WO1995017734A1 - Procede et appareil destines a la reconnaissance de structures et procede de compilation d'un dictionnaire en vue de la reconnaissance de structures - Google Patents

Procede et appareil destines a la reconnaissance de structures et procede de compilation d'un dictionnaire en vue de la reconnaissance de structures Download PDF

Info

Publication number
WO1995017734A1
WO1995017734A1 PCT/JP1994/002160 JP9402160W WO9517734A1 WO 1995017734 A1 WO1995017734 A1 WO 1995017734A1 JP 9402160 W JP9402160 W JP 9402160W WO 9517734 A1 WO9517734 A1 WO 9517734A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
pattern
divided
dictionary
creating
Prior art date
Application number
PCT/JP1994/002160
Other languages
English (en)
French (fr)
Inventor
Takao Kobayashi
Original Assignee
Birds Systems Research Institute, Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Family has litigation
First worldwide family litigation filed litigation Critical https://patents.darts-ip.com/?family=18371229&utm_source=google_patent&utm_medium=platform_link&utm_campaign=public_patent_search&patent=WO1995017734(A1) "Global patent litigation dataset” by Darts-ip is licensed under a Creative Commons Attribution 4.0 International License.
Application filed by Birds Systems Research Institute, Inc. filed Critical Birds Systems Research Institute, Inc.
Priority to KR1019950703453A priority Critical patent/KR100242458B1/ko
Priority to US08/500,995 priority patent/US5689584A/en
Priority to EP95903008A priority patent/EP0686933A4/en
Publication of WO1995017734A1 publication Critical patent/WO1995017734A1/ja

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10Character recognition
    • G06V30/18Extraction of features or characteristics of the image
    • G06V30/18086Extraction of features or characteristics of the image by performing operations within image blocks or by using histograms
    • G06V30/18095Summing image-intensity values; Projection and histogram analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/211Selection of the most significant subset of features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10Character recognition
    • G06V30/19Recognition using electronic means
    • G06V30/191Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06V30/1912Selecting the most significant subset of features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10Character recognition

Definitions

  • the present invention relates to a pattern recognition method and apparatus for recognizing characters and various patterns input with reference to a predetermined dictionary, and a method of creating a pattern recognition dictionary used for the same.
  • pattern recognition for recognizing letters, numbers, and various patterns is performed by using a dictionary that stores standard patterns and comparing the read patterns with the patterns in the dictionary. I have. Therefore, if the types of characters, numbers, and various patterns to be recognized increase, the size of the dictionary for storing them increases, and the recognition processing time increases accordingly. Therefore, it is desired to reduce the time required for pattern recognition.
  • the pattern recognition method includes a method based on pattern matching and a method based on feature point extraction.
  • a pattern matching method uses a scanner or the like to optically read printed characters and handwritten characters. It matches with a plurality of standard patterns stored in a pattern recognition dictionary to find the standard pattern with the highest similarity. The name is determined as the name of the input pattern.
  • a dictionary in which the vertical and horizontal distribution of each part of a character and the relationship between a character segment and an adjacent character segment are recorded as a character feature in advance is created.
  • the characteristic points are similarly obtained for the read characters, and the characteristic points are compared.
  • the character corresponding to the feature point having the highest similarity is determined as a read character.
  • a pattern recognition method using a neurocomputer includes, for example, an input layer consisting of neurons corresponding to dots constituting a two-dimensional pattern such as a character, a numeral, and various patterns to be recognized, and an output layer consisting of a new mouth corresponding to a recognition output. And an intermediate layer that performs connection by weighting between them. The weight of the intermediate layer is adjusted by the backpropagation rule, etc., and upon completion of learning, the pattern input to the input layer is adjusted. A recognition result such as a pattern name is output from the output layer.
  • an object of the present invention is to provide a pattern recognition method and a pattern recognition method that reduce the dictionary creation time, reduce the recognition processing time even when a large number of character types are mixed, and improve the recognition rate.
  • An object of the present invention is to provide a device and a dictionary creation method.
  • a further object of the present invention is to make it possible to obtain similarities for all categories in recognition processing ⁇ -. Disclosure of the invention
  • the pattern recognition method comprises the steps of: (a) dividing an area to be recognized of an input pattern into N areas, assigning a corresponding area number to each area, and creating N areas. (B) calculating a feature amount for each of the N divided regions according to a predetermined criterion, and creating a feature vector having the N feature amounts as elements; Searching for the largest or smallest one of the N elements of the feature vector, creating a first feature set consisting of one segmented area number corresponding to the searched element; Two are searched in order from the largest or the smallest, and a second feature set including a combination of two divided area numbers corresponding to the searched two elements is created.
  • (N-1) Creating a total of (N-1) feature sets by creating a feature set; and (d) performing steps (a) to (c) in advance for various model patterns. Copy the obtained special set to the model
  • the similarity between the feature set of the input pattern and the feature set in the dictionary is determined for each category, and the category having the maximum similarity is determined. Determining a name as one category of the input pattern.
  • the pattern recognition device includes: pattern input means for inputting a pattern to be recognized; dividing a recognition target region of the input pattern into N regions, and corresponding divided regions. Means for assigning a number to create N divided regions; calculating a feature amount for each of the N divided regions according to a predetermined criterion, and using the N feature amounts as elements Means for creating a feature vector to be searched for; searching for the largest one or the smallest one among the N elements of the feature vector, and from one divided area number corresponding to the found element. A first feature set is created, and then two are searched in order from the largest one or the smallest one, and a second one consisting of a combination of two divided area numbers corresponding to the searched two elements is obtained.
  • (N— 1) Means for creating a total of (N-1) feature sets by creating up to the (N-1) th feature set consisting of a combination of the divided area numbers of the pattern input means, the divided area creating means, A means for storing a feature set obtained by applying the feature set creating means to various example patterns in advance together with one category of the example pattern; and a dictionary storing means. Means for obtaining the similarity between the feature set of the input pattern and the feature set in the dictionary for one category, and determining the category having the highest similarity as the category of the input pattern. I do.
  • the method for creating a pattern recognition dictionary includes the steps of: (a) dividing a recognition target area of an input model pattern into N areas, assigning a corresponding divided area number to each of the areas, For creating divided areas (B) calculating a feature quantity for each of the N divided areas according to a predetermined criterion, and creating a feature vector having the N feature quantities as elements; (C) Search the largest or smallest of the N elements of the feature vector, and create a first feature set consisting of one divided area number corresponding to the searched element.
  • a step of creating a total of (N-1) feature sets by creating up to an (N-1) th feature set including a combination of (N-1) divided area numbers, and d) Predetermined memory by associating each of the feature sets with one category of the example pattern. And the step of storing in the.
  • FIG. 1 is a diagram for explaining the principle of the present invention.
  • FIG. 2 is a block diagram showing a hardware configuration of the pattern recognition device according to one embodiment of the present invention.
  • FIGS. 3A, 3B, and 3C are diagrams for explaining the extraction of feature vectors.
  • FIGS. 4A, 4B and 4C are diagrams for explaining the outline of the feature vector.
  • FIG. 5 is a diagram for explaining a pattern recognition dictionary.
  • FIG. 7 is a diagram for explaining compression of dictionary data.
  • Figure 8 is a flowchart showing the procedure of the feature vector creation process.
  • FIGS. 9 to 13 are flowcharts showing the procedure of dot weighting processing of the input pattern.
  • FIG. 14A, 14B, 14C, 14D, 14E and 14F are shown in FIG. It is a figure which illustrates a mode that evening changes by a weighting process.
  • FIG. 15 is a diagram illustrating an input pattern.
  • FIG. 16 is a diagram showing a pattern obtained by performing weighting processing on the input pattern shown in FIG.
  • FIG. 17 is a diagram for explaining a dot value allocated to each area in the dot measurement processing in the divided area for the pattern shown in FIG. 14F.
  • FIG. 18 is a diagram showing a feature vector obtained by the dot measurement processing for the dot distribution shown in FIG.
  • FIGS. 19 and 20 are flowcharts showing the procedure of the dot measurement process in the divided area.
  • FIGS. 21 to 23 are flowcharts showing the procedure of the dictionary creation process.
  • FIGS. 24 and 25 are flowcharts showing the procedure of the input pattern recognition process.
  • FIGS. 26 and 27 are flowcharts showing the procedure of the dictionary search process. BEST MODE FOR CARRYING OUT THE INVENTION
  • FIG. 1 is a diagram for explaining the principle of the present invention.
  • the pattern recognition device basically includes a pattern input unit 1, a pattern recognition dictionary 2, a pattern recognition processing unit 3, as shown in FIG. It is composed of
  • the pattern input unit 1 is configured by a scanner capable of reading a pattern or by a memory storing image data.
  • the dictionary 2 for pattern recognition has a configuration in which a feature set obtained from various pattern names such as katakana, hiragana, kanji and other character names, numeric names, and asterisks is stored in memory in association with the category names. Yes.
  • a feature set can be obtained from model patterns for a plurality of character types belonging to the same category and stored.
  • the pattern recognition processing unit 3 uses an arithmetic processing function such as a microprocessor to generate the dictionary 2 from the input pattern from the pattern input unit 1. The similarity between the feature set of this input pattern and the feature set stored in dictionary 2 is determined for each category, and the category with the highest similarity is determined as the input pattern. It performs processing to determine the category name.
  • FIG. 2 is a block diagram showing a hardware configuration of the pattern recognition device according to one embodiment of the present invention.
  • reference numeral 10 denotes a CPU comprising a general-purpose microprocessor.
  • Reference numeral 11 denotes a ROM for storing a program of the CPU 10 and a dictionary for recognizing a pattern.
  • Reference numeral 12 denotes a RAM for storing temporary data in the operation and control processing of the CPU 10.
  • Reference numeral 13 denotes a scanner that scans a pattern to be recognized by optical means, and reference numeral 14 denotes a control circuit for the scanner.
  • Reference numeral 15 denotes a hard disk drive for storing a pattern recognition dictionary, a pattern recognition result, and the like as a file
  • reference numeral 16 denotes a control circuit for the hard disk drive
  • reference numeral 17 denotes a control circuit for the hard disk drive.
  • reference numeral 18 indicates the flexible disk device. 2 shows a control circuit corresponding to this.
  • Reference numeral 19 denotes a printer for printing the result of pattern recognition processing and the like
  • reference numeral 20 denotes a control circuit for the printer.
  • the CPU 10, the ROM 11, the RAM I 2 and the control circuits 14, 16, 18 and 20 are connected by a system bus 30. Yes.
  • the dictionary for pattern recognition will be described.
  • a dictionary obtains a feature set for an example pattern input via the pattern scanner 13 and stores the feature set in a memory (for example, in correspondence with one category such as a character name, a numeric name, and a pattern name). This is stored in ROM1 1).
  • Such a feature set is obtained from the feature vector.
  • the feature vector divides the recognition target area of the sample pattern into N pieces with the same area or the same number of dots, and uses the number of dots divided by the area of each divided area as elements. Alternatively, N other feature values are obtained from the model parameters and used as elements.
  • This feature, Vector V This feature, Vector V,
  • V ⁇ V 1, V 2, V 3,..., V i,..., V N ⁇ 1)
  • the set whose element is the position on the vector of the selected I elements (this corresponds to the number of the divided area) is called the feature set T,.
  • I 1 to N_1
  • (N-1) feature sets are created.
  • a sequence ⁇ ,, ⁇ 2 , ⁇ , ⁇ -, in which the (N-1) feature sets T, are arranged in the order of I is referred to as a feature set sequence.
  • the characteristic set corresponding to I 1 Is ⁇ 1 ⁇
  • a feature set ⁇ , (I 1 to N—1), whose elements are the I divided area numbers corresponding to the selected I vector elements, is obtained.
