Abstract: A method for fractal-darwinian object generation consists of the steps of: preparing a fractal object library including predetermined objects and associated rules of property and context, forming objects and comparing the formed objects with the objects in the fractal object library. By using the property rules, a local classification likelihood is allocated to each formed object. Thereupon, by using the context rules for each object, a respective fractal classification likelihood is formed. For optimisation of the fractal classification likelihood, alteration rules are applied to the objects. The above method is carried out iteratively, whereby a process of gradual optimisation takes place.

Abstract: Two different recognition processing units respectively perform a plurality of different recognition processing operations for the pattern to be recognized input from a character input unit, and acquire a plurality of recognition candidates and recognition scores. An accuracy calculation unit converts the recognition scores obtained by the two recognition processing units into accuracies based on a scale common to the plurality of different recognition processing operations using parameters stored in a parameter storage unit. A determination unit generates a recognition candidate sequence by determining the order of the recognition candidates obtained by the two recognition processing units on the basis of the accuracies converted by the accuracy calculation unit.

Abstract: A system and method for real-time multi-view (i.e. not just frontal view) face detection. The system and method uses a sequence of detectors of increasing complexity and face/non-face discriminating thresholds to quickly discard non-faces at the earliest stage possible, thus saving much computation compared to prior art systems. The detector-pyramid architecture for multi-view face detection uses a coarse-to-fine and simple-to-complex scheme. This architecture solves the problem of lengthy processing that precludes real-time face detection effectively and efficiently by discarding most of non-face sub-windows using the simplest possible features at the earliest possible stage. This leads to the first real-time multi-view face detection system which has the accuracy almost as good as the state-of-the-art system yet 270 times faster, allowing real-time performance.

Abstract: A hierarchical-segmenting pre-processor hierarchically segments the axis of each coordinate of a characteristic value used for classifying a pattern to decide a threshold value for each dimension, and generates hierarchized learning patterns which indicate a segment of the segmented coordinate axis to which each coordinate of the characteristic value of the learning pattern belongs. A developing variable determining unit selects a dimension (variable) having the highest classification efficiency in each node of a classification tree. A classification tree generator generates a classification tree based on the dimension (variable) obtained by the developing variable determining unit and a threshold value set in the dimension.

Abstract: An image-processing method, and an apparatus for carrying out the method, that is used particularly for detecting and recognizing objects in traffic. Locally-softened feature images of an image scene are correlated with feature images of patterns for detecting and recognizing objects in real scenes with the use of images. A plurality of patterns ordered in a pattern tree structure is correlated with the image scene. To extensively prevent erroneous detections, the features are subdivided into different feature types.

Abstract: The value of a reference feature vector calculated from a feature vector of an input pattern is transformed into a candidate category set, by a mapping described in a candidate table. Then, pattern recognition is performed using the candidate category set. By suitably setting the mapping, a high-speed process is performed while maintaining a recognition accuracy.

Abstract: A method of machine learning utilizing binary classification trees combined with Bayesian classifiers in which a training component includes creating nodes, utilizing the expectation maximization algorithm to create statistical kernels and mixtures for leaf nodes and kappa statistics to identify nodes as leaf nodes or branch nodes, creating hyperplanes for branch nodes, splitting the training data for branch nodes into subsets to be provided to subtrees, and in which an operating component traverses a binary classifier tree with a feature vector to be classified by determining for each branch node whether the feature vector lies to the left or the right of the branch node hyperplane and finally classifies the feature vector upon arriving at a leaf node by computing log likelihoods for the feature vector for each mixture in the leaf node and determining the classification of the feature vector according to the highest log likelihood, in which the preferred embodiment described is classification of internal com

Abstract: Methods for segmenting audio-video recording of meetings containing slide presentations by one or more speakers are described. These segments serve as indexes into the recorded meeting. If an agenda is provided for the meeting, these segments can be labeled using information from the agenda. The system automatically detects intervals of video that correspond to presentation slides. Under the assumption that only one person is speaking during an interval when slides are displayed in the video, possible speaker intervals are extracted from the audio soundtrack by finding these regions. Since the same speaker may talk across multiple slide intervals, the acoustic data from these intervals is clustered to yield an estimate of the number of distinct speakers and their order. Clustering the audio data from these intervals yields an estimate of the number of different speakers and their order.

