Abstract: An intrusion prevention/detection system filter (IPS filter) performance evaluation is provided. The performance evaluation is performed at both the security center and at the customer sites to derive a base confidence score and local confidence scores. Existence of new vulnerability is disclosed and its attributes are used in the generation of new IPS filter or updates. The generated IPS filter is first tested to determine its base confidence score from test confidence attributes prior to deploying it to a customer site. A deep security manager and deep security agent, at the customer site, collect local confidence attributes that are used for determining the local confidence score. The local confidence score and the base confidence score are aggregated to form a global confidence score. The local and global confidence scores are then compared to deployment thresholds to determine whether the IPS filter should be deployed in prevention or detection mode or sent back to the security center for improvement.

Abstract: In general, the subject matter described in this specification can be embodied in methods, systems, and program products. A plurality of electronic training images that are each classified as displaying substantially pictures is obtained. A plurality of local image features in each of the plurality of electronic training images is identified. A plurality of weak classifiers are recursively applied to the local image features. During each iteration a weak classifier that accurately classifies the local images features is selected. After each selection of a weak classifier features that were misclassified by the selected weak classifier are given greater weight than features that were classified correctly by the selected weak classifier. For each selected weak classifier a hillclimbing algorithm is performed to attempt to improve the weak classifier. A strong classifier that is a weighted combination of the selected weak classifiers on which hillclimbing algorithms have been performed is produced.

Abstract: A cell phone having distributed artificial intelligence services is provided. The cell phone includes a neural network for performing a first pass of object recognition on an image to identify objects of interest therein based on one or more criterion. The cell phone also includes a patch generator for deriving patches from the objects of interest. Each of the patches includes a portion of a respective one of the objects of interest. The cell phone additionally includes a transmitter for transmitting the patches to a server for further processing in place of an entirety of the image to reduce network traffic.

Abstract: A plurality of images inputted in an image signal input portion are divided into a plurality of regions by an image dividing portion, and a feature value in each of the plurality of regions is calculated by a feature value calculation portion and divided into a plurality of subsets by a subset generation portion. On the other hand, a cluster classifying portion classifies a plurality of clusters generated in a feature space into any one of a plurality of classes on the basis of the feature value and occurrence frequency of the feature value. And a classification criterion calculation portion calculates a criterion of classification for classifying images included in one subset on the basis of a distribution state of the feature value in the feature space of each of the images included in the one subset.

Abstract: Disclosed are methods, devices, and computer program products for red-eye detection in an image. In one example embodiment, a method for detecting red-eye objects in an image includes several acts. First, a set of candidate red-eye objects identified in the image is received. Then, features are extracted from the candidate red-eye objects and, with a plurality of classifiers, a false red-eye object is eliminated from the set of candidate red-eye objects based on the extracted features. First and second ones of the plurality of classifiers are optimized for classifying objects in a first range of sizes using first and second ones of the extracted features, respectively. Furthermore, third and fourth ones of the plurality of classifiers are also optimized for classifying objects using the first and second ones of the extracted features, respectively, but for objects in a second range of sizes.

Abstract: An image-based pattern recognizer and a method and apparatus for making such a pattern recognizer are disclosed. By employing positional coding, the meaning of any feature present in an image can be defined implicitly in space. The pattern recognizer can be a neural network including a plurality of stages of observers. The observers are configured to cooperate to identify the presence of features in the input image and to recognize a pattern in the input image based on the features. Each of the observers includes a plurality of neurons. The input image includes a plurality of units, and each of the observers is configured to generate a separate output set that includes zero or more coordinates of such units.

Abstract: This invention documents the efforts on the research and development of a miniaturized GPS/MEMS IMU integrated navigation system. A miniaturized GPS/MEMS IMU integrated navigation system is presented; Laser Dynamic Range Imager (LDRI) based alignment algorithm for space applications is discussed. Two navigation cameras are also included to measure the range and range rate which can be integrated into the GPS/MEMS IMU system to enhance the navigation solution.

Abstract: Various technologies and techniques are disclosed that improve cursive handwriting recognition. Cursive handwriting input is received from a user. The system performs a hierarchical prototype search as part of a recognition operation. A same space search is performed against a mixed database that has both print and cursive samples. A same space search is also performed against a cursive database that has only cursive samples. The results of these two same space searches are merged into a combined alternate list. The combined alternate list is then used as a constraint for the dynamic time warp searches that are performed against the mixed and cursive databases, respectively. The results of the dynamic time warp searches are also merged into a final combined alternate list, and the combined alternate list is used to make a recognition decision regarding the user's handwritten input.