  • ⁇ 5 ⁇ 6, 7, 10, 11, 13 ⁇
  • a feature set T, (I 1 to N-1), whose elements are the I divided region numbers corresponding to the selected I vector elements, is obtained.
  • FIG. 3C shows an example in which a pattern recognition target area is divided by concentric circles and radial lines.
  • the number of black dots in the divided area is counted, the number of the area with the largest count value of the eight outermost areas is set to 1, and the area numbers 2 to 24 is assigned, the number of dots in each area is counted, and a feature vector consisting of 24 elements is created. Then, one, two,..., two or three feature vector elements are selected in order from the top or bottom.
  • the X, X, and X are the pattern recognition target areas, where the number of horizontal dots is X and the number of vertical dots is Y.
  • the function f (x, y) also represents between grid points. If a point (x, y) is defined as a contour point, a point (x ', y') where f (x, y) ⁇ f ( ⁇ ', ⁇ ') is very close to the point (x, y) ) Exists. That is, for the closest points (x, y) and (x'.y '), the point where f (x, y) ⁇ f ( ⁇ ', ⁇ ') is a contour point.
  • V i f Ri (2 ⁇ f (x, y) ⁇ 1) xd (x, y, f) dxdy (4)
  • R i is a rectangular divided area
  • JJ Ri () dxdy indicates the area of the rectangular divided area. Also, 1 ⁇ i ⁇ N.
  • the sequence ⁇ , ⁇ ⁇ —, of these sets arranged in the order of I is the feature set sequence.
  • 4A, 4B, and 4C are diagrams for explaining the outline of the feature vector, and show the elements of the feature vector V with the dimension N set to 4 ⁇ (1 in the figure) ) To (4) indicate division area numbers. If the feature vectors shown in Figs. 4A, 4B and 4C are Va, Vb and Vc, respectively,
  • V a (5 0, 0 0, 1 0 0, 2 0 0, 1 1 0)
  • V b (30, 29, 28, 27)
  • V c (— 3, 1, 2, — 1, -4)
  • T a a ⁇ 1, 2, 3 ⁇
  • T b 2 ⁇ 1, 2 ⁇
  • T b 3 ⁇ 1, 2, 3 ⁇
  • T c 3 ⁇ 1, 2, 3 ⁇
  • V A l (100, 90, 80, 70, 60, 50)
  • V A 2 (40, 50, 45, 33, 35, 34)
  • V A 3 (1980, 12, 2000, 1, 0, 2)
  • V A 4 (96, 95, 94, 99, 98, 97)
  • V B l (24, 22, 30, 32, 28, 26)
  • V B 2 (24, 22, 64, 60 52, 56)
  • V B 3 (154, 155, 175, 174, 165, 164)
  • V B 4 (-60, -5, -4, -3, -2, -1)
  • the feature set sequences TA l to TA 4, TB to ⁇ 4 created from the feature vectors VA l to VA 4 and V ⁇ 1 to VB 4 are obtained from the largest feature vector elements in order from 1 One to five elements are sequentially selected, such as two elements, and the next element, and the area numbers corresponding to the selected elements are combined.
  • TA1 ⁇ 1 ⁇ , ⁇ 1,2 ⁇ , ⁇ 1,2, 3 ⁇ , ⁇ 1,2, 3,4 ⁇ , ⁇ 1,2, 3, 4, 5 ⁇
  • TA2 ⁇ 2 ⁇ , ⁇ 2,3 ⁇ , ⁇ 1,2, 3 ⁇ , ⁇ 1,2,3, 5 ⁇ , ⁇ 1,2, 3, 5, 6 ⁇
  • ⁇ 3 ⁇ 3 ⁇ , ⁇ 1, 3 ⁇ , ⁇ 1,2,3 ⁇ , ⁇ 1,2, 3, 6 ⁇ , ⁇ 1,2, 3, 4, 6 ⁇
  • ⁇ 1 ⁇ 4 ⁇ , ⁇ 3, 4 ⁇ , ⁇ 3, 4, 5 ⁇ , ⁇ 3,4, 5, 6 ⁇ , ⁇ 1, 3, 4, 5, 6 ⁇
  • ⁇ 2 ⁇ 3 ⁇ , ⁇ 3, 4 ⁇ , ⁇ 3, 4, 6 ⁇ , ⁇ 3,4, 5, 6 ⁇ , ⁇ 1,3,4, 5, 6 ⁇
  • ⁇ 3 ⁇ 3 ⁇ , ⁇ 3, 4 ⁇ , ⁇ 3, 4, 5 ⁇ , ⁇ 3, 4,5, 6 ⁇ , ⁇ 2, 3, 4, 5, 6 ⁇
  • ⁇ 4 ⁇ 6 ⁇ , ⁇ 5, 6 ⁇ , ⁇ 4, 5, 6 ⁇ , ⁇ 3, 4, 5, 6 ⁇ , ⁇ 2, 3, 4, 5, 6 ⁇ .
  • a dictionary is created by storing these feature sets in correspondence with one category name in a memory.
  • FIG. 5 shows part of a dictionary based on the above example.
  • the feature set ⁇ 3 ⁇ is a category named “AJ”
  • the dictionary creates a feature set from the example pattern, creates a record by associating the feature set with the category name, and creates a record.
  • the time required to create a dictionary is proportional to the number of example patterns.
  • the time required for creating a dictionary can be significantly reduced compared to the conventional example, and when adding an example pattern, it is only necessary to modify and add a part of the dictionary.
  • the time required for pattern recognition processing can be reduced.
  • FIG. 6 is a diagram for explaining the data structure of the pattern recognition dictionary, and is an example in which one category is represented by a bit position.
  • the category name LX when different category names LX,, LX 2 correspond to the same feature set, the category name LX 'stored in the dictionary by taking the logical sum of the category names LX,, LX 2 It can be.
  • one record of the dictionary is formed by a combination of a category name represented by a bit string having the same number of bits as the number of category names to be recognized and a feature set. Will be.
  • the element position of the feature vector that is, the divided region number, can be represented by the bit position in the bit string having the number of bits equal to the number of divided regions.
  • each element of “1” to “6” in the above-described six-division is represented by a bit position in a 6-bit bit string, and in that case, the special set ⁇ 2, 3 ⁇ is “ 0 0 0 1 1 0 ".
  • one category corresponds to one name.
  • the data compression / decompression technique applied to various data processing can be applied. For example, as shown in FIG. 7, the LX data is rearranged and compressed by deleting duplicates to create new LX data. Then, a table is created in which the value of ⁇ ⁇ and the pointer information to L X data are set. These tables may be further compressed if they do not interfere with the recognition process.
  • the feature set ⁇ can be represented by an address.
  • FIG. 8 is a schematic flowchart showing a processing procedure for creating a feature vector.
  • This feature vector creation routine is called in a dictionary creation process and an input pattern recognition process described later. It should be noted that the vector creation processing can employ various modifications in accordance with the use of individual pattern recognition systems and the like, and the following examples are representative.
  • Step i 0 2 it is determined whether to perform weighting on the pattern data.
  • a dot weighting process of the pattern data is performed (step 104), and then a measurement process is performed on the weighted dot (step 106). If weighting is not performed, the dot data measurement process is immediately performed (step 106).
  • the above is an outline of the feature vector creation process. Next, details of the weighting process and the measurement process will be described.
  • step 104 shows flowcharts of the weighting process, and are examples of the process in step 104 of FIG.
  • 0 is assigned to a variable I (designating a dot position in the recognition area) (step 202).
  • I designating a dot position in the recognition area
  • step 202 the number of horizontal dots in the recognition area of the input pattern is X and the number of vertical dots is Y, whether or not I is smaller than XXY, that is, I indicates the position below the last dot position in the recognition area. It is determined whether or not it is (step 204). If I ⁇ XXY, the process proceeds to step 210, and if I ⁇ XxY, the process proceeds to step206.
  • step 206 the input pattern P [I] is multiplied by a predetermined intermediate value MIDDLEVAL to create a weighted pattern Q [I].
  • This input pattern P [I] is, for example, a white dot.
  • This intermediate value is desirably selected so as to prevent the white (“0”) area from eventually becoming a negative value as a result of the weighting. In the present embodiment, too, Set to "1 6". However, it is not limited to this.
  • step 206 ends for all I, that is, for all dots
  • the process proceeds to step 210 (FIG. 10).
  • the pattern Q [I] shown in the figure has been created.
  • Figure 14 4
  • step 210 the intermediate value MIDDLEVAL is assigned to a variable J.
  • 0 is substituted for a variable I, and a predetermined flag FLG is set to 0 (step 2 12).
  • I and XXY are compared (step 2 14).
  • step 2 2 2 I + X and X x Y are compared, and if I + X ⁇ X x ⁇ , the process proceeds to step 2 26, and if ⁇ + ⁇ ⁇ ⁇ Moves to step 222.
  • step 236 the value of the variable I is incremented and the process proceeds to step 216 to execute the above-mentioned processing for the next dot. Back up.
  • the pattern shown in Fig. 14B changes as shown in Fig. 14C at the time of the first execution of step 238, and the white dot portion adjacent to the boundary line has "15". Is set.
  • step 2308 it is determined whether or not the flag FLG is "0". If the flag is "1", decrement J (step 240) and loop back to step 212.
  • the fact that the flag FLG is "0" in step 238 means that the processing has been performed up to FIG. 14D in the case of FIG. 14C described above as an example. Then, in that case, the process proceeds to step 242.
  • step 2 42 the intermediate value MIDDLEVAL is substituted for the variable J.
  • step 244 the number of horizontal dots X is substituted for the variable I, and the flag FLG is set to "0".
  • step 250 Q [I-X] is compared with J. If Q [I-X] is less than J, the process proceeds to step 264, and if Q [I-X] ⁇ J, then If the dot above Q [I] is equal to the variable J, go to step 25.2.
  • step 2 52 Q [I + X] is compared with J. If Q [I + X] ⁇ J, the process proceeds to step 2 64, and if Q [I + X] ⁇ J, If the dot below Q [I] is equal to the variable J, Move to step 2 5 4.
  • step 2 56 (Fig. 13), Q [1-1] is compared with J, and if QC ⁇ -J, the process proceeds to step 2 64, where Q [I-1] If ⁇ J, that is, if the left neighbor of Q [I] is equal to J, go to step 258.
  • step 26 2 J + 1 is substituted into Q [I] because Q [I] and its immediately above, one-fifth, left and right neighbors are all equal to J. That is, when the value of the variable J is 16 (the first value), 17 is assigned to Q [I]. Then, the flag FLG is set to 1 and the flow shifts to step 264.
  • step 264 the same processing is performed for the next dot, and the variable I is incremented, and the process loops back to step 246.
  • step 266 it is determined whether or not the flag FLG is "0". If the flag is not "0”, J is incremented (step 268), and the loop back to step 244 is performed. . If the flag FLG is “0” in step 266, the pattern of FIG. This means that the processing up to the final pattern of 1.4F has been completed, and this weighting process ends. In this way, from the input pattern P [I] (Fig. 14A), a pattern Q [I] (Fig. 14F) weighted according to the distance from the boundary between white and black can be obtained. .
  • FIG. 15 shows an example of a further input pattern having a larger number of dots than FIG. 14A.
  • an input pattern P corresponding to the letter T is shown.
  • step 106 in FIG. 8 the procedure for measuring the dots in the divided area (step 106 in FIG. 8), which is performed to determine the feature vector from the input pattern or the pattern after the weighting processing, is described in detail. explain.
  • this measurement processing the number of black dots or the weighted dot value in the divided area created by dividing the recognition target area into a plurality of pieces is added.
  • N dimension of the feature vector
  • step 306 the process proceeds to step 306.
  • step 306 0 is substituted for the element V [I] of the feature vector V, and in step 308, I is incremented and looped back to step 304. That is, each element V [I] of the feature vector V is set to 0 in the initial state by the steps 304, 306 and 308.
  • step 310 the variable J is set to 0.
