Abstract: A system and method are disclosed for determining the pose angle of an object in an input image. In a preferred embodiment, the present system comprises a pose estimator having a prototype projector, a regression estimator, and an angle calculator. The prototype projector is preferably adapted to reduce the input image dimensionality for faster further processing by projecting the input pixels of the image onto a Self-Organizing Map (SOM) neural network. The regression estimator is preferably implemented as a neural network and adapted to map the projections to a pattern unique to each pose. The angle calculator preferably includes a curve fitter and an error analyzer. The curve fitter is preferably adapted to estimate the pose angle from the mapping pattern. The error analyzer is preferably adapted to produce a confidence signal representing the likelihood of the input image being a face at the calculated pose.

Abstract: A new neural model for direct classification, DC, is introduced for acoustic/pictorial data compression. It is based on the Adaptive Resonance Theorem and Kohonen Self Organizing Feature Map neural models. In the adaptive training of the DC model, an input data file is vectorized into a domain of same size vector subunits. The result of the training (step 10 to 34) is to cluster the input vector domain into classes of similar subunits, and develop a center of mass called a centroid for each class to be stored in a codebook (CB) table. In the compression process, which is parallel to the training (step 33), for each input subunit, we obtain the index of the closest centroid in the CB. All indices and the CB will form the compressed file, CF. In the decompression phase (steps 42 to 52), for each index in the CF, a lookup process is performed into the CB to obtain the centroid representative of the original subunit. The obtained centroid is placed in the decompressed file.

Abstract: A new neural model for direct classification, DC, is introduced for acoustic/pictorial data compression. It is based on the Adaptive Resonance Theorem and Kohonen Self Organizing Feature Map neural models. In the adaptive training of the DC model, an input data file is vectorized into a domain of same size vector subunits. The result of the training (step 10 to 34) is to cluster the input vector domain into classes of similar subunits, and develop a center of mass called a centroid for each class to be stored in a codebook (CB) table. In the compression process, which is parallel to the training (step 33), for each input subunit, we obtain the index of the closest centroid in the CB. All indices and the CB will form the compressed file, CF. In the decompression phase (steps 42 to 52), for each index in the CF, a lookup process is performed into the CB to obtain the centroid representative of the original subunit. The obtained centroid is placed in the decompressed file.

Abstract: A divide-and-conquer (DAC) method and system improve the detection of abnormalities, like lung nodules, in radiological images via the use of zone-based digital image processing and artificial neural networks. The DAC method and system divide the lung zone into different zones in order to enhance the efficiency in detection. Different image enhancement techniques are used for each different zone to enhance nodule images, as are different zone-specific techniques for selecting suspected abnormalities, extracting image features corresponding to selected abnormalities, and classifying the abnormalities as either true or false abnormalities.

Abstract: An image processing apparatus and method and a provision medium arranged to perform image learning and image recognition in a short time. An image difference detector computes a differential image between an image stored in a frame buffer and an image stored in another frame buffer, and also computes the centroid of the differential image. A information collector forms RGB histogram data and binary data of a peripheral area about the centroid obtained by the image difference detector. A category former formed by a Kohonen network forms a category based on the RGB histogram data and binary data. A category statistic processor performs statistical processing of the categories output from the category former, and outputs a processing result to a learner formed by a recurrent neural network.

Abstract: An image recognizing apparatus and method is provided for recognizing behavior of a mobile unit accurately with an image of external environment acquired during the mobile unit is moving.

Abstract: A system for performing image-based diagnosis of a machine includes a database containing a plurality of historical images taken from a plurality of machines, a diagnostic unit configured to diagnose a new artifact image from the machine and to communicate historical and non-historical images or data associated with the system to a remote facility. The plurality of historical images include a plurality of ideal images generated from the plurality of machines using all possible machine settings and a plurality of artifact images generated from the plurality of machines, each of the artifact images having known faults associated therewith and a corresponding corrective action for repairing the faults. The diagnostic unit includes a diagnostic image processor and a diagnostic fault isolator.

Abstract: An expert system and software method for image recognition optimized for the repeating patterns characteristic of organic material. The method is performed by computing parameters across a two dimensional grid of pixels (rather than a one dimensional scan) with intensity values for each pixel having precision of eight significant bits. The parameters are fed to multiple neural networks, one for each parameter, which were each trained with images showing the tissue, structure, or nucleus to be recognized and trained with images likely to be presented that do not include the material to be recognized. Each neural network then outputs a measure of similarity of the unknown material to the known material on which the network was trained. The outputs of the multiple neural networks are aggregated by an associative voting matrix. A sub-neural network is used for each identified mode of data degradation in the input data.

