Chair: Jenq-Neng Hwang, University of Washington (USA)
S. Garcia-Salicetti, Institut National des Telecommunications
P. Gallinari, LAFORIA-IBP UA CNRS
B. Dorizzi, Institut National des Telecommunications
A. Mellouk, LAFORIA-IBP
D. Fauchon, Institut National des Telecommunications (FRANCE)
We present a neural prediction system for on-line writer- independent character recognition as a first step towards a word recognition system. The input feature vectors contain the pen trajectory information, recorded by a digitizing tablet. Each letter is modeled by a variable number of predictive Neural Networks, depending on its length. Successive parts of a letter are modeled by different Multilayer Neural Networks, only transitions from each one to itself or to its right neighbors being permitted. To deal with the great variability of cursive handwriting, we introduce a holistic approach for both Learning and Recognition, combining Neural Networks and Dynamic Programming techniques. Our system is able to recognize strongly distorted and truncated letters, obtained by automatic segmentation of 10000 words from 10 different writers. Even on such databases, unappropriate to character recognition (letters in it were not recorded as handwritten isolated characters), quite good recognition rates are obtained.
Patricia Davies, Purdue University (USA)
Brian R. Silverstein, Purdue University (USA)
This paper addresses two types of problems which prove difficult for traditional classifiers: having very limited training data for at least one class, and having classes with a large amount of overlap. Issues discussed will include the 1) use of nearest neighbor methods and neural nets for classification of data which is completely inseparable by linear and quadratic classifiers, 2) dealing with training sets of unequal size from each class.
Ce Zhu, Shantou University
Jun Wang, Southeast University (PEOPLES REPUBLIC OF CHINA)
Taijun Wang, Southeast University (PEOPLES REPUBLIC OF CHINA)
Although the family of LVQ algorithms have been widely used for pattern classification and have achieved a great success, the rigirous theoretical studies on the classification performance of LVQ algorithms have seldom been made. In this paper, the asymptotical performance of LVQ1, LVQ2 and LVQ2.1 algorithms have been studied thoroughly, and three significant conclusions have been achieved respectively. Furthermore, a simple modification scheme to LVQ2 algorithm has been developed and analyzed on the asymptotical performance, which can produce the optimal or nearly-optimal classifier in the stable equilibrium state for the classification problems with classes overlapping.
M. Petroni, McGill University
A.S. Malowany, McGill University
C.C. Johnston, McGill University
B.J. Stevens, University of Toronto (CANADA)
The analysis of infant cry vocalizations has been the focus of a number of efforts over the past thirty years. Since the infant cry is one of the only means that an infant has for communicating with its care-giving environment, it is thought that information regarding the state of an infant, such as hunger or pain, can be determined from an infant's cry. To date, research groups have determined that adult listeners can differentiate between different types of cries auditorily, and at least one group has attempted to automate this classification process. This paper presents the results of another attempt at automating the discrimination process, this time using artificial neural networks (ANNs). The input data consists of successive frames of one of two parametric representations generated from the first second of a cry following the application of either an anger, fear, or pain stimulus. >From tests conducted to date, it is determined that ANNs are a useful tool for cry classification and merit further study in this domain.
Batuhan Ulu, Ohio State University
Stanley C. Ahalt, Ohio State University
Richard A. Mitchell, Wright Laboratories (USA)
In many model- based Automatic Target Recognition (ATR) systems the size of the model catalog can be a critical factor in determining the viability of the system. In this paper we examine an ATR system which uses synthetic High Range Resolution (HRR) Radar data to determine how classification performance is affected by the compression of the HRR model catalog. For this purpose the data is preprocessed, clustered and classified using Nearest Neighbor and Radial Basis Function (RBF) classifiers. The effect of compression on classification performance is examined through simulations for both of these classification schemes. For the data in question we show that significant (100:1 or greater) compression can be achieved with little degradation in classification performance.
Datong Wei, University of Michigan (USA)
Berkman Sahiner, University of Michigan (USA)
Heang-Ping Chan, University of Michigan (USA)
Nicholas Petrick, University of Michigan (USA)
A convolution neural network (CNN) was used for classification of masses and normal tissue on mammograms. A generalized CNN was developed that uses multiple images derived from a single region of interest (ROI) as input. CNN input images were obtained from the ROIs using (i) averaging and subsampling; and (ii) texture feature extraction methods on smaller sub-regions inside the ROI. In (ii), features computed over different sub-regions were arranged as texture-images, and subsequently used as inputs to the CNN. Results indicate that using texture-images improves classification accuracy.
Huang Deshuang, Beijing Institute of Technology (PEOPLES REPUBLIC OF CHINA)
Mao Erke, Beijing Institute of Technology (PEOPLES REPUBLIC OF CHINA)
Han Yueqiu, Beijing Institute of Technology (PEOPLES REPUBLIC OF CHINA)
This paper studies the mechanism for classification of Feedforward Neural Networks form the geometric viewpoints. It is pointed out that the MLPNs realize hyperplance divisions in the pattern space, and the FLN realize hypercurved divisions. We give a form of Generalized Function Link Nets (GFLN), and discuss the application of a special GFLN to recognition of radar targets, and give several experimental results.
K.P. Cohen, University of Wisconsin - Madison (USA)
Y.H. Hu, University of Wisconsin - Madison (USA)
W.J. Tompkins, University of Wisconsin - Madison (USA)
J.G. Webster, University of Wisconsin - Madison (USA)
We have developed and trained a Fuzzy Neural Network (FNN) to detect individual breaths using information from multiple independent noninvasive ventilation sensors. We derive input features from simultaneous recordings from impedance and inductance plethysmographs, and a pneumotachometer while healthy adults performed several different combinations of ventilation and motion. We first tested our FNN using membership functions, rules and consequent sets derived using a heuristic approach. Using all features, on 4 subjects we found that the average rate of combined false- positive and false-negative detections was 5.1%. When we trained our FNN using a gradient descent algorithm, the average rate of combined false-positive and false-negative detections was reduced to 2.6%.
Yuh-Fwu Guu, Southern Methodist University (USA)
Behrouz Peikari, Southern Methodist University (USA)
This paper presents a new logpolar sampling procedure for recognition of handwritten numerals. It is shown that this approach requires less computation than the logpolar sampling method employed by Duren and Peikari. Furthermore, in addition to the ability of transforming rotational variation to translational variation, it can also reduce scale variation. Logpolar sampling is used as a pre-processing stage in conjunction with various neural network structures. The results show that it can be used with a two layered sparsly connected neural network to obtain a better recognition rate than previous works. A normalization method based on boundary crossings is also introduced, it is shown that it requires even less computations than logpolar sampling and has the ability of reducing the deformation effect found in handwritten characters. Over 16500 characters are used in conducting the experiments, recognition rates of 96.24% and 95.91% are obtained using logpolar sampling and normalization method respectively.
Y. Cai, University of Windsor (CANADA)
H.K. Kwan, University of Windsor (CANADA)
In this paper, a fuzzy inference network (FIN) is proposed. The proposed FIN preserves the advantages of both fuzzy classification algorithm and neural networks. It can learn membership functions directly from training samples and classify patterns according to the membership values. An efficient self-organizing learning algorithm is also presented.