  • J YxVY
  • VY indicates the number of vertical divisions of the recognition area
  • J YXVY
  • step 3 1 8 the expression
  • step 320 I is incremented and step Loop back to 3 16.
  • step 3 22 J is incremented, and loop back to step 3 12 is performed.
  • FIGS. 21 to 23 are flowcharts showing the procedure of the dictionary creation processing. Note that the number of dimensions of the feature vector V is N, and as shown in Fig. 6, this routine creates a dictionary that stores a plurality of records consisting of the feature set TX and the category name LX. is there.
  • step 402 the counter CNT for specifying the address of the record is cleared to "0" (step 402).
  • step 404 a pattern file (for example, stored in the hard disk 15) is opened (step 404).
  • step 406 it is determined whether or not the processing has been completed for all pattern data in the file (step 406).
  • one pattern data of the sample pattern is taken out (step 408).
  • a predetermined variable CODE is set to one numerical value from 0 to: L-1 corresponding to one of the category and one to one (step 410).
  • L is the number of one category.
  • a process of creating a feature vector V composed of N elements is performed (step 412).
  • the initial value 0 is substituted into the feature set T (step Step 4 14), and the variable I is set to 0 (Step 4 16).
  • step 420 the variable MAX J is set to 0, and J is set to 1, and the process proceeds to step 422.
  • step 424 the element V [J] of the feature vector V is compared with the maximum value V (MAX J) so far, and if V [J]> VCM AXJ), step 4 2 Proceed to step 6. If V [J] ⁇ V [MAXJ], proceed to step 428.
  • step 426 since the current element V [J] is larger than the previous maximum value V CMAX J], the current value of J is substituted for the variable MAX J, and the process proceeds to step 428.
  • step 428 J is incremented and loop back to step 422.
  • step 430 11 is substituted for V [MAX J] so that the largest element searched this time is not detected next time the largest element is searched.
  • the elements of the vector are made not to take a negative value.
  • the 1st bit of the J-th MAX in the feature set T is 1.
  • the least significant bit (LSB) shall be bit 0. This processing is performed by using the left shift operator " ⁇ ", one of the C language bit processing operators.
  • the feature set T Is the contents of the feature set table TX [CNT]. Also, the bit of the C OD Eth is set to 1 (however, the least significant bit (LSB) is set to bit 0), and the bit string with the other bits set to 0 is set in the category information table LX CCNT). The contents of.
  • step 434 the variable I is incremented, and the process loops to step 418.
  • the elements of the feature vector V are found in order from the largest one, and the process of creating one record consisting of the feature set and the category information is repeated, and the total (N-1) Create a record for When the processing is completed for one pattern data, the process returns to step 406 to check whether there is still a pattern to be processed.
  • the pattern file is closed (step 4336).
  • a table including the feature set TX and the category information LX is sorted (step 438).
  • other sorting methods can be used.
  • it is determined whether the current dictionary creation processing is for new creation of a dictionary file or for adding data to an already created dictionary (step 440). In the case of a new creation, the contents in the memory are written to a file (step 442).
  • the merge processing shown in Fig. 6 is executed simultaneously.
  • the dictionary and the contents of the memory are merged and written out to a file (step 444).
  • a dictionary for pattern recognition is created.
  • a feature vector similar to that in the dictionary creation described above is created for the input pattern (step 502).
  • step 5 12 the feature set T is set to the initial value of 0.
  • step 5 14 the variable K is set to 0 and the process proceeds to step 5 16.
  • Steps 518 to 528 are the same as steps 420 to 430 in the dictionary creation processing described above, and a feature set is obtained in these steps. Then, a dictionary search process to be described later is performed (step 530), K is incremented (step 532), and the process loops back to step 516.
  • step 5 3 4 the similarity SC 0 RES [I] becomes the maximum I Then, one category corresponding to the I is determined as one category of the input pattern, and the recognition processing of the input pattern ends.
  • FIGS. 26 and 27 are flowcharts showing the procedure of the dictionary search process executed in step 530 in the input pattern recognition process described above.
  • initialization is performed by substituting 0 for a variable I ST ART and a predetermined value T B LMAX for a variable I END.
  • the predetermined value TB LMAX indicates the number of records in the dictionary.
  • step 606 (I START + I END), 2 is substituted for the variable I W. That is, the sum 12 of the start address and the end address is defined as the intermediate address IW.
  • step 610 IW is substituted for IEND and loopback is performed to step 604.
  • step 612 IW + 1 is substituted for ISTART, and the process loops back to step 604.
  • step 614 0 is substituted for I, and the flow advances to the next step 6 16.
  • step 62 the similarity score S CORES [I] is decremented by 1, and the process proceeds to step 62.
  • I is incremented and loop back to step 6 16. The above is the procedure of the dictionary search process.
  • Category Feature set sequence of one example pattern for each person
  • each category has a feature set sequence for a plurality of sample patterns corresponding to different fonts. That is, for a given category c, there are m example patterns, and the feature set matrix, ) T c)
  • T ML (c) and T (c) T (c) are formed, the similarity is defined as follows, and this is called [, width space similarity].
  • I is 1 to N— 1
  • the present invention calculates the width space similarity based on the above-mentioned equation (6) for the feature set obtained from the input pattern and the feature set stored in the character recognition dictionary, and sets the model having the maximum similarity. It is determined that one category of the pattern is one category of the input pattern.
  • V X 1 (6, 8 8 9 9 9 9, — 55, 77, -44 4)
  • V X 2 (2 5, 1 6 3 4, 6 1, 5 2, 4 3)
  • ⁇ 2,3,5 ⁇ does not link to any single category.
  • the similarity between the input pattern P X1 and the category A is 4 Z 5.
  • the similarity of the input pattern P X 1 to the category B is 1/5.
  • the similarity of the input pattern P X 2 with the category A is 3/5
  • the similarity of the input pattern P X 2 with the category B is 4-5. Accordingly, the input pattern P X 1 is recognized as the category A, and the input pattern P X 2 is recognized as the category B. It is also possible to display one recognition candidate category in the order of similarity.
  • similarities SA (c >, SB (c) obtained by the methods A and B are obtained, and the synthesized similarity S (c) is calculated as:
  • the present invention obtains the “width space similarity” between the feature set of the input pattern and the feature set of the dictionary, and determines one category having the maximum similarity as the category of the input pattern. This was judged to be one person, and it was confirmed that the recognition rate was high.
  • the records in the dictionary are arranged in the order of the feature set, and can be searched by, for example, the bisection method or the address expression method. Therefore, it is only necessary to refer to a part of the dictionary without calculating the similarity for the entire feature set, which has the advantage of reducing the time required for recognition.