Abstract: A method of performing statistical pattern recognition in a set of data having predetermined dimensions involving the steps of performing feature selection on the set of data to determine a selected feature; performing pattern recognition using the set of data with the selected feature to determine a recognized pattern; and outputting the recognized pattern.

Abstract: A color measuring system includes a color classifying apparatus and a color nonuniformity checking apparatus capable of measuring color nonuniformity which simplify the configuration, reduce cost, withstand mechanical vibration, and are capable of classifying the colors of objects well without limiting the light source even when its spectrum varies, and which are capaple of further improving the color classification accuracy.

Abstract: A hybrid surface mount component inspection system which includes both vision and infrared inspection techniques to determine the presence of surface mount components on a printed wiring board, and the quality of solder joints of surface mount components on printed wiring boards by using data level sensor fusion to combine data from two infrared sensors to obtain emissivity independent thermal signatures of solder joints, and using feature level sensor fusion with active perception to assemble and process inspection information from any number of sensors to determine characteristic feature sets of different defect classes to classify solder defects.

Abstract: A method and an arrangement are presented for pattern recognition on the basis of statistics. According to the method, for an object to be recognized on the basis of a complete set of two-class or multiclass classifiers, the association with each target class of the class set is estimated with a numerical value that is produced by cascaded use of polynomial classifiers. According to the invention, on a learning sample in which all class patterns to be recognized are sufficiently represented, there is a selection, from all the two-class or multiclass classifiers by way of their estimation vector spectrum, of those two-class or multiclass classifiers with estimations contributing the most to minimize a scalar quantity calculated over the estimation vector spectrum and having high separating relevance. The selected two-class or multiclass classifiers are subsequently used to form, via an expanded learning sample, estimation vectors from which expanded characteristic vectors are produced by polynomial linking.

Abstract: A pattern matching encoding device for executing pattern matching encoding of binary still images and a pattern matching decoding device corresponding to the pattern matching encoding device are proposed. In the pattern matching encoding, each input pattern extracted from the input image is matched against library patterns in the library, and the input pattern is encoded using a matched library pattern as a reference pattern if the matched library pattern is found. The pattern matching encoding device comprises a pattern segmentation section for segmenting each of selected library patterns and the input pattern into two or more parts and thereby generating segmented library patterns and segmented input patterns, a matching section for matching each of the segmented input patterns against corresponding segmented library patterns, and a pattern combination section for generating a new library pattern by combining the segmented library patterns each of which has matched one of the segmented input patterns.

Abstract: In an artificial intelligence system for image recognition, a global image of an object is input from a camera or other optical pick-up device, and is processed in a global image processing means, which performs analytical processing on the global image by extracting global characteristics of the input image and evaluating consistency of the extracted global characteristics. Simultaneously, the image data is processed in a local image processing means which undertakes analytical processing on a plurality of local images defining local portions of the image to be recognized. The local image processing means is constructed by plural modules, each further defined by sub-modules, which conduct respective analyses corresponding to local images having characteristics useful in recognizing the global image, wherein each local processor extracts characteristics of an input local image and evaluates consistency of the extracted characteristic with the object to be recognized.

Abstract: A learning system learns terms in the context of a set of documents. During an accumulation phase, the learning system accumulates contextual data for a term in the form of a categorization schema. The categorization schema, which is based on a classification hierarchy, classifies the term in categories such that the classifications are based on uses of the terms in the set of documents During a computational phase, the learning system analyzes the categorization schema, and selects, if sufficient contextual data has been accumulated, a single category in the classification system to classify the term. A content processing system, which understands the thematic content of documents, is used in conjunction with the learning system.