Abstract: A system, method and device for grading of meat such as bovine, porcine, sheep, horse or poultry meat among others. The device of this invention is a portable tool, which is approached toward a meat specimen to be analyzed and captures an image. The device then objectively relates the image to meat quality parameters by means of an image analyzing method. The device and method solve, in a practical, fast and satisfactory way, the problem of determining meat quality parameters such as texture, color, and contained intramuscular fat percentage.

Abstract: A data processing apparatus includes a feature-value calculating unit that calculates an image feature value indicating a feature of image data, a case database including a case set including a correspondence of image feature values and functions, and an optimum-function predicting unit that predicts an optimum function based on the case database and the image feature value calculated by the feature-value calculating unit. Due to the optimum-function predicting unit, work efficiency of a user can be improved.

Abstract: A detecting method for measuring gaps between two parts of an object is provided. The detecting method selects a reference image and a measured image, merges the reference image and the measured image to form an image of the object, grids the reference image and the measured image to obtain triangle mesh surfaces, obtains boundary points and outline points of the measured image, and obtains triangles on a joint portion of the reference image and measured points on a corresponding joint portion of the measured image. The method further compares each measured point with an obtained triangle to obtain gap values of the joint portion between the reference image and the measured image, and outputs an analysis report of the gap values on a display device.

Abstract: Down-sampling of an image may be performed in the DCT domain. A multiple layered network is used to select transform matrices for down-sampling a DCT image of size M×N to a DCT image of size I×J. A spatial domain down-sampling method is selected and applied to the DCT image to produce a down-sampled DCT reference image. A learning with forgetting algorithm is used to apply a decay to the elements of the transform matrix and select a transform matrices which solve an optimization problem. The optimization problem is a function of the visual quality of images obtained using the transform matrices and the computational complexity associated with using the transform matrices. The visual quality is a measure of the difference between the down-sampled DCT image obtained using the transform matrices and the visual quality of the DCT reference image obtained using a spatial domain down-sampling method.

Abstract: A system and method for automatically placing a rotated row in a multi-up imposition. Calculations can be performed, which are required to execute an automated multi-up imposition. The dimensions of the input document can be then evaluated against the available area in the imposed layout to determine whether there is sufficient space to add a rotated row to the layout (e.g., the “Dutch Row” in Dutch imposition). If it is determined that there is sufficient space for a rotated row, a rotated page row can then be added to the imposed layout.

Abstract: An image-based pattern recognizer and a method and apparatus for making such a pattern recognizer are disclosed. By employing positional coding, the meaning of any feature present in an image can be defined implicitly in space. The pattern recognizer can be a neural network including a plurality of stages of observers. The observers are configured to cooperate to identify the presence of features in the input image and to recognize a pattern in the input image based on the features. Each of the observers includes a plurality of neurons. The input image includes a plurality of units, and each of the observers is configured to generate a separate output set that includes zero or more coordinates of such units.

Abstract: A method for determining the presence or absence of malignant features in medical images, wherein a plurality of base comparison or training images of various types of lesions taken of actual patient is examined by one or more image reading experts to create a first database array. Low-level features of each of the lesions in the same plurality of base comparisons or training images are determined using one or more image processing algorithms to obtain a second database array set. The first and second database array set are combined to create a training database array set which is input to a learning system that discovers/learns a classifier that maps from a subset of the low-level features to the expert's evaluation in the first database array set. The classifier is used to determine the presence of a particular mid-level feature in an image of lesion in a patient based solely on the image.

Abstract: A system and method for printing target colors includes a print-engine interface and a neural network component. The print-engine interface is in operative communication with a print engine of a printing system. The neural network component is calibrated to the print engine for printing a target color on a substrate. The neural network is in operative communication with the print-engine interface and communicates a parameter associated with printing the target color on the substrate utilizing the print engine.

Abstract: An apparatus for efficiently recognizing a part of a body shown in each of plural axial images for one series of axial images obtained by imaging an object to be inspected with a modality. The apparatus includes: apart determining unit for tentatively determining a part of a body shown in each of plural axial images; and a part correcting unit for correcting the part tentatively determined for at least one axial image by the part determining unit, based on information on the plural axial images.