  • the similarity calculation is performed for one person in all categories, there is an advantage that it is excellent in applicability, such as being applicable to a large classification in pattern recognition.
  • the dot printer prints out the half-size alphanumeric characters of the alphabet, 52 characters, and 10 numeric characters, and the resolution is 300.
  • JIS 1st level Kanji 295 6 character types are printed out with a laser printer, read by the aforementioned 300 dpi scanner, and input as a sample pattern for dictionary creation and recognition rate measurement.
  • a pattern Was. In that case, the perception rate was 99.797%. In this case, the recognition time per character was about 0,19 seconds.
  • dictionaries were created and recognition rates were measured for handwritten characters 0 to 9 and X. In that case, the recognition rate was 98.86% for a total of 109,944 characters. That is, a practically sufficient recognition rate could be obtained.
  • the dictionary for pattern recognition extracts N features from the model pattern to form a feature vector, forms N-1 feature sets from the feature vector, and divides them into the model pattern. It is stored in memory in association with one category name, and multiple feature names can be associated with one feature set. Therefore, even when there are many example patterns and one category, the time required for dictionary creation is short, and the dictionary can be created economically. Furthermore, even when adding a model pattern, there is an advantage that the dictionary can be easily added because only a part of the dictionary needs to be modified and added. Industrial applicability
  • the present invention can be applied to any pattern recognition by appropriately selecting a feature vector creation method. That is, the present invention can be applied not only to an optical character reading device (OCR) as an input device of a computer, but also to a medical diagnosis system and a speech recognition system (using a speech waveform as a pattern). .
  • OCR optical character reading device
  • a high recognition rate means that the pattern used to create the dictionary is always correct, and that in various cases of character recognition (handwritten characters, printed characters, numbers, alphanumeric characters, kanji, etc.) Both are proven by the fact that good recognition rates were actually obtained. ing

Description

明 細 書 パターン認識方法及び装置並びにパターン認識用辞書作成方法 技術分野
本発明は、 所定の辞書を参照して入力された文字や各種パ夕一ン を認識するパターン認識方法及び装置、 並びにそれに用いるパター ン認識用辞書の作成方法に関する。 背景技術
一般的に、 文字、 数字、 各種パターン等を認識するためのパター ン認識は、 標準のパターンを格納した辞書を用い、 読み取ったパ夕 ーンと辞書のパターンとを照合することにより行われている。 した がって、 認識対象とされる文字、 数字、 各種パターン等の種類が多 くなると、 それらを格納する辞書が大型化し、 それに伴って認識処 理時間が長くなるという問題がある。 そこで、 パターン認識の所要 時間を短縮することが要望されている。
従来技術に係るパターン認識方法には、 パターン照合による方法. 特徴点抽出による方法等がある。 パターン照合による方法は、 例え ば、 印刷文字や手書き文字等をスキャナ等により光学的に読み取り. パターン認識用辞書に格納された複数の標準パターンと照合し、 類 似度が最大となる標準パターンの名称を入力パターンの名称と判定 するものである。
また、 特徴点抽出による方法は、 例えば、 文字の各部の垂直方向 及び水平方向の分布や文字素片と隣接文字素片との関係等を文字の 特徴として記録した辞書をあらかじめ作成しておき、 読み取った文 字についても同様に特徴点を求め、 それぞれの特徴点を比較して、 類似度が最大となる特徴点に対応する文字を読み取り文字と判定す るものである。
さらに、 ニューロコンピュータを用いたパターン認識方法も知ら れている。 このパターン認識方法は、 例えば、 認識すべき文字、 数 字、 各種パターン等の二次元パターンを構成する ドッ トにそれぞれ 対応するニューロからなる入力層と、 認識出力に対応するニュー口 からなる出力層と、 それらの間の重みづけによる接続を行う中間層 とを有し、 バッ クプロパゲ一ショ ン則等によって中間層の重みづけ を調整し、 学習完了により、 入力層に入力されたパターンに対して 出力層からパターン名称等の認識結果を出力するものである。
従来のパターン照合による認識方法においては、 入力パターンの 大きさや位置等と、 辞書に格納されている標準パターンの大きさや 位置等とを同等とするための前処理を必要とし、 かつ、 入力パター ンと辞書に格納されている全標準バタ一ンとを照合する必要があり- したがって、 前処理と照合処理とに要する時間が長くなるという欠 点があった。
また、 特徴点抽出による認識方法においては、 認識すべき文字の すべての特徴点について比較する必要があり、 かつ、 文字、 数字、 各種パターンのそれぞれに対する特徴点の数が膨大となる。 そのた め、 高い認識率を得よう とすれば辞書が大型化するという問題があ り、 結果として認識処理時間が長くなるという欠点があった。
また、 標準パターンと共にゴシック体ゃイタ リ ツク体等の文字種 が混在している文字、 数字等の場合においては、 各文字種毎にパ夕 一ン照合又は特徴点の抽出 · 比較を行う必要があり、 同一文字名に 対して各文字種に対応する数の照合又は比較等の処理が必要となり . 認識に要する時間が長く なるという欠点があった。 また、 逐次比較 を行う代わりに、 大分類を行って認識すべきパターン種類を絞り込 むことが考えられているが、 認識率を低下させないように大分類を 行うための最適な手段が実現されていない。
また、 ニューロコ ンピュータを用いたパターン認識方法は、 学習 を 1 0〜数 1 0 0 0回繰り返す必要があり、 かつ、 認識可能なパ夕 一ン数も少ないので、 現在は実用化されていない。
かかる実情に鑑み、 本発明の目的は、 辞書作成時間を短縮し、 多 数の文字種が混在するような場合でも認識処理時間を短縮し、 かつ, 認識率の向上を図った、 パターン認識方法及び装置並びに辞書作成 方法を提供することにある。 また、 本発明の更なる目的は、 認識処 理においてすべてのカテゴリーに対する類似度が得られるようにす る ^-とにある。 発明の開示
本発明に係るパターン認識方法は、 (a) 入力パターンの認識対象 領域を N個の領域に分割し、 それぞれに対応する分割領域番号を付 与して、 N個の分割領域を作成するステップと ; (b) 該 N個の分割 領域の各々に対して、 所定の基準に従って特徴量を算出し、 該 N個 の特徴量を要素とする特徴べク トルを作成するステップと ; (c) 該 特徴べク トルの N個の要素の中の最大のもの又は最小のものを検索 し、 該検索された要素に対応する 1個の分割領域番号からなる第 1 の特徴集合を作成し、 次いで、 最大のもの又は最小のものから順に 2個を検索し、 該検索された 2個の要素に対応する 2個の分割領域 番号の組み合わせからなる第 2の特徵集合を作成し、 以下同様にし て (N— 1 ) 個の分割領域番号の組み合わせからなる第 (N— 1 ) の特徴集合まで作成することにより、 合計 (N— 1 ) 個の特徴集合 を作成するステップと ; (d) 前記ステップ(a) から(c) までを予め 各種の手本パターンに対して実施して得られた特徵集合を該手本パ ターンのカテゴリ一名とともに記録した辞書を参照することにより . 該入力パターンの特徴集合と該辞書内の特徴集合との類似度をカテ ゴリ一名対応に求め、 該類似度が最大となるカテゴリ一名を該入力 パターンのカテゴリ一名と判定するステップと ; を具備する。
また、 本発明に係るパターン認識装置は、 認識されるべきパ夕一 ンを入力するパターン入力手段と ; 入力されたパターンの認識対象 領域を N個の領域に分割し、 それぞれに対応する分割領域番号を付 与して、 N個の分割領域を作成する手段と ; 該 N個の分割領域の各 々に対して、 所定の基準に従って特徴量を算出し、 該 N個の特徴量 を要素とする特徴べク トルを作成する手段と ; 該特徴べク トルの N 個の要素の中の最大のもの又は最小のものを検索し、 該検索された 要素に対応する 1個の分割領域番号からなる第 1 の特徴集合を作成 し、 次いで、 最大のもの又は最小のものから順に 2個を検索し、 該 検索された 2個の要素に対応する 2個の分割領域番号の組み合わせ からなる第 2の特徴集合を作成し、 以下同様にして (N— 1 ) 個の 分割領域番号の組み合わせからなる第 (N— 1 ) の特徴集合まで作 成することにより、 合計 (N— 1 ) 個の特徴集合を作成する手段と 前記パターン入力手段、 分割領域作成手段、 特徴べク トル作成手段. 及び特徴集合作成手段を予め各種の手本パターンに対して適用して 得られた特徵集合を該手本パターンのカテゴリ一名とともに記録し た辞書を記憶する手段と ; 該入カバターンの特徴集合と該辞書内の 特徴集合との類似度をカテゴリ一名対応に求め、 該類似度が最大と なるカテゴリ一名を該入力パターンのカテゴリ一名と判定する手段 と ; を具備する。
また、 本発明に係るパターン認識用辞書作成方法は、 (a) 入力さ れた手本パターンの認識対象領域を N個の領域に分割し、 それぞれ に対応する分割領域番号を付与して、 N個の分割領域を作成するス テツプとパ (b) 該 N個の分割領域の各々に対して、 所定の基準に従 つて特徴量を算出し、 該 N個の特徴量を要素とする特徴べク トルを 作成するステップと ; (c) 該特徴べク トルの N個の要素の中の最大 のもの又は最小のものを検索し、 該検索された要素に対応する 1個 の分割領域番号からなる第 1 の特徴集合を作成し、 次いで、 最大の もの又は最小のものから順に 2個を検索し、 該検索された 2個の要 素に対応する 2個の分割領域番号の組み合わせからなる第 2の特徵 集合を作成し、 以下同様にして (N— 1 ) 個の分割領域番号の組み 合わせからなる第 (N— 1 ) の特徵集合まで作成することにより、 合計 (N— 1 ) 個の特徴集合を作成するステップと (d) 該特徵集 合の各々 と該手本パターンのカテゴリ一名とを対応させて所定のメ モリに格納するステップと ; を具備する。 図面の簡単な説明
図 1 は、 本発明の原理を説明するための図である。
図 2は、 本発明の一実施例に係るパターン認識装置のハー ドゥエ ァ構成を示すプロッ ク図である。
図 3 A、 3 B及び 3 Cは、 特徴ベク トルの抽出について説明する ための図である。
図 4 A、 4 B及び 4 Cは、 特徴ベク トルの概要を説明するための 図である。
図 5は、 パターン認識用辞書を説明するための図である。
図 6は、 パターン認識用辞書のデータ構造を説明するための図で あ Q 0
図 7は、 辞書データの圧縮について説明するための図である。 図 8は、 特徴べク トル作成処理の手順を示すフローチヤ一 トであ O 図 9 〜 1 3は、 入力パターンの ドッ トの重みづけ処理の手順を示 すフローチヤ一トである。
図 1 4 A , 1 4 B , 1 4 C, 1 4 D , 1 4 E及び 1 4 Fは、 ノ、。夕 一ンが重みづけ処理により変化する様子を例示する図である。
図 1 5は、 入力パターンを例示する図である。 ^ 図 1 6は、 図 1 5に示される入力パターンを重みづけ処理して得 られるパターンを示す図である。
図 1 7は、 図 1 4 Fに示されるパターンに対する分割領域内 ドッ ト計測処理に関し、 各領域に配分される ドッ トの値を説明するため の図である。
図 1 8は、 図 1 7に示される ドッ トの配分についての ドッ ト計測 処理により得られる特徴べク トルを示す図である。
図 1 9及び 2 0は、 分割領域内ドッ ト計測処理の手順を示すフ口 一チヤ一トである。
図 2 1 〜 2 3は、 辞書作成処理の手順を示すフローチヤ一 トであ る
図 2 4及び 2 5は、 入力パターン認識処理の手順を示すフローチ ヤー トである。
図 2 6及び 2 7は、 辞書サーチ処理の手順を示すフローチヤ一ト である。 発明を実施するための最良の形態
本発明をより詳細に説述するために、 添付の図面に従ってこれを 説明する。
図 1 は、 本発明の原理を説明するための図である。 本発明に係る パターン認識装置は、 基本的には、 この図に示すように、 パターン 入力部 1 と、 パターン認識用辞書 2 と、 パターン認識処理部 3 とか ら構成される。 パターン入力部 1 は、 パターンの読取りが可能なス キヤナにより構成されるか、 又は画像データを格納したメモリ によ り構成される。 また、 パターン認識用の辞書 2は、 片仮名、 平仮名, 漢字等の文字名、 数字名、 星印等の各種パターン名より得られる特 徴集合をそのカテゴリー名と対応付けてメモリに格納した構成を有 する。 なお、 辞書 2には、 同一カテゴリ一名に属する複数の文字種 について手本パターンから特徴集合を求めて格納することができる ( この手本パターンの追加をする必要がない場合は、 リー ドオンリ メ モリ (R O M ) によって辞書 2を形成することも可能である。 また, パターン認識処理部 3は、 マイクロプロセッサ等の演算処理機能を 用いて、 パターン入力部 1 からの入力パターンから、 辞書 2の作成 時と同様な処理によって特徴集合を求め、 この入力パターンの特徴 集合と辞書 2に格納された特徴集合との類似度をカテゴリ一名対応 に求め、 類似度が最大となるカテゴリ一名を入力パターンのカテゴ リー名と判定する処理を行う ものである。
図 2は、 本発明の一実施例に係るパターン認識装置のハ一 ドゥエ ァ構成を示すブロック図である。 この図において、 符号 1 0は、 汎 用マイクロプロセッサからなる C P Uを示す。 符号 1 1 は、 C P U 1 0のプログラム及びバタ一ン認識用辞書を格納する R O Mを示す, 符号 1 2は、 C P U 1 0の演算 ·制御処理過程における一時的なデ 一夕を格納する R A Mを示す。 符号 1 3は、 光学的手段で認識対象 のパターンを走査するスキャナを示し、 符号 1 4は、 そのスキャナ に対する制御回路を示す。 符号 1 5は、 パターン認識用辞書、 パ夕 一ン認識結果等をフアイルとして格納するハー ドディスク装置を示 し、 符号 1 6 は、 そのハー ドディスク装置に対する制御回路を示す, 符号 1 7は、 同様にそのようなファイルを格納するフレキシブルデ イ スク装置を示し、 符号 1 8は、 そのフ レキシブルディ スク装置に 対する制御回路を示す。 符号 1 9 は、 パターン認識処理結果等をプ リ ン トァゥ トするためのプリ ン夕装置を示し、 符号 2 0は、 そのプ リ ンタ装置に対する制御回路を示す。 C P U 1 0、 R OM 1 1、 R AM I 2並びに制御回路 1 4、 1 6、 1 8及び 2 0は、 システムバ ス 3 0によって接続される。 い.