Abstract: A fast way for determining the best subset test for a nominal attribute in a decision tree. When a nominal attribute has n distinct values, the prior art requires computing the impurity functions on each of the 2.sup.n-1 -1 possible subset partitioning of the n values and finding the minimum case among them. This invention guarantees the minimum impurity test on the attribute by computing only (n-1) impurity function computations. This reduction of computational complexity makes it practically possible to find the true best tests for many real data mining application, where a binary decision tree is used as the classification model.

Abstract: An image processing device and method for identifying an input image and a copier including such an image processing device and method. The image processing device forms a gradated image from the input image, determines whether the gradated image includes a specified pattern, extracts a portion of the input image corresponding to the specified pattern, and compares the extracted portion of the input image with a reference pattern to determine a degree of resemblance between the extracted portion of the input image and the reference pattern.

Abstract: Character recognition with an improved recognition ratio is provided without requiring special operations to be performed before character recognition is performed. A preliminary character recognition operation is performed in which the feature vector of an input character is compared to a recognition dictionary that contains a reference vector for each category. The candidate category to which the input character belongs is determined, and the feature vector and recognition result for each input character is saved. The input characters judged to have been recognized with high reliability are selected with reference to their recognition results. The feature vector of each selected input character is used to predict a writer-specific feature vector of a category different from the candidate category to which the selected input character belongs.

Abstract: A handwritten pattern recognition system for recognizing an input pattern is provided. The system has a plurality of parameter determining units, each determining the value of a desired parameter for an input pattern to be recognized. The system also includes a pattern match determiner which produces match values for each parameter of the input pattern with its corresponding parameter of each reference parameter. The match determiner also produces an overall match value for each reference pattern. A pattern classifier selects the reference pattern whose parameter set is "closest", by some matching criterion, to that of the input pattern.

Abstract: In a decision-forest classifier in accordance with the invention, a decision forest including multiple decision trees is used to classify "seen" training data and "unseen" data. Each individual tree performs an initial classification based on randomly selected subsets of the data. The classification outcomes by the individual trees are combined using a discriminant process in the decision-forest classier to render the ultimate classification decision.

Abstract: Fast, low-overhead implementations of a powerful, reliable image matching engine based on the Hausdorff distance are disclosed. In one such implementation, a method is provided in which a processor receives two inputs. The first input is a pattern to be recognized in an image; the second, a digital image in which the pattern is to be recognized. The digital image is preprocessed with the processor using various morphological dilation operations so as to produce a set of preprocessed digital images. Thereafter, the processor performs a hierarchical search for the pattern in the digital image. The hierarchical search is performed over a search space, and includes a series of decisions, each decision indicating whether a portion of the search space can be eliminated from the search. Each decision is made by performing a plurality of comparisons between the pattern and the preprocessed digital images of the set and analyzing the results of these comparisons.

Abstract: A character and/or character-string retrieving method with retrieves a plurality of patterns at a time by using a single deterministic finite automaton prepared from a plurality of different patterns. There is also a method for optimizing the number of states for the above-mentioned retrieving method, and a storage medium having records of programs and data necessary for executing the above-mentioned character and/or character-string retrieving and a state number optimizing method. A plurality of regular expressions r.sub.1, r.sub.2, . . . , r.sub.n to be simultaneously retrieved by pattern matching are prepared, and then augmented to form an augmented regular expression ((r.sub.1)#.sub.1,).vertline.((r.sub.2)#.sub.2).vertline. . . . ((r.sub.n)#.sub.n). A deterministic finite automaton is constructed so that it treats states including positions corresponding to #.sub.1, #.sub.2, . . . , n, thereby simultaneously retrieving a plurality of regular expression patterns by distinguishing matches from one another.

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: The invention detects areas of interest at low magnification, locating possible abnormal cells or other cells of interest using image processing and statistical pattern recognition techniques. Next, at high magnification, the areas identified at low magnification are re-examined. The information from the low magnification and high magnification scans is collated and a determination is made about the slide--whether it is normal, abnormal, contains endocervical component, and so forth. The invention also provides a method and apparatus to train object feature and slide feature classifiers. The invention provides an automated cytology system that can process training slides for use in a feed back classifier development environment. The invention also can classify endocervical groups of cells.

