacked generalization
Abstract:
A neural network-based classifier has been applied to the problem of automatic target recognition (ATR) using forward-looking infrared (FLIR) imagery. The target classifier consists of several neural networks that form a committee for classification. Each neural network in the committee receives inputs from features extracted from only a local region of a target, known as a receptive field, and is trained independently from other committee members. The classification results of the individual neural networks are combined to determine the final classification. Our experiments show that this committee of networks classifier is superior to a fully connected neural network classifier in terms of complexity (number of weights to be learned) and performance (classification rate). The proposed classifier shows a high noise immunity to clutter or target obscuration due to the independence of the individual neural networks in the committee. Performance of the proposed classifier is further improved by the use of multi-resolution features and by the introduction of a higher level neural network on the top of committee, a method known as stacked generalization.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 331 Poster
Evolutionary CT image reconstruction from a small number of projections
Zensho Nakao, Midori Takashibu and Yen-Wei Chen
Keywords: CT image genetic algorithms gray level image reconstruction from projections
Abstract:
An evolutionary algorithm for reconstructing CT gray images from a small number of projections is presented; the algorithm reconstructs two-dimensional unknown images from four one-dimensional projection data. A Laplacian constraint term is included in the fitness function of the genetic algorithm for handling smooth images more efficiently, and the evolutionary process reconstructs images into finer ones by partitioning the images gradually, thereby increasing the chromosome size exponentially as the generation proceeds. Results obtained are compared to those obtained by the well-known algebraic reconstruction technique (ART), and it was found that the evolutionary method is more effective thatn ART when the number of projection directions is very limited, based on both the projection error and pixel value error estimators.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 592 Poster
Fast inner-outer point evaluation in a generalization polytopic domain
Osvaldo Agamennoni and Pablo S. Mandolesi
Keywords: inner-outer point evaluation generalization polytopic domain interpolation
Abstract:
In this paper the generalization domain or applicability domain of a given model is addressed. The generalization domain is related with the interpolation domain that is usually defined through the polytope formed by the training data. Some considerations about the relationship between the interpolation and the generalization domains are given. A fast algorithm to tests in real time if a model input is inside or outside the interpolation domain is presented. This is a more general problem commonly encountered in many areas, i.e., check if a point is inside or not a given polytope. An example to evaluate points of a sphere from a closed ball is discussed to show the performance of the algorithm.
_____
ICNN97 Pattern Recognition & Image Processing Session: PRP2 Paper Number: 287 Poster
Neural trees for image segmentation
Iren Valova and Yukio Kosugi
Keywords: Neural trees image segmentation magnetic resonance images
Abstract:
In this paper we report the application of neural trees for image segmentation of magnetic resonance (MR) images. The network, built up during training, effectively partitions the feature space into subregions and each final subregion is assigned a class label according to the data routed to it. As the tree grows, the number of training data for each node decreases, which results in less weight update epochs and decreases the time consumption.
A key point in the proposed algorithm is choosing the right neuron to divide the data. A general training coefficient set for all tree nodes may offer poor division. A different training coefficient should be chosen in order to gain purity of classes. A set of candidate - neurons is installed and the winner is taken to fill the node space in our tree.
The network performance is compared to the multilayered perceptron (MLP) over the white/gray matter MRI segmentation problem.