パターン認識用辞書について説明する。 かかる辞書は、 例えばパ ターンスキャナ 1 3を介して入力した手本パターンについて、 特徵 集合を求め、 文字名、 数字名、 パターン名等のカテゴリ一名と対応 させてその特徴集合をメモリ (例えば、 R OM 1 1 ) に格納したも のである。 かかる特徵集合は、 特徴べク トルから求められる。 また. その特徴べク トルは、 手本パターンの認識対象領域を等面積若しく は等ドッ ト数となるように N個に分割し、 各分割領域の ドッ ト数ゃ 面積等を要素とするか、 又は手本パ夕一ンより N個の他の特徴量を 求めて要素とするものである。 この特徴ベク トル Vは、
V = { V 1 , V 2 , V 3 , ···, V i , ···, V N } … 1 ) と表すことができる。
そして、 特徴べク トル Vの要素 V , 〜 v N の大きいものから順に. 又は小さいものから順に、 I個 ( I = 1 〜N— 1 ) を選ぶ。 選ばれ た I個の要素のべク トル上の位置 (これは分割領域の番号に一致す る。 ) を要素とする集合を特徴集合 T , と称する。 I = 1〜N_ 1 として、 (N— 1 ) 個の特徴集合が作成されることとなる。 この (N - 1 ) 個の特徴集合 T , を I の順番に並べた列 Τ , , Τ2 , ー, ΤΝ-, を特徴集合列と称する。 例えば、 ( 1 ) 式において各要素が 大きいものから順に配列されているとした場合、 すなわち V , , V V 3 , …, V Ν の順に小さ くなるとした場合、 I = 1 に対応する特 徵集合は { 1 } 、 I = 2に対応する特徴集合は { 1 , 2 } 、 1 = 3 に対応する特徴集合は { 1 , 2 , 3 } 、 以下同様にして最後に、 I =N— 1 に対応する特徴集合は { 1 , 2, 3 , ···, Ν - 1 } となる < このように計 (Ν— 1 ) 個の特徴集合が得られる。 この各特徴集合 と手本パターンのカテゴリ 一名とを対応させてメモリ に格納するこ とにより、 パターン認識用辞書が作成される。
例えば、 図 3 Αに示す文字 「A」 について、 認識対象領域が X ド ッ ト X Yドッ トで構成される場合において、 縦に 4等分、 横に 4等 分するときには、 1 6個の等面積の分割領域が形成される。 そして. 各分割領域に領域番号 1〜 1 6を付与し、 文字部分を黒として、 各 分割領域内の黒ドッ ト数を計数する。 この計数値をそれぞれ V , 〜 V 16とすると、 ( 1 ) 式の特徴べク トル Vが得られる。 そして V 1 〜v 16の中から大きいものから順に順次 I (= 1〜N— 1 ) 個を選 ぶ。 次いで、 選ばれた I個のべク トル要素に対応する I個の分割領 域番号を要素とする特徴集合 Τ, ( I = 1〜N— 1 ) を求める。 こ れらを Iの順番に並べた集合の列 T, 〜TN-i ( = T, 〜T15) を 特徴集合列とする。
この文字 「A」 について、 分割領域内の黒ドッ ト数が大きいもの から順に領域番号 1〜 1 6を並べたとき、
6, 7, 10, 11, 13, 16, 2, 3, 9,12, 5, 8, 14, 15, 1, 4 であるとすると、 特徵集合列 〜T15は、
Τ, = {6 }
Τ2 = {6,7}
Τ 3 = {6, 7, 10}
Τ 4 = {6, 7, 10, 11}
Τ 5 = {6, 7, 10, 11, 13}
Τ 15= {1,2,3,5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16} となる。 これらをカテゴリ "τ名 A,と対応付けてメモリに格納する。 また、 図 3 Bは、 黒ドッ ト数が同一となるように、 認識対象領域 を先ず縦に分割し、 次に各分割領域についても、 黒ドッ ト数が同一 となるようにそれぞれ横に分割し、 各分割領域に領域番号 1〜 2 0 を付与した場合を示す。 そして、 各分割領域の面積をそれぞれ V , 〜v2。とすると、 ( 1 ) 式の特徴べク トル Vが得られる。 そして、 面積の大きいものから順に、 或いは面積の小さいものから順に、 順 次 I (= 1〜N— 1 ) 個を選ぶ。 次いで、 選ばれた I個のベク トル 要素に対応する I個の分割領域番号を要素とする特徴集合 T , ( I = 1〜N— 1 ) を求める。 これらを Iの順番に並べた集合の列 T , 〜!^ ( = T, 〜Τ19) を特徴集合列とする。
また、 図 3 Cは、 同心円と放射方向の線とによってパターンの認 識対象領域を分割した例を示す。 この例においては、 分割領域内の 黒ドッ ト数を計数し、 一番外側の 8個の領域について計数値の一番 大きいものの領域の番号を 1 とし、 そこを基準位置として順に領域 番号 2〜 2 4を付与し、 各領域の黑ドッ ト数を計数して 2 4個の要 素からなる特徴べク トルを作成する。 そして、 上位又は下位から順 に、 1個、 2個、 …、 2 3個の特徴ベク トル要素を選ぶ。 選ばれた I個のべク トル要素に対応する I個の分割領域番号を要素とする特 徵集合 Τ , ( Ι = 1〜Ν— 1 ) を求める。 これらを Iの順番に並べ た集合の列 Τ, 〜ΤΝ-, ( = Τ, 〜Τ23) を特徵集合列とする。 こ の方式も、 各種のパターンについて認識を行う場合に適用すること ができる。
次に、 ドッ トの重みづけ処理について説明する。 図 3 Αにおいて 横の ドッ ト数を Xとし、 縦の ドッ ト数を Yとした X XY ( ドッ ト) のパターン認識対象領域について、
f (X, y) e { 0 , 1 } … ( 2 ) l o こ こで、 0≤ xく Xs 0≤ y < Y. x, は整数
と定義する。 従って、 関数 f (x,y) は格子点を表すことになる。 また、 ( 2 ) 式を実数上に拡張し、
f (X, y) R e { 0 , 1 } - ( 3 ) こ こで、 0≤ x <X, 0≤ y < Y . χ, yは実数
と定義する。 従って、 関数 f (x,y) は格子点間も表すことになる。 また、 点(x,y) を輪郭点とすると、 その点(x,y) の極く近く に f (x,y) ≠ f (χ' ,γ' ) となる点(x' ,y' ) が存在する。 即ち、 極く近い 点(x,y) , (x'.y') について f (x,y) ≠ f (χ',γ') である点は輪郭 点である。
また、 点(x,y) から最も近い輪郭点までの距離を d (x,y,f) と表 し、 区間 [0,X), [0, Y) をそれぞれ ηχ , η Ύ 等分し、 nx x n r =N個の分割領域を形成し、 特徴べク トル Vの要素 V i を、
V i = fRi ( 2 ■ f(x,y)- 1 ) xd(x,y,f) dxdy … ( 4 ) と表すことができる。 但し、 R iは長方形の分割領域で、
JJRi ( ) dxdyはその長方形の分割領域の面積分を示す。 また、 1 ≤ i ≤ Nである。
前述の ( 4 ) 式の ( 2 · f (X, y) — 1 ) は、 f (x, y) が 0又は 1 であるから、 f (X, y) = 1のときは 1、 f (X, y) = 0のときは一 1 となる。 すなわち、 ( 4 ) 式は、 最も近い輪郭点からの距離に応じ て、 重み付けして特徴べク トル Vの要素 を求める場合を示すも のである。 そして、 前述のように、 特徴べク トル Vを求め、 要素 V の大きいものから順に、 或いは小さいものから順に、 順次 I (= 1 〜N— 1 ) 個を選ぶ。 そして、 選ばれた I個のべク トル要素に対応 する I個の分割領域番号を要素とする特徴集合 T , ( I = 1〜N— 1 ) を求める。 これらを Iの順番に並べた集合の列 Τ, 〜ΤΝ— , を 特徴集合列とする。 図 4 A、 4 B ,び.4 Cは、 特徴べク トルの概要を説明するための 図であり、 それぞれ次元 Nを 4 とした特徴べク トル Vの要素を示す < 図中の ( 1 ) 〜 ( 4 ) は、 分割領域番号を示す。 図 4 A、 4 B及び 4 Cにそれぞれ示された特徴ベク トルを V a, V b , V c とすると-
V a = ( 5 0, 0 0 , 1 0 0, 2 0 0, 一 1 0 )
V b = ( 3 0 , 2 9 , 2 8 , 2 7 )
V c = (— 3, 一 2, — 1 , - 4 )
となる。
特徴べク トル V aの中の大きいものから順に、 順次 1〜N— 1個- 即ち、 1個、 2個、 3個を選び、 特徴集合 T a , 、 T a 2 、 T a 3 を求めると、
T a , = { 1 }
T a 2 = { 1 , 3 }
T a a = { 1 , 2 , 3 }
となる。 同様に、 特徴べク トル V bから特徴集合を求めると、
T b , = { 1 }
T b 2 = { 1 , 2 }
T b 3 = { 1 , 2 , 3 }
となる。 また、 特徴べク トル V cから特徴集合を求めると、
T c a = { 3 }
T c 2 = { 2 , 3 }
T c 3 = { 1 , 2 , 3 }
と 7よる。
次に、 例えば、 カテゴリ一名 「Α」 , 「 B」 について、 複数種類 の印刷文字、 手書き文字等の各手本パターンを 4種類入力し、 認識 領域を 6個に分割し (すなわち、 特徴べク トルの次元を 6 とする) . 分割領域番号を 1〜 6 として、 それぞれの分割領域の ドッ ト数を計 数した結果、 次のようになった場合を仮定する
領域番号 1 2 3 4 5 6
II V A l = ( 100, 90, 80, 70, 60, 50)
V A 2 = ( 40, 50, 45, 33, 35, 34)
V A 3 = ( 1980, 12, 2000, 1, 0, 2)
V A 4 = ( 96, 95, 94, 99, 98, 97)
V B l = ( 24, 22, 30, 32, 28, 26)
V B 2 = ( 24, 22, 64, 60 52, 56)
V B 3 = ( 154, 155, 175, 174, 165, 164)
V B 4 = ( -60, - 5, -4, -3, -2, -1)
特徴べク トル V A l 〜V A 4 , V Β 1 〜V B 4から作成される特 徴集合列 T A l 〜T A 4 , T B ト Τ Β 4は、 特徴べク トルの要素 の大きいものから順に、 1個、 それと次の要素との 2個というよう に、 順次要素を 1 〜 5個選択し 、 その選択された要素に対応する領 域番号を組合せたものであり、
TA1= {1} , {1,2} , {1,2, 3} , {1,2, 3,4} , {1,2, 3, 4, 5}
TA2= {2} , {2,3} , {1,2, 3} , {1,2,3, 5} , {1,2, 3, 5, 6}
ΤΑ3= {3} , {1, 3} , {1,2,3} , {1,2, 3, 6} , {1,2, 3, 4, 6}
{4, 5} , {4, 5, 6} , 11, 4, 5, 6} , {1,2, 4, 5, 6}
ΤΒ1= {4} , {3, 4} , {3, 4, 5} , {3,4, 5, 6} , {1, 3, 4, 5, 6}
ΤΒ2= {3} , {3, 4} , {3, 4, 6} , {3,4, 5, 6} , {1,3,4, 5, 6}
ΤΒ3= {3} , {3, 4} , {3, 4, 5} , {3, 4,5, 6} , {2, 3, 4, 5, 6}
ΤΒ4= {6} , {5, 6} , {4, 5, 6} , {3, 4, 5, 6} , {2, 3, 4, 5, 6} となる。
これらの特徴集合とカテゴリ一名とを対応させてメモリに格納す ることにより、 辞書が作成される。 図 5は、 上述の例に基づく辞書 の一部を示す。 例えば、 特徴集合 { 3 } は、 カテゴリ一名 「AJ , 「 B J :に'共通であることを示している。 このように、 辞書は、 手本 パターンより特徴集合を作成し、 特徴集合とカテゴリー名とを対応 させてレコー ドを作成し、 そのレコー ドを特徴集合の順番に並べる ものである。 そして、 1 つの特徴集合に対して複数のカテゴリ一名 が対応するのを可能とする。 したがって、 辞書作成に要する時間は. 手本パターン数に比例した時間で済み、 辞書作成のための所要時間 を従来例に比較して大幅に短縮することができる。 また、 手本バタ ーンを追加するときも、 辞書の一部を修正、 追加するだけで済むか ら、 簡単に追加することができる。 また、 このような辞書構造を有 することにより、 パターン認識処理における所要時間を短縮するこ とが可能となる。