Abstract: A technique for the enhancement of degraded document images to improve their display quality characteristics and image recognition accuracy. Images believed to be representative of the same symbol which occur in different positions over a image source (e.g., a facsimile page) are clustered together. Using the symbols within a particular cluster, an average character image outline for that cluster of symbols is derived and thereafter used to refine the matching of symbols within the cluster and to determine a final representative symbol for that cluster. The final representative symbols from the various resulting clusters are then used to replace all matching images throughout the image source. Advantageously, the display quality and recognition accuracy of the image source is enhanced after application of the present invention due to the resulting improvement of the images in the image source.

Abstract: In a decision-forest classifier in accordance with the invention, a decision forest including multiple decision trees is used to classify "seen" training data and "unseen" data. Each individual tree performs an initial classification based on randomly selected subsets of the data. The classification outcomes by the individual trees are combined using a discriminant process in the decision-forest classier to render the ultimate classification decision.

Abstract: The invention pertains to a method and a device (60) for carrying out the quality control of an object (10). According to this method, at least a partial region of the object is scanned by means of at least one optical scanning system (40), and the gray scale values of the scanned points are recorded and stored. The gray scale values of at least part of the points are evaluated by an evaluation unit (64). In order to attain a faster display of the object (10), the invention proposes that at least part of the recorded gray scale values are stored in a storage unit (62), that at least one respective point is classified as being possibly defective or flawless based on its gray scale value in a filter unit (63) that is arranged in front of the evaluation unit (64), and that at least the points that were classified as being possibly defective are evaluated by the evaluation unit (64) based on the measured data stored in the storage unit (62).

Abstract: A word shape token-based document classification system prepares a plurality of sets of training data degraded by image quality and selects the optimum training data set by examining scores from a relevance measurement. The system achieves high accuracy from a wide range of image quality.

Abstract: Pattern recognition apparatus and method which provides plural pattern recognition modes, one of which is selected in accordance with characteristics of an unknown pattern to be recognized. The pattern recognition apparatus includes a pattern recognizer having plural recognition modes for comparing an unknown pattern with at least one known pattern, and a selector for selecting one of the plural recognition modes based on characteristics of the unknown pattern. In one aspect, the selection of one of the plural recognition modes is based on font size, font type and image quality, and in another aspect, selection of the recognition mode is based on the results of prior pattern recognition operations. The pattern recognition apparatus may be provided in a stand-alone device, and this device may interface to a local area network bus.

Abstract: A method and apparatus for hierarchical character recognition processing of ambiguous and noisy characters which produces highly reliable results at high levels of hierarchical processing. The invention first applies a universal classifier system (which may comprise one or more universal classifiers) to input image data, and identifies "suspicious" characters. The image data for suspicious characters is then applied to a "specialist" classifier that is designed to handle only a narrow and well-defined set of recognition cases. This hierarchical processing architecture and method results in increased accuracy of recognition. The method is particularly applicable to handwritten characters and to distorted and noisy machine-printed characters.

Abstract: Method of, and apparatus for, optical character recognition of predominate and non-predominate color coded characters based on hypothesis testing utilizing both geometric optical character recognition and color coded optical character recognition techniques. In this hypothesis testing technique, geometric OCR is used to generate a null hypothesis that a target character has been recognized, and an alternative hypothesis that the target character has not been recognized. Color coded OCR is used to select either the null hypothesis or the alternative hypothesis. Alternatively, the roles of geometric OCR and color coded OCR may be reversed such that color coded OCR is used for hypothesis generation, and geometric OCR is used for hypothesis selection.

Abstract: A method and apparatus for optical identification an unknown character from a plurality of known characters. Each of the known characters includes a predominant color and a nonpredominant color in preselected proportions. The unknown character has at least one geometric feature and a plurality of pixels including a predominant color and a nonpredominant color. The method and apparatus of the present invention include an examination of at least one geometric feature of the unknown character. A hypothetical identity for the unknown character is generated in response to the examination of at least one geometric feature of the unknown character. A portion of the plurality of pixels of the unknown character is sampled and a proportion between the predominant color and the nonpredominant color within the sampled portion of the plurality of pixels is determined from the sampled pixels.

