Abstract: A 2D image supplied from an image input unit including a wide view lens is sampled into a discrete form by an array sensor, and then mapped to a multi-resolution space by a 2D filter. The feature of the supplied image is detected, and then the mapped image is transformed to a local pattern about the detected feature, and then the coordinates of the position of the feature and the code word of the local pattern are formed into a set which is then encoded. The code is supplied to each cell of a stochastic automaton. The quantity of visual information is calculated in accordance with the quantity of mutual information between different cells of the stochastic automaton consisting of cells in blocks, the coordinates of the position of the feature and the distance from the feature to the optical axis so as to control the optical axis of the image input unit in such a manner that the quantity of visual information is maximized.

Abstract: An automated method, and a computer storage medium storing instructions for executing the method, for analysis of image features in lung nodule detection in a chest radiographic image represented by digital data, including preprocessing the image to identify candidate nodules in the image; establishing a region of interest (ROI) including a candidate nodule identified in the preprocessing step; performing image enhancement of the candidate nodule within the ROI; obtaining a histogram of accumulated edge gradients as a function of radial angles withing the ROI after performing the image enhancement; and determining whether the candidate nodule is a false positive based on the obtained histogram. A 64.times.64-pixel region of interest (ROI) centered at the candidate location is used. The contrast of the ROI is improved by a two-dimensional background subtraction. A nodule shape matched filter is used for enhancement of the nodular pattern located in the central area of the ROI.

Abstract: An image processing apparatus includes a feature data extracting circuit for detecting feature data indicative of density characteristics of a document based on image signals inputted from an input terminal, a density correction table selecting circuit composed of a neural circuit network which is learned beforehand so as to recognize image characteristics based on the feature data, and a density correcting circuit for selecting a density correction table in accordance with image characteristics based on a selection signal from the density correction table selecting circuit so that the density of image signals is corrected based on the density correction table. As a result, characteristics of the document are extracted, so that the density of the image signals can be corrected based thereon.

Abstract: A denomination recognition apparatus and method uses a programmed microelectronic CPU (14) to execute program instructions (19) stored in PROM (18) to read in pixel data from an optical imaging section including LEDs (15) and photodiodes (16) for generating signals which can be converted to a first image of a currency note (13) being transported along the path of travel. The CPU (14) receives position and skew data detected by external sensors for sensing the position and skew of the note (13).

Abstract: A signal processing apparatus and concomitant method for learning and using fidelity metric as a control mechanism and to process large quantities of fidelity metrics from a visual discrimination measure (VDM) to a manageable subjective image quality ratings. The signal processing apparatus incorporates a VDM and a neural network. The VDM receives input image sequences and generates fidelity metrics, which are received by a neural network. The neural network is trained to learn and use the fidelity metrics as a control mechanism, e.g., to control a video encoder.

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: 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 hybrid neural network system for object recognition exhibiting local image sampling, a self-organizing map neural network, and a hybrid convolutional neural network. The self-organizing map provides a quantization of the image samples into a topological space where inputs that are nearby in the original space are also nearby in the output space, thereby providing dimensionality reduction and invariance to minor changes in the image sample, and the hybrid convolutional neural network provides for partial invariance to translation, rotation, scale, and deformation. The hybrid convolutional network extracts successively larger features in a hierarchical set of layers. Alternative embodiments using the Karhunen-Loeve transform in place of the self-organizing map, and a multi-layer perceptron in place of the convolutional network are described.

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: Apparatus and method for determining a location and span of an object in an image. The determined location and span of the object are used to process the image to simplify a subsequent classification process.

Abstract: A character recognition system is described, in particular a system suitable for use in monitoring cargo container codes or vehicle number plates. An image of the code is first analyzed to extract potential characters. As part of this process, long horizontal and vertical line segments are filtered out. The extracted potential characters are then input to a two-level character recognition means. The first level comprises a neural network classifier that classifies a character into a smaller set of possible characters; and then the second level comprises another neural network classifier which identifies which character among the smaller set of possible characters the extracted character is.