図 6は、 パターン認識用辞書のデータ構造を説明するための図で あり、 カテゴリ一名をビッ ト位置で表現した場合の例である。 前述 のように、 同一の特徴集合に対して、 異なるカテゴリ一名 L X , , L X 2 が対応するとき、 カテゴリ 一名 L X , , L X 2 の論理和をと つて辞書に格納するカテゴリ 一名 L X ' とすることができる。 この ような処理を行うことにより、 認識すべきカテゴリ一名の個数と等 しいビッ ト数のビッ ト列で表されるカテゴリ一名と特徴集合との組 合せによって辞書の 1 レコー ドが形成されることになる。 また、 特 徴集合 T Xについても、 特徴べク トルの要素位置すなわち分割領域 番号を、 分割領域数に等しいビッ ト数のビッ ト列におけるビッ ト位 置で表現することができる。 例えば、 前述の 6分割の場合における " 1 " 〜 " 6 " の各要素は、 それぞれ 6 ビッ トのビッ ト列における ビッ ト位置で表現され、 その場合、 特徵集合 { 2 , 3 } は、 " 0 0 0 1 1 0 " と表現される。
図 6に示す、 特徴集合 T Xとカテゴリ一名 L Xとを対応させた辞 書は、 分割領域数 = 6、 カテゴリ一数 = 1 0の例を示すものである, 例えば、 特徴集合 TX = 1 ("000001") にはカテゴリ一名 L X = 2 ("0000000010") が対応し、 この場合、 1個のカテゴリ一名が対応 することになる。 特徴集合 TX = 2 ("000010") にはカテゴリ一名 L X = 5 ("0000000101") が対応し、 " 1 " のビッ ト位置で示され る二つのカテゴリ一名が対応することになる。
なお、 カテゴリ一数が多く、 すなわちカテゴリー名 L Xのビッ ト 数が多くなる場合は、 各種のデータ処理に適用されているデータの 圧縮 復元技法を適用することができる。 例えば、 図 7に示すよう に、 L Xデ一夕を並べ換え、 重複するものを削除することにより圧 縮して、 新たな L Xデータを作成する。 そして、 Τ Χの値と L Xデ 一夕へのポイ ンタ情報とを組とするテーブルを作成する。 また、 認 識処理に差し支えなければ、 これらのテーブルをさらに圧縮しても よい。
また、 特徴集合 ΤΧをァ ドレスによって表現することができる。 その場合、 例えば、 図 6における Τ X = し 2 , 3 , 5, ···, 6 3 は、 ア ドレス 1 , 2 , 3 , 5 , ···, 6 3 として表現され、 ア ドレス 1 には L X = 2、 ア ドレス 2には L X = 5 というように格納し、 ま た、 ア ドレス 4 には対応するカテゴリ一名 L Xがないため、 ァ ドレ ス 4には L X = 0を格納する。
次に、 特徴ベク トルの具体的な作成手順について説明する。 図 8 は、 特徴べク トル作成の処理手順を示す概略フローチヤ一トである, この特徴べク トル作成ルーチンは、 後に説明する辞書作成処理及び 入力パターン認識処理において呼び出されるものである。 なお、 特 徵べク トル作成処理は、 個々のパターン認識システムの用途等に応 じて各種の変形例を採用できるものであり、 以下の実施例は、 代表 的なものを示す。
まず、 パターンデータについて重みづけを行うか否かを判定する (ステップ i 0 2 ) 。 重みづけを行う場合には、 パターンデータの ドッ トの重みづけ処理 (ステップ 1 0 4 ) を行ってから、 その重み づけ処理された ドッ トについての計測処理を行う (ステップ 1 0 6 ) また、 重みづけを行わない場合には、 直ちに、 パターンデータの ド ッ トの計測処理を行う (ステップ 1 0 6 ) 。 以上が特徴べク トル作 成処理の概略である。 次に、 重みづけ処理及び計測処理の詳細につ いて説明する。
図 9〜 1 3は、 重みづけ処理のフローチャー トを示し、 図 8のス テツプ 1 0 4 における処理の一例である。 まず、 変数 I (認識領域 内の ドッ ト位置を指定する) に 0を代入する (ステップ 2 0 2 ) 。 次いで、 入力バタ一ンの認識領域の横の ドッ ト数を X、 縦の ドッ ト 数を Yとして、 Iが X X Yより小さいか否か、 すなわち I が認識領 域内の最終ドッ ト位置以下を指しているか否かを判定する (ステツ プ 2 0 4 ) 。 I ≥ X X Yの場合にはステップ 2 1 0に移行し、 I く X x Yの場合はステップ 2 0 6に移行する。
ステップ 2 0 6においては、 入力パターン P 〔 I 〕 に所定の中間 値 MIDDLEVAL を乗算して、 重みづけされたパターン Q 〔 I〕 を作成 する。 なお、 この入力パターン P 〔 I〕 は、 例えば、 白 ドッ トを
" 0 " 、 黑ドッ トを " 1 " としたものである。 したがって、 中間値 を " 1 6 " とすると、 黒の " 1 " のドッ トが " 1 6 " の値に重みづ けされる。 次いで、 I をインク リ メ ン トして (ス.テツプ 2 0 8 ) 、 ステップ 2 0 4 にループバッ クする。 この中間値は、 白 ( " 0 " ) の領域が重みづけの結果、 最終的に負の値となることを回避するよ うに選定することが演算処理上望ましく、 本実施例においてもその ために " 1 6 " に設定される。 しかし、 これに限定きれるものでは ない。
全ての Iすなわち全ドッ トについてステップ 2 0 6の処理が終了 し、 ステップ 2 0 4 において I = X X Yが成立すると、 ステップ 2 1 0 (図 1 0 ) に移行する。 例えば、 図 1 4 Aに示すような入力パ ターン P 〔 I 〕 (X = 9 , Y= l l すなわち 1 = 0〜9 8 ) が入力 されたときには、 ステップ 2 1 0実行時点において図 1 4 Βに示す パター ン Q 〔 I 〕 が作成されている。 なお、 図 1 4 Αにおいて、
" 1 " は黒ドッ ト、 "0 " は白 ドッ トを示す。
ステップ 2 1 0においては、 中間値 MIDDLEVAL を変数 Jに代入す る。 次いで、 変数 I に 0を代入し、 所定のフラグ F L Gを 0に設定 する (ステップ 2 1 2 ) 。 次いで、 I と X X Yとを比較する (ステ ップ 2 1 4 ) 。 I = X x Yの場合にはステップ 2 3 8に移行し、 I く X X Υの場合にはステップ 2 1 6に移行する。 ステップ 2 1 6で は、 Q 〔 I 〕 と 0 とを比較し、 Q 〔 I 〕 ≠ 0の場合にはステップ 2 3 6 に移行し、 Q 〔 I 〕 = 0の場合にはステップ 2 1 8に移行する。 ステップ 2 1 8では、 Ι — Χと 0 とを比較し、 I — Xく 0の場合 にはステップ 2 2 2に移行し、 I — X≥ 0の場合にはステップ 2 2 0に移行する。 ステップ 2 2 0では、 Q 〔 Ι — Χ〕 と J とを比較し、 Q C I - X) = Jの場合にはステップ 2 3 4 に移行し、 Q C I - X) ≠ Jの場合にはステップ 2 2 2に移行する。 すなわち、 処理中の ド ッ トの値 Q 〔 I 〕 が 0でその一つ上の ドッ トの値 Q 〔 I 一 X〕 が J に等しいときには、 Q 〔 I〕 を変更すべくステップ 2 3 4に移行す ステップ 2 2 2では、 I + Xと X x Yとを比較し、 I + X≥ X x Υの場合にはステップ 2 2 6に移行し、 Ι + Χく Χ Χ Υの場合には ステップ 2 2 4 に移行する。 ステップ 2 2 4では、 Q 〔 Ι + Χ〕 と J とを比較し、 Q 〔 I + X〕 = Jの場合にはステップ 2 3 4に移行 し、 Q 〔 I + X〕 ≠ Jの場合にはステップ 2 2 6に移行する。 すな わち、 処理中の ドッ トの値 Q 〔 I 〕 が 0でその一^ D下の ドッ トの値 Q 〔 I + X〕 が Jに等しいときには、 Q 〔 I〕 を変更すべく ステツ プ 2 3 4に移行する。
また、 ステップ 2 2 6 (図 1 1 ) では、 I ZXの余りすなわち I %Xと 0 とを比較し、 余りが 0の場合すなわち I %X = 0の場合に はステップ 2 3 0に移行し、 余りが 0でない場合すなわち I % X≠ 0の場合にはステップ 2 2 8に移行する。 なお、 " " は、 C言語 における剰余 (モジュロ) 演算子である。 ステップ 2 2 8では、 Q 〔 1 — 1〕 と Jとを比較し、 Q 〔 I 一 1〕 = Jの場合にはステップ 2 3 4に移行し、 Q 〔 1 — 1〕 ≠ Jの場合にはステップ 2 3 0に移 行する。 すなわち、 処理中の ドッ トの値 Q 〔 I〕 が 0でその左隣の ドッ トの値 Q 〔 I 一 1〕 が Jに等しいときには、 Q 〔 I〕 を変更す ベく ステップ 2 3 4に移行する。
ステップ 2 3 0では、 I ZXの余りすなわち I %Xと X— 1 と比 較し、 I %X = X— 1の場合にはステップ 2 3 6に移行し、 I %X ≠X- 1の場合にはステップ 2 3 2に移行する。 ステップ 2 3 2で は、 Q 〔 1 + 1〕 と Jとを比較し、 Q 〔 1 + 1〕 = Jの場合すなわ ち Q 〔 I〕 の右隣が Jと等しい場合にはステップ 2 3 4に移行し、 Q 〔 I + 1〕 ≠ Jの場合にはステップ 2 3 6に移行する。 すなわち. 処理中の ドッ トの値 Q 〔 I〕 が 0でその右隣の ドッ トの値 Q C I + 1〕 が Jに等しいときには、 Q 〔 I〕 を変更すべくステップ 2 3 4 に移行する。
ステップ 2 3 4においては、 Q 〔 I〕 に J一 1を代入し、 かつ、 フラグ F L Gを " 1 " にセッ トする。 これは、 入力パターンの白と 黒との境界線に隣接した白 ドッ 卜の重みづけを J— 1の値 (最初は. 1 6 - 1 = 1 5 ) とすることを意味する。
また、 ステップ 2 3 6では、 次の ドッ トに関して上述の処理を実 行すべく、 変数 Iの値をインク リ メ ン トしてステップ 2 1 4にル一 プバッ クする。
全ての ドッ トに関して以上の処理が完了し、 ステップ 2 1 4にお いて I = X X Yが成立すると、 ステップ 2 3 8に移行する。 前述の 図 1 4 Bのパターンは、 第 1回目のステップ 2 3 8実行時点におい て図 1 4 Cに示すように変化しており、 境界線に隣接した白 ドッ ト 部分には "1 5 " がセッ トされている。
ステップ 2 3 8では、 フラグ F L Gが "0 " か否かを判定する。 フラグが "1 " である場合には、 Jをデク リ メ ン ト して (ステップ 2 4 0 ) 、 ステップ 2 1 2にループバッ クする。 ステップ 2 3 8に おいてフラグ F L Gが " 0 " であることは、 前述の図 1 4 Cを例に して言えば、 図 1 4 Dまで処理がなされたことを意味する。 そして、 その場合には、 ステップ 2 4 2に移行する。
ステップ 2 4 2 (図 1 2 ) においては、 変数 Jに中間値 MIDDLEVAL を代入する。 次のステップ 2 4 4では、 変数 Iに横の ドッ ト数 Xを 代入するとともに、 フラグ F L Gを "0" とする。 次いで、 ステツ プ 2 4 6においては、 I と X X Y— Xとを比較し、 I =X xY— X の場合にはステップ 2 6 6に移行し、 I く XxY— Xの場合にはス テツプ 2 4 8に移行する。 ステップ 24 8では、 Q 〔 I〕 と Jとを 比較し、 Q 〔 I〕 = Jの場合にはステップ 2 5 0に移行し、 Q 〔 I〕 ≠ Jの場合にはステップ 2 6 4に移行する。