Abstract: A neural network has reduced requirements for storing intermodal weight values, as a result of a dual-precision training process. In the forward propagation of training samples, low-resolution weight values are employed. During back-propagation of errors to train the network, higher-resolution values are used. After training, only the lower resolution values need to be stored for further run-time operation, thereby reducing memory requirements.

Abstract: This invention provides a system and method for integrating a plurality of diagnostic information from a multiple of sources. In this invention, there is a site specific database which contains site specific information for a machine. A plurality of diagnostic related information is obtained from the machine. A diagnostic router collects the site specific information for the machine and the plurality of diagnostic related information and generates a current incident record therefrom. An approved incident record database contains a plurality of approved incident records obtained from a plurality of machines. An integrator finds approved incident records from the approved incident record database that most closely match the current incident record.

Abstract: A neural network pattern recognition system for remotely sensing and identifying chemical and biological materials having a software component having an adaptive gradient descent training algorithm capable of performing backward-error-propagation and an input layer that is formatted to accept differential absorption Mueller matrix spectroscopic data, a filtering weight matrix component capable of filtering pattern recognition from Mueller data for specific predetermined materials and a processing component capable of receiving the pattern recognition from the filtering weight matrix component and determining the presence of specific predetermined materials. A method for sensing and identifying chemical and biological materials also is disclosed.

Abstract: An apparatus and methods for path identification in a communication network, including a plurality of network elements, the plurality of network elements further including a first network element, for issuing a first signal requesting a path identification operation, and at least one second network element, for receiving the first signal and for issuing a second signal. The second signal contains information of one of network configuration, network topology and distance between the first network element and the at least one second network element.

Abstract: A method, apparatus, and article of manufacture employing lexicon reduction using key characters and a neural network, for recognizing a line of cursive text. Unambiguous parts of a cursive image, referred to as “key characters,” are identified. If the level of confidence that a segment of a line of cursive text is a particular character is higher than a threshold, and is also sufficiently higher than the level of confidence of neighboring segments, then the character is designated as a key character candidate. Key character candidates are then screened using geometric information. The key character candidates that pass the screening are designated key characters. Two-stages of lexicon reduction are employed. The first stage of lexicon reduction uses a neural network to estimate a lower bound and an upper bound of the number of characters in a line of cursive text. Lexicon entries having a total number of characters outside of the bounds are eliminated.

Abstract: The present invention relates to a hierarchical artificial neural network (HANN) for automating the recognition and identification of patterns in data matrices. It has particular, although not exclusive, application to the identification of severe storm events (SSEs) from spatial precipitation patterns, derived from conventional volumetric radar imagery. To identify characteristic features a data matrix, the data matrix is processed with a self organizing network to produce a self organizing feature space mapping. The self organizing feature space mapping is processed to produce a density characterization of the feature space mapping. The self organizing network is preferably completely unsupervised. It may, under some circumstances include a supervised layer, but it must include at least an unsupervised component for the purposes of the invention. The “self organizing feature space” is intended to include any map with the self organizing characteristics of the Kohonen Self Organizing Feature Map.

Abstract: In a method for a self-organizing neural network for representing multidimensional, nonlinear imaging functions onto simpler imaging functions use divider-membranes are employed for achieving an error free representation of the imaging function via the learning sample, allowing for a high level of generalization. Kohonen cell borders coincide with a required imaging function. The neural network can independently determine a number of neurons necessary for an error-free solution of a problem. A readout of the neural network can occur through the calculation of the minimum of the squares of the distances.

Abstract: A classification tree which allows direct recoginition of an input pattern such as image or sound without extra processing such as pre-processing of unprocessed pattern data having high order characteristic variables is prepared. Information processing method and apparatus conduct hierarchical pre-processing for hierarchically pre-processing a learning pattern, prepares a classification tree based on the learning pattern processed by the hierarchical pre-processing and conducts the recognition by using the classification tree.