Abstract: A method an system for using a local attention threshold to aid in the detection of clustered abnormalities in digitized medical images is disclosed. The local attention threshold is applied to locate spots within a predetermined distance from previously identified spots. More specifically, seed pixels are identified by applying a first seed threshold function to the output of a shift-invariant neural network and adaptive threshold. The seed pixels are then segmented into spots by applying a segmentation threshold function to each seed pixel. False-positive spots are removed using various techniques. Additional seed pixels are then identified by applying a local attention threshold to pixels within a predetermined distance to previously identified spots. The local attention threshold disclosed is less selective for pixels which are closer to the nearest spot than for pixels which are further from the nearest spot.

Abstract: A method for obtaining a representation of the textures of a geological structure, characterized in that images characteristic of the sedimentology of the environment are formed, parameters corresponding to the nature of the images are estimated at every point of each image and in a spatial domain around the point so as to determine a texture vector for each of the points and to obtain a set of texture vectors. The method also includes the steps of selecting texture vectors representative of the characteristic textures of the geological environment in the set of texture vectors; and using a neural network formed of cells distributed in two dimensions which contains as many cells as characteristic textures. The selected texture vectors are used to submit the neural network to a learning process so that a final topology map of the textures characteristic of the geological environment is obtained.

Abstract: A computerized method for the detection and characterization of disease in an image derived from a chest radiograph, wherein an image in the chest radiograph is processed to determine the ribcage boundary, including lung top edges, right and left ribcage edges, and right and left hemidiaphragm edges. Texture measures including RMS variations of pixel values within regions of interest are converted to relative exposures and corrected for system noise existing in the system used to produce the image. Texture and/or geometric pattern indices are produced. A histogram(s) of the produced index (indices) is produced and values of the histograms) are applied as inputs to a trained artificial neural network, which classifies the image as normal or abnormal.

Abstract: A method and system for detecting and displaying clustered microcalcifications in a digital mammogram, wherein a single digital mammogram is first automatically cropped to a breast area sub-image which is then processed by means of an optimized Difference of Gaussians filter to enhance the appearance of potential microcalcifications in the sub-image. The potential microcalcifications are thresholded, clusters are detected, features are computed for the detected clusters, and the clusters are classified as either suspicious or not suspicious by means of a neural network. Thresholding is preferably by sloping local thresholding, but may also be performed by global and dual-local thresholding. The locations in the original digital mammogram of the suspicious detected clustered microcalcifications are indicated. Parameters for use in the detection and thresholding portions of the system are computer-optimized by means of a genetic algorithm.

Abstract: This invention describes a technology to improve the feature based comparison of images. The images are captured and their significant features are extracted. For a comparison only the feature values have to be compared instead of the images themselves. This leads to a significant reduction of storage space and calculation time needed. The reliability of the comparison is improved greatly by including also the individual variation ranges of the feature values and using a specialized neural net for classification.

Abstract: A method of processing a digital image comprising the steps of:providing an input digital image having an input size;providing the output size of a desired output digital image;inputting the provided input digital image and the provided desired output digital image size to a weight estimation sub-system which operates to generate several weight derivative estimates;inputting the provided input digital image to an x-direction derivative estimation sub-system which produces an array of x-direction derivative estimates;inputting the provided input digital image, the weight derivative estimates and the x-direction estimates to an x-direction interpolated value estimation subsystem which produces in intermediate digital image which has the same number of rows as said input digital image while having the same number of columns as said output digital image;operating on said intermediate digital image by a y-direction derivative estimation sub-system to produce y-direction derivative estimates;inputting the intermedi

Abstract: A system for deriving final display parameters for a wide range of MR images consists of a feature generator utilizing both histogram and spatial information computed from an input MR image, a wavelet transform within the feature generator for compressing the size of the feature vector, a competitive layer based neural network for clustering MR images into different subclasses, a bi-modal linear estimation network and a radial bases function network based non-linear estimator for each subclass, as well as a data fusion system using estimates from both estimators to compute the final display parameters.