ステップ 2 5 0では、 Q 〔 I—X〕 と Jとを比較し、 Q 〔 I—X〕 く Jの場合にはステップ 2 6 4に移行し、 Q 〔 I一 X〕 ≥ Jの場合 すなわち Q 〔 I〕 の一つ上の ドッ トが変数 Jと等しい場合にはステ ップ 2 5 2に移行する。
ステップ 2 5 2では、 Q 〔 I +X〕 と Jとを比較し、 Q 〔 I +X〕 < Jの場合にはステップ 2 6 4に移行し、 Q 〔 I +X〕 ≥ Jの場合 すなわち Q 〔 I〕 の一つ下の ドッ トが変数 Jと等しい場合にはステ ップ 2 5 4に移行する。
ステップ 2 5 4では、 I /Xの余りすなわち I % Xと 0とを比較 し、 I %X= 0の場合にはステップ 2 6 4に移行し、 I %X≠ 0の 場合にはステップ 2 5 6に移行する。 ステップ 2 5 6 (図 1 3) で は、 Q 〔 1— 1〕 と Jとを比較し、 Q C ί - Π く Jの場合にはス テツプ 2 6 4に移行し、 Q 〔 I一 1〕 ≥ Jの場合すなわち Q 〔 I〕 の左隣が Jと等しい場合にはステツプ 2 5 8に移行する。
ステップ 2 5 8では、 I ZXの余りすなわち I %Xと X— 1 とを 比較し、 I %X = X— 1の場合にはステップ 2 6 4に移行し、 1 % X≠ X - 1の場合にはステツプ 2 6 0に移行する。 ステップ 2 6 0 では、 Q 〔 1 + 1〕 と Jとを比較し、 Q 〔 1 + 1〕 く Jの場合には ステップ 2 6 4に移行し、 Q 〔 1 + 1〕 ≥ Jの場合すなわち Q 〔 I〕 の右隣が Jと等しい場合にはステップ 2 6 2に移行する。
ステップ 2 6 2では、 Q 〔 I〕 並びにその一つ上、 一- 5下、 左隣 及び右隣の全てが Jに等しいため、 Q 〔 I〕 に J+ 1を代入する。 すなわち、 変数 Jの値が 1 6 (最初の値) のときには、 Q 〔 I〕 に 1 7を代入する。 そして、 フラグ F L Gを 1 とし、 ステップ 2 6 4 に移行する。
ステップ 2 6 4では、 次のドッ トに関し同様の処理を実行すベく、 変数 Iをイ ンク リ メ ン トして、 ステップ 2 4 6にループバッ クする。
ステップ 2 4 6において最初に I = X X Y— Xが成立するときに は、 前述の図 1 4 Dを例にして言えば、 図 1 4 Eに示すパターンに まで重みづけ処理が進行している。
ステップ 2 6 6では、 フラグ F L Gが "0 " か否かを判定し、 "0" でない場合には Jをインク リ メ ン トして (ステップ 2 6 8 ) 、 ステップ 2 4 4にループバッ クする。 また、 ステップ 2 6 6でフラ グ F L Gが "0 " である場合には、 前述の図 1 4 Eのパターンが最 終的な図.1. 4 Fのパターンにまで処理されたことを意味し、 本重み づけ処理を終了する。 このようにして、 入力パターン P 〔 I〕 (図 1 4 A) から、 白と黒との境界線からの距離に従って重みづけされ たパターン Q 〔 I〕 (図 1 4 F) を得ることができる。
図 1 5は、 図 1 4 Aより も ドッ ト数の多い更なる入カバターンの 例であり、 その図においては、 英字 Tに相当する入力パターン P
〔 I〕 が示されている。 この入力パターン P 〔 I〕 について前述の 重みづけ処理 (中間値 MIDDLEVAL を 1 6 とする) を実施すると、 処 理結果は、 図 1 6に示すものとなる。 すなわち、 入力パターン P 〔 I〕 の "1 " の領域の境界からの距離に応じて 1 7, 1 8 , 1 9 の重みづけが行われ、 また、 "0 " の領域の境界からの距離に応じ て 1 5, 1 4 , ···, 7 , 6の重みづけが行われることになる。
次に、 入力パターン又は重みづけ処理後のパターンから特徴べク トルを求めるために実施する、 分割領域内のドッ トの計測処理 (図 8のステップ 1 0 6 ) について、 その処理手順を詳細に説明する。 この計測処理は、 認識対象領域を複数に分割することにより作成さ れた分割領域内の黒の ドッ ト数又は重みづけされた ドッ トの値の加 算処理を行う ものである。
なお、 横の ドッ ト数 Xを横の分割数 VXで割ったときに整数と成 らない場合、 及び縦のドッ ト数 Yを縦の分割数 VYで割ったときに 整数とならない場合については、 境界線にかかる ドッ トの値を面積 に比例して各領域に配分する。 そして、 その計算においては、 小数 点以下の計算をしなくても済むように、 境界線上のドッ トの値を面 積比に応じて除算する代わりに、 境界線上にないドッ トの値を VX XVY倍して数えるようにする。 その結果、 求められた特徴べク ト ル Vの値は、 VX XV Y倍されているが、 大小関係だけがわかれば よいので、 そのまま使用することができる。 例えば、 図 1 4 Fのパ ターン (X= 9 , Y= l l ) を VX= 2, VY= 3 として,分割する 場合には、 図 1 7に示すようなデータについて加算処理を実行する こととなり、 図 1 8に示される特徴べク トルが得られる。
そのような計測処理の具体的な処理手順は、 図 1 9及び 2 0のフ ロ ーチャー トに示される。 まず、 変数 Iを 0 とし (ステップ 3 0 2 ) I と N (N =特徴べク トルの次元) とを比較し (ステップ 3 0 4 ) 、 I = Nの場合にはステップ 3 1 0に移行し、 I く Nの場合にはステ ップ 3 0 6に移行する。 ステップ 3 0 6では、 特徴べク トル Vの要 素 V 〔 I〕 に 0を代入し、 次のステップ 3 0 8では、 Iをインク リ メ ン トしてステップ 3 0 4にループバッ クする。 すなわち、 ステツ プ 3 0 4 , 3 0 6及び 3 0 8により、 特徴べク トル Vの各要素 V 〔 I〕 を初期状態の 0 とする。
ステップ 3 1 0では、 変数 Jを 0 とする。 次のステップ 3 1 2で は、 J = Yx VYか否かを判定する。 なお、 VYは認識領域の縦の 分割数、 VXは横の分割数を示し、 従って、 特徴べク トルの次元 N は、 N = VX xVYとなる。 J = YXVYの場合には、 本計測処理 を終了する。 J <Y xVYの場合には、 変数 Iを 0 とし (ステップ 3 1 4 ) 、 1 = か否かを判定し (ステップ3 1 6 ) 、 1 = X xVXの場合にはステップ 3 2 2に移行し、 I く X XVXの場合 はステップ 3 1 8に移行する。
ステップ 3 1 8では、 式
V[(J/Y)*VX+ I/X]*-V[(J/Y)*VX+ I/X] + Q[(J/VY)*X+ I/VX] による演算を実行する。 これは、 前述のように、 横の ドッ ト数 Xを 横の分割数 VXで割った時に整数とならない場合、 及び縦の ドッ ト 数 Yを縦の分割数 VYで割った時に整数とならない場合について考 慮して加算処理を行う ものである。
そして、 ステップ 3 2 0では、 Iをインク リ メ ン トしてステップ 3 1 6にル一プバッ クする。 また、 ステップ 3 2 2では、 Jをイ ン ク リ メ ン ト して、 ステップ 3 1 2にループバッ クする。
以上の処理により、 分割数 VX, V Yで X X Yの認識領域內を分 割した場合の各分割領域における ドッ ト数又は重みづけされた ドッ トの値が加算されることとなる。 例えば、 図 1 4 Fのパターン (X = 9 , Y= l 1 ) を VX = 2, VY= 3 として分割する場合には、 図 1 7に示すようなデータについての加算処理が実行されて、 その 結果、 図 1 8 に示される特徴ベク トルが得られることが容易に理解 されるであろう。
次に、 前述の特徴べク トル作成処理を使用して辞書を作成する手 順について説明する。 図 2 1〜 2 3は、 かかる辞書作成処理の手順 を示すフローチャー トである。 なお、 特徴べク トル Vの次元数は N であり、 本ルーチンは、 図 6に示すように、 特徴集合 TXとカテゴ リー名 L Xとからなるレコー ドを複数個格納した辞書を作成するも のである。
まず、 レコー ドのア ドレスを指定するカウンタ C NTをク リアし て " 0 " とする (ステップ 4 0 2 ) 。 次に、 パターンファイル (例 えば、 ハー ドディスク 1 5に格納されている) をオープンする (ス テツプ 4 0 4 ) 。 次に、 そのファイル内の全てのパターンデータに ついて処理を完了したか否かを判定する (ステップ 4 0 6 ) 。
完了していない場合には、 手本パターンのパターンデータを 1 つ 取り出す (ステップ 4 0 8 ) 。 そして、 所定の変数 C O D Eに、 当 該カテゴリ一名と 1対 1 に対応する、 0〜: L— 1 の内の 1つの数値 をセッ トする (ステップ 4 1 0 ) 。 ここで、 Lはカテゴリ一名の数 である。 そして、 前述の手順により、 N個の要素からなる特徴べク トル Vの作成処理を行う (ステップ 4 1 2 ) 。 手本パターンの特徴 べク トルが作成されると、 特徵集合 Tに初期値の 0を代入し (ステ ップ 4 1 4 ) 、 変数 Iを 0 とする (ステップ 4 1 6 ) 。
次いで、 I と N— 1 とを比較し (ステップ 4 1 8 ) 、 I =N— 1 の場合 (すなわち、 後述するステップ 4 3 4における Iのインク リ メ ン トの結果、 I =N— 1 となった場合) にはステップ 4 0 6にル ープバッ ク し、 I く N— 1の場合にはステップ 4 2 0に進む。
ステップ 4 2 0では、 変数 MAX Jを 0 とし、 かつ、 Jを 1 とし てステップ 4 2 2へ進む。 ステップ 4 2 2では、 Jと Nとを比較し、 J =Nの場合にはステップ 4 3 0に移行し、 J <Nの場合にはステ ップ 4 2 4に移行する。 ステップ 4 2 4においては、 特徴べク トル Vの要素 V 〔J〕 と今までの最大値 V (MAX J) とを比較し、 V 〔J〕 > V CM A X J ) の場合にはステップ 4 2 6に進み、 V 〔J〕 ≤ V 〔M A X J〕 の場合にはステップ 4 2 8に進む。 ステップ 4 2 6においては、 今回の要素 V 〔J〕 が前回までの最大値 V CMAX J〕 より大きいため、 変数 MAX Jに現在の Jの値を代入し、 ステ ップ 4 2 8に進む。 ステップ 4 2 8では、 Jをインク リ メ ン ト して ステップ 4 2 2にループバッ クする。
ステップ 4 3 0においては、 次に最大の要素を探すときに今回検 索された最大の要素が検出されないように V 〔MAX J〕 に一 1を 代入する。 なぜならば、 本実施例の特徴ベク トル作成処理において は、 前述したように、 べク トルの要素が負の値をとらないようにさ れているからである。 また、 特徴集合 Tの第 MAX J番目のビッ ト を 1 とする。 ただし、 最下位ビッ ト (L S B) をビッ ト 0 とする。 この処理は、 C言語のビッ ト処理用演算子の 1つである左シフ ト演 算子 " < < " を使用すれば、
T T+ ( 1 < <MAX J)
で表される。
次のステップ 4 3 2においては、 こう して求められた特徴集合 T を特徴集合テーブル TX 〔CNT〕 の内容とする。 また、 第 C OD E番目のビッ トを 1 とし (ただし、 最下位ビッ ト ( L S B ) をビッ ト 0 とする。 ) 、 他のビッ トを 0 としたビッ ト列をカテゴリー情報 テーブル L X CCNT) の内容とする。 これらの処理は、 下の式で 示される。
TX 〔CNT〕 — T
L X 〔CNT〕 - 1 くく C OD E
そして、 カウンタ CNTをインク リ メ ン ト してステップ 4 3 4に進 む。 ステップ 4 3 4では、 変数 Iをインク リ メ ン ト してステップ 4 1 8にループノくッ クする。