Abstract: An image processing apparatus for performing accurate discrimination between specific originals and ordinary originals. CCD 101 inputs RGB signals, and color-space matching judgment circuit 106 and pattern matching judgment circuit 113 judge the similarity in the color distribution and pattern between the input RGB data and a predetermined specific image data. CPU 114 judges the similarity and print-signal generator 105 generates a change signal in accordance with the judged similarity. The input RGB signals are changed based on the generated change signal and outputted to external color printer 111.

Abstract: A method and apparatus for consolidation of equivalence classes of scanned symbols as may be used for image data compression. The present invention performs run-length symbol extraction and classifies symbols based on both horizontal and vertical run length information. An equivalence class is represented by an exemplar. Feature-based classification criteria for matching an exemplar is defined by a corresponding exemplar template. The feature-based classification criteria all use quantities that can be readily computed from the run endpoints. Reducing the number of equivalence classes is achieved through a process called equivalence class consolidation. Equivalence class consolidation utilizes the symbol classifier to identify matched exemplars indicating equivalence classes which may be merged. For a consolidated equivalence class, the exemplar matching the most symbols is selected as the representative for the class.

Abstract: A method and system for identifying boundaries of characters in handwritten text by classifying segment strokes provides improved performance in a handwriting recognition system. A segment stroke is a portion of handwritten text which includes a boundary between two characters. The segment stroke is recognized by the same method used to recognized characters. Recognition of a segment stroke is accomplished by training a learning machine to act as a classifier which implements a discriminant function based on a polynomial expansion.

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: This invention provides a character recognition apparatus comprising: dictionary storage means for storing a plurality of dictionaries; deriving means for deriving a candidate character and an operation result from a character image data by using a first dictionary stored in said storage means; judging means for judging whether or not if the operation result derived by said deriving means satisfies a predetermined reference value; and controlling means for controlling to recognize said character again by using a second dictionary different from said first dictionary used for deriving said operation result, if said judging means judges that said operation result does not satisfy said reference value.

Abstract: A method and apparatus for classification of scanned symbol into equivalence classes as may be used for image data compression. The present invention performs run-length symbol extraction and classifies symbols based on both horizontal and vertical run length information. An equivalence class is represented by an exemplar. Feature-based classification criteria for matching an exemplar is defined by a corresponding exemplar template. The feature-based classification criteria all use quantities that can be readily computed from the run endpoints. Reducing the number of equivalence classes is achieved through a process called equivalence class consolidation. Equivalence class consolidation utilizes the symbol classifier to identify matched exemplars indicating equivalence classes which may be merged. For a consolidated equivalence class, the exemplar matching the most symbols is selected as the representative for the class.

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: Lower accuracy picture coincidence detecting means gives positional relationship between picture data of higher resolution picture and lower resolution picture as corresponding pixel position predicted with marker position as reference, i.e., initial position. Block matching processing means sets a search range corresponding to a block of an arbitrary range set within the lower resolution picture in accordance with the initial position to globally search, within the search range, block on the higher resolution picture corresponding to the block of the arbitrary range to find out a candidate of coincidence area to further interpolate between respective pixels of lower resolution picture and higher resolution picture to thereby evaluate, with high accuracy, similarity between the set block and block within the search range to detect a picture in which coincidence is obtained.

Abstract: Fiduciary mark detection system usable with an over-the-belt optical character recognition (OCR) reader for ascertaining the position and orientation of text within the destination address on parcels moving along a conveyor, comprising charged coupled device (CCD) array, analog-to-digital (A/D) converter, general purpose computer, and software program. Image processing software program comprising projection histograms, convolution filtering, correlation, radial variance, first moments of inertia, edge image analysis, Hough method, detection confidence testing. Orientation defining mark comprising fluorescent ink placed approximately in the center of the destination address block on a parcel is non obstructive of the underlying text and may be affixed to parcel any time prior to scanning. Fiduciary mark comprising two circles of different diameter oriented such that a vector from the center of large circle to center of small circle is oriented in the same direction as underlying text.