Abstract: Image content classification methods, systems and computer programs repeatedly scan an image having an array of image pixels, with at least one random neural network. Each scan corresponds to one of multiple texture patterns. A corresponding texture pattern is compared to each of multiple image portions for each of the multiple scans. A value is assigned to each image portion, corresponding to the texture pattern having the highest coincidence. An array of pixels corresponding to the assigned values for the image portions may then be displayed. Highly accurate results may be obtained, at high speed, without the need for lengthy expert analysis.

Abstract: An image processing apparatus includes an input unit for entering a plurality of color image signals, an image processing unit for subjecting the plurality of entered color image signals to processing based upon an algorithm of a cellular neural network, and a comparison decision unit for deciding output color data based upon results of processing by the image processing unit. Since input multivalued color image data based upon the algorithm of a neural network are converted to output color image data, it is possible to obtain a high-quality output color image that is faithful to the input color image.

Abstract: A regularization system and method for image restoration in homogeneous or inhomogeneous environments. The system and method includes features similar to a neural network with intermediate levels of structure including a pixel having processing capabilities; clusters consisting of a plurality of interconnected pixels and also having processing capabilities; and an image space comprised of a plurality of interconnected pixels and clusters and also having processing capabilities. The system and method also include means for assigning a regularization parameter to each pixel depending on the local variance of intensity of pixels; decomposing the image space into clusters of pixels, each cluster having the same regularization parameter; imposing a blurring function on each pixel; rapidly forming a regularized image by simultaneous local and global encoding of a regularization matrix onto each pixel directed through a process of gradient energy decent; and a means of assessing the output image.

Abstract: The present invention is a classification method and apparatus for classifying an input into one of a plurality of possible outputs. The invention is also a method and apparatus for adjusting a neuron encompassing a plurality of feature vectors. The invention characterizes the spatial distribution of the feature vectors. The invention then spatially adjusts the neuron in accordance with that characterization.

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: In automatic evaluation of cereal kernels or like granular products handled in bulk, the kernels are conveyed on a vibrating conveyor belt (15). Owing to the vibrations, the kernels are spread and settled in grooves (14) in the belt so as to be oriented in essentially the same direction. A video camera (40) produces digital images of all the kernels on the belt. The kernels are identified in the images, and for each kernel input signals are produced and then sent to a neural network based on picture element values for the picture elements representing each kernel. A neural network then determines which of a plurality of predetermined classes that each kernel belongs.

Abstract: Neural networks are used to classify objects automatically by means of Doppler-broadened radar echo signals. The classification device KK contains a neural network (NET, NET2) which has an input layer (IL) of input nodes (IN1, . . . , IN57) for features (M) of the Doppler-broadened radar echo signals, and an output layer (OL) of output nodes (ON1, ON2, ON3) for predetermined classes to which the objects can be allocated. The neural network (NET, NET2) is adapted to the external conditions prevailing at the time of the classification operation. The adaptation takes place either via accessible input nodes (ZN1, ZN2) into which control information (SI) can be entered, and which cause the neural network (NET) to adapt to one or to several external influence factors, or via a selection device (SEL) which, from the parameters (P1, . . .

Abstract: A color correcting unit receives color separation values such as CMY values from an image input unit. Under the control of a control portion, inputs to a first conversion portion constituted by a neural network which has been trained in advance on the basis of the spectral distribution of an arbitrary illuminant are corrected so that outputs from the first conversion portion satisfy the color separation values and a predetermined requirement. The input values to the first conversion portion, which satisfy the predetermined requirement, are sent to an image output unit. The image output unit outputs an image in accordance with these input values.

Abstract: To perform processing to increase the number of pixels of inputted image data, the inputted image data having low resolution is converted to image data having high resolution. For this conversion, one pixel of the low resolution data is interpolated, a difference value between a predetermined value and image data included in an interpolating-pixel block which corresponds to a pixel of interest of the low resolution data, is calculated, and pixel values of pixels in the interpolation-pixel block are determined in accordance with the difference value. Furthermore, multi-level image data on which pixel values have been determined is binarized by the dither method, to determine pixel values such that the smaller the absolute value of the difference value is, the larger the difference from the predetermined value.