Abstract: A method and apparatus enhance visible contrast within an acquired image for display. The contrast enhancement utilizes all N bands of an original N-band spectral image to produce an M-dimensional enhanced image for display. The method creates an enhanced image from an original image in which the visible contrast in the original image is improved. The original image includes pixels, each pixel having N spectral intensities. The display or printer device which must display the relevant information may be limited to a number of bands M which is smaller than N. Maximum contrast of objects is obtained by emphasizing differences in the N-dimensional pixels by as large differences as possible within the dynamic range of the M-band display space. When M=N=3, this means moving pixels in the display space to utilize the full color palette available on a color monitor or printer. When N>M, a mapping from N space to M space must also be accomplished.

Abstract: A method for recognizing an object image comprises the steps of extracting a candidate for a predetermined object image from an image, and making a judgment as to whether the extracted candidate for the predetermined object image is or is not the predetermined object image. The candidate for the predetermined object image is extracted by causing the center point of a view window, which has a predetermined size, to travel to the position of the candidate for the predetermined object image, and determining an extraction area in accordance with the size and/or the shape of the candidate for the predetermined object image, the center point of the view window being taken as a reference during the determination of the extraction area.

Abstract: There is disclosed a pattern identifying neural network comprising at least an input and an output layer, the output layer having a plurality of principal nodes, each principal node trained to recognize a different class of patterns, and at least one fuzzy node trained to recognize all classes of patterns recognized by the principal nodes but with outputs set out at levels lower than the corresponding outputs of the principal nodes.

Abstract: The present invention discloses a system and method for performing image-based diagnosis. In this invention, historical artifact images and corresponding actions for repairing the artifacts are acquired and stored in a database. The database of historical artifact images and corresponding actions is used to diagnose an incoming artifact image having an unknown fault.

Abstract: An image processing apparatus using a neural network having: an image supplying unit for supplying spatiotemporal data of a predetermined region including a target pixel of an image; and a neural network formed by coupling a plurality of artificial neuron models so as to have at least an input layer, a hidden layer, and an output layer, wherein in the output layer, an input/output converting process is executed by a linear function and data corresponding to a target pixel is outputted from the output layer.

Abstract: Detection of video shot boundaries using a Video Segmenting Hidden Markov Model to model the sequence of states of a video. The Video Segmenting Hidden Markov Model determines the state sequence based on feature values. Using Hidden Markov Model techniques allows for automatic learning and use of multiple features including motion vectors, audio differences and histogram differences, without the need for manual adjustments of these thresholds.

Abstract: A pattern recognition device having modifiable feature detectors (28) which respond to a transduced input signal (26) and communicate a feature activity signal (30) to allow classification and an appropriate output action (70). A memory (40) stores a set of comparison patterns, and is used by an assigner (66) to find likely features, or parts, in the current input signal (26). Each part is assigned to a feature detector (28[m]) judged to be responsible for it. An updater (42) modifies each responsible feature detector (28[m]) so as to make its preferred feature more similar to its assigned part. The modification embodies a strong constraint on the feature learning process, in particular an assumption that the ideal features for describing the pattern domain occur independently. This constraint allows improved learning speed and potentially improved scaling properties.

Abstract: An electronic device for performing convolution operations comprises shift registers for receiving binary input values representative of an original matrix, synapses for storing weights correlated with a mask matrix, and neurons for outputting a binary result dependent on the sum of the binary values weighted by the synapses. Each synapse has a conductance correlated with the weight stored and dependent upon the binary input value. Each neuron generates the binary result in dependence on the total conductance of the corresponding synapses.

Abstract: A digital image management system is described that includes a content analyzer that analyzes an image to extract content data from the image. The content data of an image include face feature data. The digital image management system also includes an image database that is coupled to the content analyzer to store pixel data of each of a number of images and the content data of each of the images. A search engine is also provided in the digital image management system. The search engine is coupled to the image database and the content analyzer to compare the content data of the images with that of an input image such that any image similar to the input image can be identified from the image database without retrieving the pixel data of the image from the image database. A method of extracting feature data of a face in an image is also described.

Abstract: A neural network based improving the performance of an omni-font classifier by using recognized characters for additional training is presented. The invention applies the outputs of the hidden layer nodes of the neural net as the feature vector. Characters that are recognized with high confidence are used to dynamically train a secondary classifier. After the secondary classifier is trained, it is combined with the original main classifier. The invention can re-adjust the partition or boundary of feature space, based on on-line learning, by utilizing the secondary classifier data to form an alternative partition location. The new partition can be referred to when a character conflict exists during character recognition.