Abstract: Provides quantitative data about a two or more dimensional image. Classifies and counts number of entities an image contains. Each entity comprises a structure, or some other type of identifiable portion having definable characteristics. The entities located within an image may have different shape, color, texture, etc., but still belong to the same classification. Alternatively, entities comprising a similar color/texture may be classified as one type while entities comprising a different color/texture may be classified as another type. May quantify image data according to set of changing criteria and derive one or more classifications for entities in image. I.e., provides a way for a computer to determine what kind of entities (e.g., entities) are in image and counts total number of entities visually identified in image. Information utilized during a training process may be stored and applied across different images.

Abstract: A method of compression of videotelephony images characterized by: creating (10) a learning base containing images; centering the learning base about zero; determining component images by principal component analysis (12); and keeping a number of significant principal components (14).

Abstract: We propose methods for generating a halftone image, in which each pixel takes one of two tone values. The generated image contains hidden data, which is present at data storage pixels chosen using a pseudo-random number generator. In a first case, the data is hidden within an existing halftone image by reversing the tone value at certain of the data storage pixels, and at pixels neighbouring the data storage pixels. In a second case, the halftone image is generated from a grey-scale image, and data is hidden during this conversion process.

Abstract: A method for directed machine learning includes receiving features including intensity data and location data of an image, condensing the intensity data and the location data into a feature vector, processing the feature vector by a plurality of classifiers, each classifier trained for a respective trained class among a plurality of classes, outputting, from each classifier, a probability of the feature vector belong to the respective trained class, and assigning the feature vector a label according to the probabilities of the classifiers, wherein the assignment produces a segmentation of the image.

Abstract: The present invention is an embedded audience measurement platform, which is called HAM. The HAM includes hardware, apparatus, and method for measuring audience data from image stream using dynamically-configurable hardware architecture. The HAM provides an end-to-end solution for audience measurement, wherein reconfigurable computational modules are used as engines per node to power the complete solution implemented in a flexible hardware architecture. The HAM is also a complete system for broad audience measurement, which has various components built into the system. Examples of the components comprise demographics classification, gaze estimation, emotion recognition, behavior analysis, and impression measurement.

Abstract: An image-based pattern recognizer and a method and apparatus for making such a pattern recognizer are disclosed. By employing positional coding, the meaning of any feature present in an image can be defined implicitly in space. The pattern recognizer can be a neural network including a plurality of stages of observers. The observers are configured to cooperate to identify the presence of features in the input image and to recognize a pattern in the input image based on the features. Each of the observers includes a plurality of neurons. The input image includes a plurality of units, and each of the observers is configured to generate a separate output set that includes zero or more coordinates of such units.

Abstract: Disclosed are systems and methods for reconstructing a string comprising characters given multiple strings that may contain one or more errors. In embodiments, pairwise comparisons of strings within a set of candidate strings that may contain errors is performed so that measures related to similarity between the strings in each pair may be calculated. In embodiments, an alignment and voting method using at least some of the similarity measures may be employed to generate a reconstructed string. In embodiments, the reconstructed string may be validated, and in embodiments, its characters may be converted to another encoding, if necessary.

Abstract: A method for coding pixels or voxels of a digital or digitalized two dimensional or three dimensional image includes the steps of: providing a digital image in which each pixel or voxel is defined by a variable in a grey scale image, or by HSV or RGB values in a color image; generating a vector univocally associated to the target pixel or voxel and having components generated as a function of the values of the pixels or voxels in a window that includes the target pixel or voxel and neighboring pixels or voxels. The function of the values of the target pixel or voxel and of the pixels or voxels in the window corresponds to the characteristic parameters of the numerical matrix representing the pixels or voxels of the window or of a transformation of the numerical matrix.

Abstract: Each video segment in a plurality of video segments is annotated with an indicator of the likelihood that the respective video segment shows a particular feature. The plurality of video segments forms an episode of interest from a given video domain. Initial feature probabilities are calculated for respective ones of the plurality of video segments using a machine learning algorithm. Each initial feature probability indicates the likelihood that its respective video segment shows the particular feature. Refined feature probabilities are determined for respective ones of the plurality of video segments by finding the most probable state sequence in a finite state machine. This is accomplished at least in part using the determined initial feature probabilities. Finally, each of the video segments in the plurality of vides segments is annotated with its respective refined feature probability.