Abstract: Apparatus, and an accompanying method, for use in, e.g., a neural network-based optical character recognition (OCR) system (5) for accurately classifying each individual character extracted from a string of characters, and specifically for generating a highly reliable confidence measure that would be used in deciding whether to accept or reject each classified character. Specifically, a confidence measure, associated with each output of, e.g., a neural classifier (165), is generated through use of all the neural activation output values. Each individual neural activation output provides information for a corresponding atomic hypothesis of an evidence function. This hypothesis is that a pattern belongs to a particular class. Each neural output is transformed (1650) through a pre-defined monotonic function into a degree of support in its associated evidence function.

Abstract: In a pattern recognition system provided with a pattern recognition processing unit having a network structure constructed of an input layer for inputting a feature parameter of a subject under recognition as input information an intermediate layer for processing the input information and an output layer for outputting a processed result output values of respective output nodes for constituting the output layer, corresponding to the input information, are compared with each other by the pattern recognition processing unit, and a classification item corresponding to the output node whose output value is maximum is stored into a storage unit as a recognized result with respect to the input information. The pattern recognition system is also provided with a reliability evaluating unit for setting a threshold with respect to an output value of each of the output nodes and for evaluating reliability of the recognized result based upon the output values of the respective output nodes.

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: A biometric recognition system involves two phases: creation of a master pattern set of authorized users biometric indicia and authentication using a classification neural network. To create the master pattern set, an image of an authorized biometric user's indicia is divided into a plurality of regions of interest or "features". The system determines which features are the most useful for identification purposes. Master patterns are then created from these master features, thus creating a master pattern set. During authentication, a sample pattern set of a user to be authenticated is similarly created. A neural network is used to compare the sample pattern set with the master pattern set to determine whether the user should be authenticated.

Abstract: An image processing apparatus inputs an image signal obtained by scanning a document in which a character region, a photographic region and a dot region are mixed and stores image data in a local block composed of a target picture element and plural picture elements surrounding the target picture element. The image processing apparatus computes first and second feature parameters P.sub.0 and P.sub.1 representing the features of each region based on the image data in the local block, and inputs the resulting first and second feature parameters P.sub.0 and P.sub.1 to an identification circuit adopting a neural network. The identification circuit outputs the region identification information O.sub.0 and O.sub.1 of the target picture element, and the filter processing circuit performs a spatial filtering process according to the region identification information O.sub.0 and O.sub.1.

Abstract: The present invention provides an apparatus and a method for classifying and recognizing image patterns using a second-order neural network, thereby achieving high-rate parallel processing while lowering the complexity. The second-order neural network, which is made of adders and multipliers, corrects positional translations generated in a complex-log mapping unit to output the same result for the same object irrespective of the scale and/or rotation of the object. The present invention enables high-rate image pattern classification and recognition based on parallel processing, which is the advantage obtained in neural network models, because consistent neural networks and consistent network structure computation models are applied to all steps from the image input step to the pattern classifying and recognizing step.

Abstract: A device for discriminating an attribute of an image in a block area contained in a document image includes a device for performing a Fourier transformation based on image data in the block area and for determining a spatial frequency spectrum relating to the image in the block area; and a neural network for outputting a discrimination result as to whether or not the attribute of the image in the block area is a halftone dot image based on the spatial frequency spectrum output from the Fourier transform device.

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: An automated computer-aided diagnosis (CAD) method and system using artificial neural networks (ANNs) for the quantitative analysis of image data. Three separate ANNs were applied for detection of interstitial disease on digitized two-dimensional chest images. The first ANN was trained with horizontal profiles in regions of interest (ROIs) selected from normal and abnormal chest radiographs. The second ANN was trained using vertical output patterns obtained from the 1.sup.st ANN for each ROI. The output value of the 2.sup.nd ANN was used to distinguish between normal and abnormal ROIS with interstitial infiltrates. If the ratio of the number of abnormal ROIs to the total number of all ROIs in a chest image was greater than a certain threshold level, the chest image was considered abnormal. In addition, the third ANN was applied to distinguish between normal and abnormal chest images where the chest image was not clearly normal or abnormal.