Abstract: A statistical classifier utilizes combined output values to determine posterior probabilities for certain output classes. In the field of handwriting recognition, compound characters are factored into classes of base letter forms and diacritical marks. A separate output activation value is produced for each base letter form and each diacritical mark. Pairs of output values, comprised of one value for a base letter form and one value for a diacritical mark, are combined to produce a posterior probability for every possible compound character, without requiring a network output for each possible class.

Abstract: A vector normalizing apparatus which normalizes an input vector or weight vectors by L.sub.2 -norm and adds norm information to the vector before it is normalized. Vector data from an input vector acquiring device (1) is displayed on an input vector display device (2) and transmitted through a vector transmitting device (3) to a normalized vector output device (4). During the transmission of the vector data, L.sub.2 -norm of the vector data is calculated by an L.sub.2 -norm calculating device (5) including a vector component squaring device (7) for squaring each of the components of the displayed vector, a summation device (8) for calculating a total sum of the squared components, and a square-root calculating device (9) for calculating L.sub.2 -norm by obtaining a square root of the result of the summation. The input vector is normalized by a vector component adjusting device (6) using the calculated value of L.sub.2 -norm. Thus, the input vector normalized by the L.sub.

Abstract: A method of estimating the chromaticity of illumination of a colored image consisting of a plurality of color-encoded pixels. The image colors are first mapped into an intensity-independent chromaticity space which is then divided into a plurality of separate regions. For each region, a first binary value is assigned to the region if the region contains no chromaticity value; or, a second binary value is assigned to the region if it does contain a chromaticity value. The assigned values are then applied as inputs to a pre-trained neural network having two output ports and at least one intermediate layer containing a plurality rality of ports connectible between selected input ports and the output ports. The chromaticity space values which characterize the input image's chromaticity of illumination are then derived at the output ports. The network is pretrained trained by initially connecting an arbitrary number of the intermediate layer ports to selected input layer ports.

Abstract: To prevent overfitting a neural network to a finite set of training samples, random distortions are dynamically applied to the samples each time they are applied to the network during a training session. A plurality of different types of distortions can be applied, which are randomly selected each time a sample is applied to the network. Alternatively, a combination of two or more types of distortion can be applied each time, with the amount of distortion being randomly varied for each type.

Abstract: A system for automatically detecting and recognizing the identity of a deformable object such as a human face, within an arbitrary image scene. The system comprises an object detector implemented as a probabilistic DBNN, for determining whether the object is within the arbitrary image scene and a feature localizer also implemented as a probabilistic DBNN, for determining the position of an identifying feature on the object such as the eyes. A feature extractor is coupled to the feature localizer and receives coordinates sent from the feature localizer which are indicative of the position of the identifying feature and also extracts from the coordinates information relating to other features of the object such as the eyebrows and nose, which are used to create a low resolution image of the object. A probabilistic DBNN based object recognizer for determining the identity of the object receives the low resolution image of the object inputted from the feature extractor to identify the object.

Abstract: An improved pattern recognition system. The invention operates on a plurality of feature vectors from a single class of data samples. The inventive system estimates a pruning radius for the feature vectors in the single class of data samples and generates a replacement class therefrom based on the estimated pruning radius. This pruning radius is used to train a classifier which in turn facilitates the recognition of a data pattern in raw data. In a specific implementation, the pruning radius is adapted based on current results from the classifier. The invention satisfies the need in the art by providing an automated technique for training classifiers for nonstationary data classes which is not limited by the need for more than one class of data.

Abstract: A statistical classifier for pattern recognition, such as a neural network, produces a plurality of output signals corresponding to the probabilities that a given input pattern belongs in respective classes. The classifier is trained in a manner such that low probabilities which pertain to classes of interest are not suppressed too greatly. This is achieved by modifying the amount by which error signals, corresponding to classes which are incorrectly identified, are employed in the training process, relative to error signals corresponding to the correct class. As a result, output probabilities for incorrect classes are not forced to a low value as much as probabilities for correct classes are raised.

Abstract: A statistical classifier that can be used for pattern recognition is trained to recognize negative, or improper patterns as well as proper patterns that are positively associated with desired output classes. A set of training samples includes both the negative and positive patterns, and target output values for the negative patterns are set so that no recognized class is indicated. The negative patterns are selected for training with less frequency than the positive patterns, and their effect on training is also modified, so that training is focused more heavily on improper patterns.

