Hitachi, Ltd.

Hatachi Ltd

Hiroshi Sakou

Masahiro Abe

Haruo Yoda

Hidenori Inouchi

Junichi Higashino

Hisao Ogata

Tsugio Takahashi

Hitoshi Matsushima

Hiromichi Fujisawa

Tatsuya Murakami

Hidefumi Masuzaki

Masaaki Fujinawa

Seiji Kashioka

Yoshihiro Shima

Masashi Koga

Kazunori Kinoshita

Kiyomichi Kurino

Darrin R Uecker

Yozo Ohuchi

Hiroshi Ouchi

To reliably discriminate the characters only from the headlines having any background textures, the present invention analyzes the projection profile of a grey image in a headline area in order to automatically set a learning area A in the area of background texture only of the periphery in the headline area, and set a learning area B in the area that includes background texture of a central portion and characters; determines a discriminant function having, as variables, characteristics that use a plurality of pixel densities in the vicinity of a pixel which is processed so that output values of the discriminant function at each of the positions in the learning area A forms a profile with an average Va, that the output values at each of the positions in the learning area B forms a profile with an average Vb, and that the sum of dispersion values of the two profiles becomes smaller than a predetermined value, in order to discriminate in which area is included the pixel which is processed in the headline area; and determines in which area of the background pattern or the characters the pixel is included depending upon whether the output value of the discriminant function is close to the value Va or to the value Vb for each of the pixels in the headline area, in order to discriminate the headline area into areas.

Texture discrimination method

A neural network (100) has an input layer, a hidden layer, and an output layer. The neural network stores weight values which operate on data input at the input layer to generate output data at the output layer. An error computing unit (87) receives the output data and compares it with desired output data from a learning data storage unit (105) to calculate error values representing the difference. An error gradient computing unit (81) calculates an error gradient, i.e. rate and direction of error change. A ratio computing unit (82) computes a ratio or percentage of a prior conjugate vector and combines the ratio with the error gradient. A conjugate vector computing unit (83) generates a present line search conjugate vector from the error gradient value and a previously calculated line search gradient vector. A line search computing unit (95) includes a weight computing unit (88) which calculates a weight correction value. The weight correction value is compared (18) with a preselected maximum or upper limit correction value (.kappa.). The line search computing unit (95) limits adjustment of the weight values stored in the neural network in accordance with the maximum weight correction value.

Learning method and apparatus for neural networks and simulator with

A file document image input can have shading removed to produce a deshaded image that is useful for highly efficient compression encoding, to be thereafter stored and transmitted in such efficient encoded form. When the image is retrieved or received, the decoded and deshaded image may have shading returned to it by combining with a shaded template image stored in the template image memory or by synthesizing shading in specified regions. By removing shading from regions, optical character recognition, shading identification or other processing such as smear prevention can be enhanced.

Document image processing system and document image processing method

An automatic pattern recognition method has a short processing time, can be applied to nonlinear separation problems, and can perform similarity calculations. The method: Divides a plurality of sample data of known categories into a plurality of classes; When the sample data in a divided class is not all in the same category, repeats dividing the sample data into subclasses until sample data in a subclass has only one category; Expresses the relationship between classes and subclasses in a tree-structure representation and determines the standard pattern for each class and subclass from the sample data contained there; and Checks which of the tree-structured classes input data of unknown category is nearest, by calculating the distance to the standard pattern of each class, and then, when the class has subclasses, performs a similar check until the lowest-level subclass is reached to determine the subclass the input data is closest to. The category of the lowest-level subclass is taken as the category of the input data.

Automatic clustering method

A learning control method reduces overall learning time by displaying data related to an appropriate determination of learning protraction and a proper restoring method. Prior to initiating the learning, the user is inquired about the current problem and a problem data set representing items associated with the problem is obtained. Evaluation data indicating a state of learning obtained during the learning on the current problem is sequentially stored and displayed. When there is a high possibility of learning protraction during the learning, a message informing the user is displayed. When the learning is stopped by the user in this case, the problem data set and evaluation data set are stored. Then, a list of restoring methods is displayed and a particular restoring method is selected by the user once the learning is stopped. The learning is restarted on the current problem in accordance with the selected restoring method.

Method of and system for controlling learning in neural network

A point pattern matching method and system for use in an object recognizing system for deciding pair combinations between a first group of n points x.sub.j (n: integer of 2 or more) (j: integer between 1 and n) in a K-dimensional space (K: integer of 2 or more) and a second group of n second points x.sub.i (i: integer between 1 and n) in the same space as the K-dimensional space. A total of n.sup.2 of neurons are provided one for each one of point pair combinations between the first group of points and the second group of points, and, when an output of each of the neurons has a value of substantially '1', it is determined that the point pair combination associated with the neuron is matched, whereas, when the output of the neuron has a value of substantially '0', it is determined that the point pair combination associated with the neuron is not matched.

Point pattern matching method and system as well as picture recognizing

A pipe-lined data processor system comprising a plurality of processors interconnected in the form of a pipe line, a plurality of memory apparatus connected to each of the processors through output data buses for supplying data to the processors or storing data from the processors, and a control board connected to the memory apparatus through input data buses for managing the operation sequences of the processors and the memory apparatus, wherein the plurality of memory apparatus comprise each a main memory for storing the data to be processed as predetermined by the processors or the data having been processed as predetermined by the processors, a buffer memory for temporarily storing the data, first control means for controlling the timing of input/output of the data between the main memory and the buffer memory, and second control means for controlling the timing of input/output of the data between the buffer memory and the input/output buses.

Pipe-lined data processor system of synchronous type having memory

A digital image analysis system is disclosed in which a digital input image formed by a raster scan method is so modified as to fill up a hole in a clump and a recess at the bottom of a clump viewed in the sub-scanning direction of the raster scan method, for the purpose of forming a control image, the state of a clump at two consecutive scanning lines of the control image (that is, the generation and termination of the clump at one of the scanning lines or the continuity of the clump at the scanning lines) is detected from the values of adjacent pixels on the two consecutive scanning lines, and a feature value of the clump is calculated on the basis of the detected state of the clump.

Digital image analysis system

A method of determining an optimum threshold for obtaining a binary image signal for use in a pattern recognition system is disclosed in which a pattern of a predetermined shape which exists at one or a plurality of positions on a surface of an object to be imaged is previously stored as a reference pattern. A threshold for obtaining a binary signal is changed stepwise. A video signal is converted into the binary signal on the basis of each value of threshold to obtain a binary image. A plurality of partial patterns which are different in position on the binary image to each other, are successively cut out of the binary image to be compared with the reference pattern. The overall degree of coincidence between the plural partial patterns and the reference pattern is calculated, and a value of threshold at which the overall degree of coincidence is highest, is selected from various values of threshold.

Growing community of inventors

Shiki, Japan

Hiroshi Sakou