Abstract: Image resizing for web-based searching is described. In one implementation, a system resizes a user-selected thumbnail image into a larger version of the image that emulates the quality of a large, original image, but without downloading the original image. First, the system extracts resizing parameters when each thumbnail image is created. Then, the system creates a codebook of primitive visual elements extracted from a collection of training images. The primitive visual elements in the codebook provide universal visual parts for reconstructing images. The codebook and a resizing plug-in can be sent once to the user over a background channel. When the user selects a thumbnail image for enlargement, the system resizes the thumbnail image via interpolation and then refines the enlarged image with primitive visual elements from the codebook. The refinement creates an enlarged image that emulates the quality of the large, original image, without downloading the original image.

Abstract: A system for computer aided detection and decision support includes an ontology of image representations for injecting meaning into and adding relationships among image contents, an image understanding and parsing module in communication with the ontology of image representations for extracting structures from an image including the image contents, and a reasoning engine based in communication with the ontology and the image understanding and parsing module for classifying the image contents, wherein the system receives the image and corresponding descriptive information.

Abstract: Various systems, methods, and programs embodied in computer-readable mediums are provided for fingerprint liveness detection. In one embodiment, a method for determining fingerprint liveness is provided that comprises receiving a fingerprint image; extracting a region of interest from the fingerprint image; generating a histogram of the region of interest; determining a statistical feature of the histogram; and determining liveness of the fingerprint image based upon the statistical feature.

Abstract: The present invention is an apparatus and method for object recognition from at least an image stream from at least an image frame utilizing at least an artificial neural network. The present invention further comprises means for generating multiple components of an image pyramid simultaneously from a single image stream, means for providing the active pixel and interlayer neuron data to at least a subwindow processor, means for multiplying and accumulating the product of a pixel data or interlayer data and a synapse weight, and means for performing the activation of an accumulation. The present invention allows the artificial neural networks to be reconfigurable, thus embracing a broad range of object recognition applications in a flexible way. The subwindow processor in the present invention also further comprises means for performing neuron computations for at least a neuron.

Abstract: A method and apparatus for generating appropriate confirmatory prompts in a speech-enabled, interactive computer system. The method can be incorporated in an interactive voice response system that includes receiving an input audio stream over a voice channel from a users, performing keyword recognition on received input audio as subsequent input audio is being received, and prompting the user with an acknowledgement of the keyword or keywords as subsequent input audio is being received. In another aspect of the method, the volume of the speech input can be continuously monitored. In a further aspect of the method, recognition results and associated confidence values can be combined to select different confirmatory prompts, and the volume is tailored to be the same as, louder than or quieter than the volume of the speech input, so that different types of confirmation can be automatically generated to produce a natural speech-enabled interface.

Abstract: A feature area extracting section extracts an area having a unique feature in an image input to an image selection support apparatus. A specific area feature collating and determining section determines whether or not the area having a feature and extracted by the feature area extracting section is a specific area. A specific area image reading section decides a rectangular area including the specific area, and reads image information of the rectangular area. The specific area image reading section has at least one of an enlargement displaying section which enlarges and displays the image information read by the specific area image reading section, a thumbnail display section which reduces and displays the input image, and an original image displaying section which enlarges and displays the input image.

Abstract: The present invention discloses an object image detection method, which uses a coarse-to-fine strategy to detect objects. The method of the present invention comprises steps: acquiring an image and pre-processing the image to achieve dimensional reduction and information fusion; using a trained filter to screen features; and sequentially using a coarse-level MLP verifier and a fine-level MLP verifier to perform a neural network image detection to determine whether the features of the image match the features of the image of a target object. The present invention simultaneously uses three mainstream image detection methods, including the statistic method, neural network method and adaboost method, to perform image detection. Therefore, the present invention has the advantages of the rapidity of the adaboost method and the accuracy of the neural network method at the same time.

Abstract: Face-based image clustering systems and methods are described. In one aspect, face regions are detected in images. At least one respective parameter value is extracted from each of the face regions. Ones of the face regions associated with parameter values satisfying a cluster seed predicate are classified as cluster seed face regions. The cluster seed face regions are clustered into one or more clusters. A respective face model is built for each of the clusters. The face models are stored. In another aspect, face regions are detected in images. At least one respective parameter value is extracted from each of the face regions. The face regions are ranked based on the extracted parameter values. The face regions are clustered in rank order into one or more clusters. Representations of ones of the clusters are rendered on a display.