Abstract: There is disclosed an apparatus for image reading or processing, that can precisely identify a particularly pattern, such as of banknotes or securities faithful copying of which is to be prohibited. The apparatus is provided with a detecting unit for detecting positional information of an original image, and a discriminating unit for extracting the pattern data of a certain part of the original image and discriminating whether the original image is the predetermined image, based on the similarity between the pattern data and the predetermined pattern.

Abstract: A character recognition apparatus for recognizing characters in a document by sequentially detecting one-character portions of images in a document image and by recognizing the character images of the one-character portions detected. The character recognition apparatus includes a character detecting section for detecting an image of one character from the document image; a character recognizing section for recognizing the detected image of one character and outputting a character code; and a control section for effecting control such that the character image of one character recognized by the character recognizing means is stored in correspondence with the character code obtained as a result of recognition, a comparison is made between the stored character image and the image of a newly detected character, and the character-code stored in correspondence with the relevant character image is read as a result of recognition in a case where the similarity of the images is sufficiently large.

Abstract: A method of sorting out candidate characters in a character recognition system comprising a character characteristic extracting step of extracting characteristics of characters on the basis of run-length information of the characters, a character distribution extracting step of determining position information of the characters on the basis of the extracted characteristics of the characters, a character classification reference data storing step of storing the determined position information of the characters as character classification reference data in storage means, and a candidate character sorting-out step of adopting the characters belonging to a particular one of the position information stored in the storage means corresponding to a characteristic of an input character to be recognized, as the candidate characters for the input character.

Abstract: The segmentation process of the invention consists in making representative objects of classes to be discriminated and for which a statistical law is estimated or is available which defines them on characteristic measurements. The data base is segmented by determining cuts between the classes taken two by two, the cuts being computed using the descriptive parameters of the statistical laws of each class. Each new object is assigned to the most probable class.

Abstract: An apparatus and method of classifying and sorting by shape crystalline objects such as synthetic diamonds in which an image of the object taken from an angle defined in relation to the object is compared to one or more templates in order to characterize the object.

Abstract: A method and system for comparing an input image with a stored image. An input image is obtained and processed. By the processing of the input image, a rough judgment of whether the input image corresponds to the stored image is obtained. If the input image does correspond to the stored image, then further processing is preformed on the input image. Characteristics of the processed input image are determined. The system then determines whether the characteristics correspond to characteristics of the stored image.

Abstract: A pattern recognition apparatus includes a pattern analyzer for analyzing a similarity between an input pattern and a learning pattern, a hypothetic pattern forming section for forming, as a hypothetic pattern, a predetermined pattern corresponding to the input pattern in accordance with an analysis result from the pattern analyzer, a comparator for comparing the hypothetic pattern formed by the hypothetic pattern forming section with the input pattern, an attribute descriptor in which a comparison result obtained by the comparator and the analysis result obtained by the pattern analyzer are described, and an inference section for, referring to a description described in the attribute descriptor, outputting a clear analysis result obtained from the pattern analyzer as a final recognition result, and starting the hypothetic pattern forming section in response to an ambiguous analysis result by the pattern analyzer.

Abstract: A method of finding a most likely match for a target facial image within a data base of stored facial images comprising determining a score for each data base image as a function of closeness of a quantization of selected facial features between each data base image and the target image and ordering the data base for sequential processing according to the potential value score in descending order, sequentially processing each data base image starting from the highest potential value score by an image comparison process to establish a correlation score for each comparison, and applying one or more decision rules to each comparison to reach a decision.

Abstract: A computer implemented process of matching two lines of text produced by optical character readers analyzes words in the two lines to identify matching words in corresponding positions in the two lines. Lines are then identified as matching based on the number of matching words in corresponding positions.