Abstract: An image processing apparatus includes an image input section, a color image/monochrome image converting section for performing image area division, a binarization circuit for binarizing a converted monochrome image, a reducing section for reducing a binary image, a boundary extracting section for extracting the boundaries between the areas of constituent elements constituting an input image, e.g., a binary image and a continuous gradation image, and a kind-of-image determining section for determining the kinds of images in partial areas defined by the extracted boundaries. For example, the image processing apparatus further includes a data compressing section.

Abstract: A pattern recognition system classifies images of patterns in which the definition of individual features of the pattern may have become blurred. The image is segmented into pieces of arbitrary size and shape, and various combinations are examined to determine those which represent the most likely segmentation of the pattern into its individual features. These individual features are then classified, according to known techniques. Through the use of a second order Markov model, not all possible combinations of pieces need to be examined, to determine the best ones. Rather, the examination of various combinations is limited in accordance with previously determined information, to thereby render the process more efficient. By combining multiple, independently determined probabilities, the accuracy of the overall operation is enhanced.

Abstract: An automated method and system for digital imaging processing of radiologic images, wherein digital image data is acquired and subjected to multiple phases of digital imaging processing. During the Pre-Processing stage, simultaneous box-rim filtering and k-nearest neighbor processing and subsequent global thresholding are performed on the image data to enhance object-to-background contrast, merge subclusters and determine gray scale thresholds for further processing. Next, during the Preliminary Selection phase, body part segmentation, morphological erosion processing, connected component analysis and image block segmentation occurs to subtract unwanted image data preliminarily identify potentials areas including abnormalities.

Abstract: A method for automatic merging two images having overlapped parts for a handy scanner. The method starts from defining a meaningful sample in a first image using fuzzy logic. Then, a two level searches based on A* algorithm are executed thereafter to find a match point in the second image. After the match point is found, the two images are stitched together using linear interpolation.

Abstract: An image reading unit picks up an image to convert the image into an electric signal. An A/D converter converts the electric signal into a digital signal, and an input image memory stores the digital signal. An image reduction processing unit reduces the image data stored in the input image memory. A reduced image memory stores the reduced image. A control point coordinate generating unit reads out the image stored in the reduced image memory to detect distortion. An image converter corrects distortion of the image data loaded from the input image memory on the basis of the information of the detected distortion. The image data whose distortion is corrected is stored in a corrected image memory. The image data whose distortion is corrected and which is stored in the corrected image memory is read out and recorded, so that when an image pickup operation of characters having distortion is to be performed, the distortion is corrected to obtain a flat image, free from distortion, like an image on a flat surface.

Abstract: A multi-stage multi-network character recognition system decomposes the estimation of a posteriori probabilities into coarse-to-fine stages. Classification is then based on the estimated a posteriori probabilities. This classification process is especially suitable for the tasks that involve a large number of categories. The multi-network system is implemented in two stages: a soft pre-classifier and a bank of multiple specialized networks. The pre-classifier performs coarse evaluation of the input character, developing different probabilities that the input character falls into different predefined character groups. The bank of specialized networks, each corresponding to a single group of characters, performs fine evaluation of the input character, where each develops different probabilities that the input character represents each character in that specialized network's respective predefined character group.

Abstract: A method and an apparatus for acquisition of calibrated three dimensional data from camera image. The apparatus for acquisition of calibrated three dimensional data from camera image includes a cameral, alight source and an image processing computer. The camera acquires a light strip image of a target object. The light illuminates light strip to the target object and provides information about the illumination angle form base line (or reference line). The image processing computer obtains image and information about the angle .theta. of light plane from base line; computes connection strength of neural network and acquires calibrated three dimensional data in neutral network based on the obtained information. The mapping relationship between a control point in the three-dimensional space and a control point projected onto the two-dimensional image plane and illumination angle of light source to control point are derived by the neural network circuit.