Abstract: A method for recognition between a first and second object represented by at least one first image and at least one second image, includes defining rectangular assemblies of random pixels; filtering the first image with first n filters obtained from the assemblies of pixels to obtain n first filtered matrices; classifying the n first filtered matrices by providing a first center and a first radius within a space of N dimensions; filtering the second image with the first n filters to obtain n second filtered matrices; classifying the n second filtered matrices by providing a second center within the space of N dimensions; and comparing the first center and first radius with the second center.

Abstract: The disclosed terrain model is a generative, probabilistic approach to modeling terrain that exploits the 3D spatial structure inherent in outdoor domains and an array of noisy but abundant sensor data to simultaneously estimate ground height, vegetation height and classify obstacles and other areas of interest, even in dense non-penetrable vegetation. Joint inference of ground height, class height and class identity over the whole model results in more accurate estimation of each quantity. Vertical spatial constraints are imposed on voxels within a column via a hidden semi-Markov model. Horizontal spatial constraints are enforced on neighboring columns of voxels via two interacting Markov random fields and a latent variable. Because of the rules governing abstracts, this abstract should not be used to construe the claims.

Abstract: A Neural Gas network used for pattern recognition, sequence and image processing is extended to a supervised classifier with labeled prototypes by extending a cost function of the Neural Gas network with additive terms, each of which increases with a difference between elements of the class labels of a prototype and a training data point and decreases with their distance. The extended cost function is then iteratively minimized by adapting weight vectors of the prototypes. The trained network can then be used to classify mass spectrometric data, especially mass spectrometric data derived from biological samples.

Abstract: A classifier (20) is trained by a feature matrix (18, 18?) made up of feature vectors (F11, . . . , Fkm). The feature vectors are generated by operating on each of a plurality (k) of training image data sets with each of a plurality (m) of image processing algorithms (121, . . . , 12m) to generate processed and segmented images. Features of the segmented regions are extracted (14) to generate the feature vectors. In this manner, the classifier is trained with data generated with a variety of image processing algorithms.

Abstract: A calculation processing apparatus, which executes calculation processing based on a network composed by hierarchically connecting a plurality of processing nodes, assigns a partial area of a memory to each of the plurality of processing nodes, stores a calculation result of a processing node in a storable area of the partial area assigned to that processing node, and sets, as storable areas, areas that store the calculation results whose reference by all processing nodes connected to the subsequent stage of that processing node is complete. The apparatus determines, based on the storage states of calculation results in partial areas of the memory assigned to the processing node designated to execute the calculation processing of the processing nodes, and to processing nodes connected to the previous stage of the designated processing node, whether or not to execute a calculation of the designated processing node.

Abstract: A device for detecting a suspicious activity, efficiently performing surveillance of a person and vehicle detected by a surveillance camera. The device uses signal generation means, a sensor for receiving a signal from the signal generation means, and image data taken by the camera. The device detects mobile bodies, such as a person and a vehicle, included in an image taken by the camera and uses them as mobile body information, uses as authentication information an ID number represented by a signal received by the sensor, defines permitted activities and prohibited activities according to authorized activities corresponding to authentication information, and issues an alarm when the device detects that an activity represented by mobile body activity information is a illegal or prohibited activity.

Abstract: An efficient, effective and at times superior object detection and/or recognition (ODR) function may be built from a set of Bayesian stumps. Bayesian stumps may be constructed for each feature and object class, and the ODR function may be constructed from the subset of Bayesian stumps that minimize Bayesian error for a particular object class. That is, Bayesian error may be utilized as a feature selection measure for the ODR function. Furthermore, Bayesian stumps may be efficiently implemented as lookup tables with entries corresponding to unequal intervals of feature histograms. Interval widths and entry values may be determined so as to minimize Bayesian error, yielding Bayesian stumps that are optimal in this respect.

Abstract: A video quality measurement (VQM) system for a video stream includes a neural network VQM module having a constructed architecture to calculate a video quality metric based on a hybrid of image metric and packet metric of the video stream. An image metric measuring module receives the video stream and calculates the image metric of the video stream. A packet metric measuring module obtains information about packet-level characteristics of the video stream to calculate the packet metric. The image metric and the packet metric are inputted to the neural network VQM module to calculate the video quality metric. The VQM system further includes a VQM test-bed that determines and validates the architecture of the neural network VQM module. Furthermore, a video quality measurement (VQM) method based on a hybrid of image metric and packet metric is also described.