Abstract: A statistical classifier is trained in a manner to remove biasing due to unequal frequencies of unigram priors. The relative frequencies of all classes in a training set of sample patterns is determined. Training patterns are then selected from the set and skipped or repeated in dependence upon the relative frequency of the class to which they belong. In this manner, the presentation of samples is balanced across the classes.

Abstract: A statistical classifier that can be used for pattern recognition is trained to recognize negative, or improper patterns as well as proper patterns that are positively associated with desired output classes. A set of training samples includes both the negative and positive patterns, and target output values for the negative patterns are set so that no recognized class is indicated. The negative patterns are selected for training with less frequency than the positive patterns, and their effect on training is also modified, so that training is focused more heavily on positive patterns.

Abstract: A solder paste brick inspection and physical quality scoring system 10 employs a neural network 70 trained with a fuzzified output vector. An image of solder paste bricks 64 on a printed circuit board 12 is acquired by a CCD camera 30. Values of a predetermined set of brick metrics are extracted from the image by a computer 28 and used as a crisp input vector to trained neural network 70. A defuzzifier 76 converts a fuzzy output vector from neural network 70 into a crisp quality score output which can be used for monitoring and process control.

Abstract: A neural network based optical character recognition technique is presented for identifying characters in a moving web. Image acquisition means defines an imaging window through which the moving web passes such that the characters printed thereon can be imaged. Classification data is extracted and accumulated for each printed web character passing through the imaging window. A light source provides transmissive illumination of the web as it is being imaged. A neural network accelerator is coupled to the image acquisition means for intelligent processing of the accumulated classification data to produce therefrom printed character classification information indicative of each corresponding character imaged. A processor is coupled to the accelerator for converting the classification information into the appropriate ASCII character code. The technique is particularly useful for reading dot-matrix-type characters on a noisy, semi-transparent background at fast real-time rates.

Abstract: A technique for automatically training a set of character templates using unsegmented training samples uses as input a two-dimensional (2D) image of characters, called glyphs, as the source of training samples, a transcription associated with the 2D image as a source of labels for the glyph samples, and an explicit, formal 2D image source model that models as a grammar the structural and functional features of a set of 2D images that may be used as the source of training data. The input transcription may be a literal transcription associated with the 2D input image, or it may be nonliteral, for example containing logical structure tags for document formatting, such as found in markup languages. The technique uses spatial positioning information about the 2D image modeled by the 2D image source model and uses labels in the transcription to determine labeled glyph positions in the 2D image that identify locations of glyph samples.

Abstract: Pattern recognition, for instance optical character recognition, is achieved by training a neural network, scanning an image, segmenting the image to detect a pattern, preprocessing the detected pattern, and applying the preprocessed detected pattern to the trained neural network. The preprocessing includes determining a centroid of the pattern and centrally positioning the centroid in a frame containing the pattern. The training of the neural network includes randomly displacing template patterns within frames before applying the template patterns to the neural network.

Abstract: In a method for extracting an object image an extraction area for extraction of a candidate for a predetermined object image from an image is determined. The center point of a view window, which has a predetermined size, is caused to travel to the position of the candidate for the predetermined object image. The extraction area is determined in accordance with the size and/or the shape of the candidate for the predetermined object image, the center point of the view window being taken as a reference during the determination of the extraction area. The extraction of the candidate for the predetermined object image is carried out by using a neural network. Even if a plurality of object images, which are to be extracted, are embedded in a given image, the object images are extracted efficiently such that an object image, which has already been extracted, may not be extracted again.

Abstract: In a method for extracting an object image, an extraction area for extraction of a candidate for a predetermined object image from an image is determined. The center point of a view window, which has a predetermined size, is caused to travel to the position of the candidate for the predetermined object image. The extraction area is determined in accordance with the size and/or the shape of the candidate for the predetermined object image, the center point of the view window being taken as a reference during the determination of the extraction area. The extraction of the candidate for the predetermined object image is carried out by using a neural network. Even if a plurality of object images, which are to be extracted, are embedded in a given image, the object images are extracted efficiently such that an object image, which has already been extracted, may not be extracted again.

Abstract: In a recognition/diagnosis method, a relation between total sums of inputs of respective cases and teacher data is listed in the order of magnitude of the total sums of inputs. Based on the value of the teacher data for the input having the maximum total sum and the number of times of change of teacher data in the table are considered, a configuration of a neural network (the number of hidden layers and the number of neurons thereof) is determined. Coupling weights are analytically calculated based on the table.