Abstract: An interval ?0, 1! of the output of neural network is equally divided into M (M being an integer of two or more), and the numbers or frequencies of data for the correct/incorrect patterns contained in the i-th interval ?(i-1)/M, i/M! are .mu.1i and .mu.0i, respectively (where, i=1 . . . M). In this case, if this network provides an output contained in the i-th interval to unknown pattern data, this pattern is stored as a likelihood conversion table so that the pattern outputs likelihood P1i, which is a category, in an equation P1i=(.mu.1i+1)/(.mu.1i+.mu.0i+2). Then, when a value contained in the i-th interval ?(i-1)/M, i/M! is output from a neural network, the likelihood convertor receives it as an input and outputs P1i which is so to speak normalized likelihood.

Abstract: A first image signal representing a radiation image of an object is obtained by exposing a stimulable phosphor sheet, on which the radiation image has been stored, to stimulating rays, which cause the stimulable phosphor sheet to emit light in proportion to the amount of energy stored thereon during its exposure to radiation, the emitted light being detected. A second image signal representing the radiation image is thereafter obtained by again exposing the stimulable phosphor sheet to stimulating rays, the light emitted by the stimulable phosphor sheet being detected. Read-out conditions, under which the second image signal is to be obtained, and/or image processing conditions, under which the second image signal having been obtained is to be image processed, are adjusted on the basis of the first image signal. A storage device stores information representing a standard pattern of radiation images.

Abstract: A system for automatically classification of human fingerprints. An unidentified fingerprint is processed to produce a direction map. The direction map is processed to generate a course direction map. The coarse direction map is input to a locally connected, highly constrained feed-forward neural network. The neural network has a highly structured architecture well-suited to exploit the rotational symmetries and asymmetries of human fingerprints. The neural network classifies the unidentified fingerprint into one of five classifications: Whorl, Double Loop, Left Loop, Right Arch and Arch.

Abstract: The invention is embodied in an image data system including a lossy image compressor having an image compression ratio in excess of 10 for producing first compressed image data from an original image, the first compressed image data specifying a corresponding one of a set of predetermined images, apparatus for computing an difference between the original image and the predetermined image specified by the first compressed image data and a lossless image compressor for compressing at least the difference to produce second compressed image data.

Abstract: An embodiment of the present invention locates facial features in an image by bandpass filtering the image and then performing morphological operations followed by a thresholding operation. This initial processing identifies candidate areas where facial features may be located. The candidate areas are evaluated by classifiers to determine if a facial feature, such as an eye or mouth, has been located.

Abstract: A system and process for performing character recognition is disclosed wherein inputted characters are compared to prototypes maintained in a predetermined database of the system to determine the best matching character. To generate the prototype database for use in recognition, training character samples are inputted and classified and prototypes, having feature value vectors, are generated for each class. The prototypes are optimized to improve the recognition capabilities of the database. The prototype feature value vectors are updated by only small amounts for abnormal prototypes that are much closer to the nearest class M than to the true class I. In addition, the updating of the prototype feature value vectors is performed so as to minimize an error in the selection of the prototypes.

Abstract: Training data is LPC analyzed to obtain a feature parameter vector sequence, which is subjected to Viterbi segmentation using reference phoneme models to separate phonemes. Each piece of phoneme data is used to estimate a mean vector of the corresponding reference phoneme model by a maximum a posteriori estimation method. The adapted phoneme model and the corresponding reference phoneme model are used to estimate a mean vector for an unadapted phoneme model through interpolation by a vector field smoothing method. Alternatively, the mean vector of the adapted phoneme model is further smoothed by the vector field smoothing method. By this, an adapted model is obtained which has, as its parameters, the mean vector obtained for each phoneme and other corresponding parameters.