Abstract: A calculation processing apparatus for executing network calculations defined by hierarchically connecting a plurality of logical processing nodes that apply calculation processing to input data, sequentially designates a processing node which is to execute calculation processing based on sequence information that specifies an execution order of calculations of predetermined processing units to be executed by the plurality of processing nodes, so as to implement the network calculations, and executes the calculation processing of the designated processing node in the processing unit to obtain a calculation result. The calculation apparatus allocates partial areas of a memory to the plurality of processing nodes as ring buffers, and writes the calculation result in the memory while circulating a write destination of data to have a memory area corresponding to the amount of the calculation result of the processing unit as a unit.

Abstract: A spot color dictionary is fine tuned or updated. Localized color production models are determined for spot colors of interest to be produced by an associated document processing system or printer. Measurements are made of colors of produced spot colors. Optionally, measurements are made of colors of test patches that are based on perturbations from the colors of the spot colors. In determining a model for the production of a target spot color, measurement data related to colors that are closer in color space to a given target color is given a higher weight than is measurement data related to colors that are further in color space from the target color. Accordingly, the model is localized to the region of color space about the target color and therefore, more accurately predicts a colorant recipe for the target color than would interpolation based on a full gamut, or more general model.

Abstract: Computer-readable media having computer-executable instructions distinguish the script type of at least one portion of a writing input. At least one sub-word of a writing line of a handwritten document is identified and is processed to determine the associated writing style that includes a cursive writing style and a hand-printed writing style. The writing line is consequently associated with a script type. The script type of a writing line is determined from the script types of the sub-words in the writing line. When the number of sub-words having a first script type is greater than the number of sub-words having a second script type, the script type of the writing line is categorized as the first script type. In addition, a script analyzer determines a writing style of at least one sub-word and selects one of a plurality of neural networks to categorize the script type of a writing line.

Abstract: Disclosed is an apparatus and method for detecting a face from an image. The apparatus and method uses color components and enables a Support Vector Machine (SVM) having superior recognition performance to previously learn face and non-face images and determine whether an image is a face image based on a learned image database by reducing the size of a feature vector of a face as compared to conventional systems. Accordingly, the apparatus converts a face image into a mosaic image having a minimum size to reduce the dimension of the feature vector, in order to rapidly and correctly detect a face image.

Abstract: A miniature autonomous apparatus for performing scene interpretation, comprising: image acquisition means, image processing means, memory means and communication means, the processing means comprising means for determining an initial parametric representation of the scene; means for updating the parametric representation according to predefined criteria; means for analyzing the image, comprising means for determining, for each pixel of the image, whether it is a hot pixel, according to predefined criteria; means for defining at least one target from the hot pixels; means for measuring predefined parameters for at least one target; and means for determining, for at least one target whether said target is of interest, according to application-specific criteria, and wherein said communication means are adapted to output the results of said analysis.

Abstract: A plurality of kinds of feature amounts are collected from image information and voice information on a person existing in a space, valid values of the collected feature amounts are calculated, feature amounts to be used for personal recognition are determined in the collected feature amounts on the basis of the calculated valid values, and personal recognition is performed by using the determined feature amounts.

Abstract: A body part is scanned to produce a first set of imaging data. A target lesion in the imaging data is identified. The body part is rescanned at a subsequent time so as to produce a second set of imaging data. The target lesion is identified in the second set of imaging data and the size of the target lesion is measured in the first and second sets of imaging data to determine two apparent image volumes corresponding to the first and second sets of imaging data. A change in size is estimated by comparing the first and second apparent lesion sizes. A variance on the change in size is estimated so as to determine a bound on the change in size measurement.

Abstract: Various technologies and techniques are disclosed that improve handwriting recognition operations. Handwritten input is received in training mode and run through several base recognizers to generate several alternate lists. The alternate lists are unioned together into a combined alternate list. If the correct result is in the combined list, each correct/incorrect alternate pair is used to generate training patterns. The weights associated with the alternate pairs are stored. At runtime, the combined alternate list is generated just as training time. The trained comparator-net can be used to compare any two alternates in the combined list. A template matching base recognizer is used with one or more neural network base recognizers to improve recognition operations. The system provides comparator-net and reorder-net processes trained on print and cursive data, and ones that have been trained on cursive-only data. The respective comparator-net and reorder-net processes are used accordingly.