Abstract: A self-organizing adaptive replicated (SOAR) for creating a replicate of human expert behavior. The SOAR can be embedded invisibly within multiple types of systems to observe, adapt and grow to emulate a user's interactive behavior and performance level. The system yields near equivalent responses to near equivalent stimuli in real time. The SOAR is based on a three layer perceptron type architecture which guarantees arbitrary M to N mapping of continuous valued spaces. The architecture uses a competitive, additive, and layer independent learning rule which insures excellent rapid learning. A self-organizing, adaptive algorithm permits the SOAR to adapt to the true classification space. The SOAR has applications in areas such a speech recognition, target detection, pattern recognition of multi-feature data, electro-mechanical subsystem control and resource allocation and optimization.

Abstract: An unsupervised back propagation method for training neural networks. For a set of inputs, target outputs are assigned l's and O's randomly or arbitrarily for a small number of outputs. The learning process is initiated and the convergence of outputs towards targets is monitored. At intervals, the learning is paused, and the values for those targets for the outputs which are converging at a less than average rate, are changed (e.g., 0.fwdarw.1, or 1.fwdarw.0), and the learning is then resumed with the new targets. The process is continuously iterated and the outputs converge on a stable classification, thereby providing unsupervised back propagation. In a further embodiment, samples classified with the trained network may serve as the training sets for additional subdivisions to grow additional layers of a hierarchical classification tree which converges to indivisible branch tips. After training is completed, such a tree may be used to classify new unlabelled samples with high efficiency.

Abstract: A technique for fingerprint classification and/or identification, in which a fingerprint is defined by areas containing patterns of ridges and valleys. At least one local pattern is determined using locations and characterizations of the fingerprint, which are indicated by a rapid change in direction of the ridges and valleys. The fingerprint is classified into types based upon the relative locations and characterizations of said local pattern(s). The fingerprint identification process can utilize minutiae location and angles as well as local pattern characterizations. Neural networks are utilized in determining the local patterns. The amount of data required to store data defining the fingerprints using the local pattern and/or minutiae techniques is significantly reduced.

Abstract: A control method of controlling a controlled system according to the invention comprises the first step of inputting a current and future target controlled variable to a first neural network model which performs learning using a past target controlled variable for the controlled system as an input signal and a past manipulated variable as a teacher signal, thereby obtaining a current virtual manipulated variable, the second step of causing a second neural network model, which have learnt to predict a behavior of the controlled system, to receive the virtual manipulated variable obtained in the first step and a controlled variable obtained from the controlled system at a current time, thereby obtaining a predicted controlled variable, the third step of obtaining an error of the predicted controlled variable obtained in the second step with respect to the target controlled variable, the fourth step of obtaining a correction amount for the virtual manipulated variable in accordance with a back propagation calcul

Abstract: A method and apparatus for determining whether a user of a system is an authorized user or an imposter by examining the keystroke characteristics of the user. The authorized user initially enters a number of user training samples on a keyboard. The user training samples are then purified to eliminate training samples which are different from other training samples. The purification can be performed by a self-organizing neural network which has input thereto, authorized user training samples, or both authorized training samples and imposter training samples. The purified user training samples are then compared to a sample to be tested to determine whether the sample is from an authorized user or an imposter. The comparison of the purified samples with the sample to be tested can be performed by a neural network such as a back propagation trained network, an ADALINE unit, a distance method or a linear classifier, discriminate function, or piecewise linear classifier.

Abstract: A pattern recognition system capable of handling the non-Gaussian feature vector distribution in the feature vector space such that the recognition rate can be improved while using the pattern matching based on the assumption of the Gaussian feature vector distribution. In the system, a nonlinear transformation for each recognizable category is applied to the feature vector extracted from the input pattern to be recognized, to obtain a transformed feature data for the input pattern, where the nonlinear transformation maps linearly inseparable distributions in a vector space containing the feature vector onto linearly separable distributions. Then, the transformed feature data for the input pattern is compared with reference feature model for each recognizable category indicating a feature vector distribution for each category, to find a category of the input pattern as that of the reference feature model closest to the transformed feature data for the input pattern.

Abstract: A signature recognition apparatus reduces the volume of training data needed and shortens the learning period. In the apparatus, a sample generating section generates sample data. A coupling load coefficient is determined based on the sample data, thereby obviating the need for additional sample data. The apparatus also uses a fuzzy net which implements a linear function in its output layer to shorten the learning period relative to the learning period required for a net implementing a non-linear function such as a sigmoid.