すなわち、 1つのパターンデータについて、 大きいものから順に 特徴べク トル Vの要素を見つけ出しては、 特徴集合とカテゴリ一情 報とからなる 1 レコー ドを作成する処理を繰り返し、 計 (N— 1 ) のレコー ドを作成する。 そして、 1つのパターンデータについて処 理が完了するとステップ 4 0 6に戻って、 まだ処理すべきパ夕一ン が存在するかを確認するのである。
パターンファイルに格納された全パ夕一ンデータについて前述の 処理が終了すると、 パターンファイルをクローズする (ステップ 4 3 6 ) 。 次いで、 特徴集合 TXとカテゴリー情報 L Xとからなるテ —ブルをソー トする (ステップ 4 3 8 ) 。 ここで、 ソー ト処理には, いわゆる度数ソー トを使用することが処理速度の面からは望ま しい, 度数ソー トを使用すると、 辞書作成時間がパターン数に比例した時 間で済むからである。 もちろん、 その他のソー ト法を使用しても差 し支えない。 次いで、 今回の辞書作成処理が辞書ファイルの新規作 成に関するものか、 又は既作成の辞書へのデータ追加に関するもの か、 を判定する (ステップ 4 4 0 ) 。 新規作成の場合には、 メモリ 上の内容をフ ァイルに書き出す (ステップ 4 4 2 ) 。 その際、 同一 の特徴集合に関しては、 図 6に示されたマージ処理を同時に実行す る。 また、 追加の場合には、 その辞書とメメモリ内容とをマージし ながらファイルに書き出す (ステップ 4 4 4 ) 。 以上の処理により パターン認識用の辞書が作成される。
次に、 図 2 4及び 2 5の'フローチャー トに基づいて、 入力パター ン認識処理の手順について説明する。 まず、 入力パターンについて 前述の辞書作成における場合と同様な特徴べク トルを作成する (ス テツプ 5 0 2 ) 。 そして、 変数 Iを 0 とし (ステップ 5 0 4 ) 、 I = L (こ こで、 L =カテゴリ一名の数) か否かを判定する (ステツ プ 5 0 6 ) 。 I =Lの場合にはステップ 5 1 2に移行する。 I ≠ L の場合には、 辞書内の当該カテゴリ一名 (変数 Iによって識別され る) と特徴集合との類似度 S C OR E S 〔 I〕 を初期値の 0 とし (ステップ 5 0 8 ) 、 Iをインク リ メ ン トして (ステップ 5 1 0 ) - ステップ 5 0 6にループバッ クする。 すなわち、 ステップ 5 0 4〜 5 1 0により、 類似度 S C OR E S C n ( I = 0〜L— 1 ) が初 期化される。
ステップ 5 1 2では特徴集合 Tを初期値の 0 とし、 次のステップ 5 1 4では変数 Kを 0 としてステップ 5 1 6に移行する。 ステップ 5 1 6では、 K = N— 1か否かを判定し、 K = N— 1の場合にはス テツプ 5 3 4に移行する。 また、 K <N— 1の場合は、 ステップ 5 1 8に移行する。
ステップ 5 1 8〜 5 2 8は、 前述の辞書作成処理におけるステツ プ 4 2 0〜4 3 0 と同一であり、 これらのステップで特徴集合丁が 求められる。 そして、 後述する辞書サーチ処理を行い (ステップ 5 3 0 ) 、 Kをイ ンク リ メ ン ト して (ステップ 5 3 2 ) 、 ステップ 5 1 6にループバッ クする。
ステップ 5 3 4では、 類似度 S C 0 R E S 〔 I〕 が最大となる I を求め、 その Iに対応するカテゴリ一名を入力パターンのカテゴリ 一名と判定して、 当該入力パターンの認識処理を終了する。
図 2 6及び 2 7は、 前述の入力パターン認識処理におけるステツ プ 5 3 0で実行される辞書サーチ処理の手順を示すフローチヤ一ト である。 まず、 ステップ 6 0 2において、 変数 I S T A R Tに 0、 変数 I E NDに所定値 T B LMAXを代入することにより初期設定 をする。 なお、 所定値 TB LMAXは、 辞書内のレコー ド数を示す < 次のステップ 6 0 4では、 I S T A R Tと I E NDとを比較し、 I S TART= I ENDの場合には辞書サーチ処理を終了し、 I ST A R T≠ I ENDの場合にはステップ 6 0 6に移行する。
この実施例におけるサーチ処理は、 2分割法に従っており、 ステ ップ 6 0 6においては、 変数 I Wに ( I START+ I END) , 2を代入する。 すなわち、 スター トア ドレスとエン ドア ドレスとの 和の 1 2を中間ア ドレス I Wとする。 次のステップ 6 0 8では、 特徴集合 Tと TX 〔 I W〕 とを比較し、 T = TX 〔 I W〕 の場合に はステップ 6 1 4へ移行し、 Tく TX 〔 I W〕 の場合にはステップ 6 1 0へ移行し、 T>TX 〔 I W〕 の場合にはステップ 6 1 2へ移 行する。
ステップ 6 1 0においては、 I E N Dに I Wを代入してステップ 6 0 4へループバッ クする。 また、 ステップ 6 1 2においては、 I STARTに I W+ 1を代入して、 ステップ 6 0 4へループバッ ク する。 ステップ 6 1 4においては Iに 0を代入して、 次のステップ 6 1 6へ移行する。
ステップ 6 1 6においては、 I と Lとを比較し、 I = Lの場合に は辞書サーチ処理を終了し、 I < Lの場合にはステップ 6 1 8へ移 行する。 ステップ 6 1 8では、 カテゴリ一名 L X 〔 I W〕 の第 I ビ ッ トが 1 となっているかをテス トする。 ただし、 最下位ビッ トを第 0 ビッ ト.とする。 換言すれば、 L X 〔 I W〕 と 1 くく I との論理積 L X 〔 I W〕 & 1 くく Iが 0か否かをテス トする。 当該論理積が 0 であるときにはステップ 6 2 2へ移行し、 0でないときにはステツ プ 6 2 0へ移行する。
ステップ 6 2 0においては、 類似度 S C O R E S 〔 I 〕 を + 1 し て、 ステップ 6 2 2へ移行する。 ステップ 6 2 2では、 I をインク リ メ ン ト してステップ 6 1 6ヘループバッ クする。 以上が辞書サー チ処理の手順である。
ここで、 特徴集合と類似度とについて詳細に解説する。 カテゴリ 一名ごとの一つの手本パターンの特徴集合列、
T > (C) , ···, ΤΝ-, (C)
(但し、 l ≤ c ≤ L、 L =カテゴリ一名の総数)
を格納したパターン認識用の辞書を用いるとする。 特徵集合、
T , * , …, ΤΝ-, *
を有する入力パターンと、 c番目のカテゴリ一名の手本パターンと の類似度 S (c) は、
S (c) = [ (π-Ι^ Σ , - ,
Figure imgf000030_0001
)) … ( 5 ) ここで、 φ (T Ti) = 0 ( ≠ Τ2 のとき)
Φ (Τι,Τ2) = 1 ( Τ, = Τ2 のとき)
Σ , Ν-1 は、 I = 1 から I = Ν— 1 までの累算 と表される。
これを、 各カテゴリ一名が、 異なる書体等に対応して、 複数の手 本パターンについてそれぞれ特徴集合列を持つ場合に拡張する。 す なわち、 あるカテゴリ一名 cについて m個の手本パターンが存在し 特徴集合のマ ト リ クス、 ) T c )
T ( c
12 ( c ) T (
.― 1
T i 1
TML ( c ) , T ( c ) T ( c ) が形成される場合、 類似度を次のように定義し、 これを [、巾空間類 似度」 と呼ぶこととする。
S (c) = [1/(N-1)]∑ N一 1
I - 1 d xmax { φ (T ,Τ ( c )
i I )} i 匪 1
( 6 ) 但し、 C , は定数 (通常は 1 )
1 は、 1 〜m
I は、 1 〜N— 1
max { } ,01は、 i は 1 〜mとしたときの
{ } 内の値の最大値
本発明は、 入力パターンから求めた特徴集合と、 文字認識用辞書 に格納された特徴集合とについて、 前述の ( 6 ) 式による巾空間類 似度を計算し、 最大の類似度となる手本パターンのカテゴリ一名が 入力パターンのカテゴリ一名である、 と判定するものである。
例えば、 未知のカテゴリ一名の入力パターン P X 1 , P X 2につ いて、 図 5に示すパターン認識用の辞書を用いて認識する場合を考 える。 入力パターン P X 1 , P X 2の特徴ベク トル VX 1 , VX 2 が、
V X 1 = ( 6 , 8 8 8 9 9 9 9 , — 5 5, 7 7 , - 4 4 4 )
V X 2 = ( 2 5 , 1 6 3 4, 6 1 , 5 2 , 4 3 )
であるとする。 入力バタ ン P X 1 の特徴べク トル VX 1 から作成 される特徴集合列は、
{3} , {2, 3} , {2,3, 5} , {1,2,3,5} , {1,2,3,4,5}
となる。 入力パターン P X 2の特徴ベク トル VX 2から作成される 特徵集合列は、 ,
{4} , {4, 5} , {4, 5, 6} , {3, 4, 5, 6} , {1, 3, 4, 5, 6}
となる。
入力パターン P X 1 の特徴集合と図 5の辞書の特徴集合とをみる -と、 入力パターン P X 1 の各特徴集合について次のことがわかる。 すなわち、
{ 3 } は、 カテゴリ一名 A, Bにリ ンクする。
{ 2 , 3 } は、 カテゴリ一名 Aにリ ンクする。
{ 2 , 3 , 5 } は、 何れのカテゴリ一名にも リ ンク しない。
{ 1 , 2 , 3 , 5 } は、 カテゴリ一名 Aにリ ンクする。
{ 1 , 2, 3, 4, 5 } は、 カテゴリ一名 Aにリ ンクする。
従って、 前述のリ ンク 1 回について 1 5点を与えることとする と、 入力パターン P X 1 のカテゴリ一名 Aとの類似度は、 4 Z 5 と なる。 また、 入力パターン P X 1 のカテゴリ一名 Bとの類似度は 1 / 5 となる。
同様に計算して、 入力パターン P X 2のカテゴリ一名 Aとの類似 度は 3 / 5 となり、 入力パターン P X 2のカテゴリ一名 Bとの類似 度は 4ノ 5 となる。 従って、 入力パターン P X 1 はカテゴリ一名 A と認識し、 入力パターン P X 2はカテゴリ一名 Bと認識する。 なお- 類似度の順に認識候補カテゴリ一名を表示することも可能である。
また、 複数種類の類似度算出方法による類似度、 例えば、 A方法 と B方法とによる類似度 S A (c〉 , S B (c) を求め、 合成した類似 度 S (c) を、
S (c) = ( S A (c) + S B (c) ) / 2
により求めることも可能である。 また、 大分類、 中分類、 小分類等 の分類によって文字認識を行う場合、 前述の認識方法を任意の段階 の分類に適用することも可能である。 これは、 本発明が総てのカテ ゴリーについて類似度を算出しているからであり、 各種の応用が可 能である。
以上説明したように、 本発明は、 入力パターンの特徴集合と、 辞 書の特徴集合との 「巾空間類似度」 を求めて、 最大の類似度を有す るカテゴリ一名を入力パターンのカテゴリ一名と判定するものであ り、 認識率が高いことが確認されている。 また、 類似度計算におい ては、 辞書の各レコー ドが特徴集合の順番に配列され、 例えば 2分 割法、 ア ドレス表現法等によってサーチすることができる。 そのた め、 辞書の全部の特徴集合に対して類似度計算を行う ことなく、 そ の一部を参照すれば済むから、 認識所要時間を短縮できる利点があ る。 また、 類似度計算は、 全カテゴリ一名に対して行う ものである から、 パターン認識における大分類等に適用することができるなど 応用性にも優れているという利点がある。
実際に、 ドッ トプリ ンタにより、 半角文字のアルフアベッ トの大 小 5 2文字と、 数字 1 0文字とをプリ ン トアウ ト し、 分解能 3 0 0
( d p i ; ドッ ト インチ) のスキャナにより読み取って、 辞書作 成用の手本パターンと、 認識率測定用の入力パターンとした。 また. 辞書に使用した文字数は 6 2文字種 X 8 0セッ ト = 4 9 6 0文字、 認識率測定用に入力した文字数は 6 2文字種 X 4 0セッ ト = 2 4 8
0文字とした。 その結果、 辞書作成に要する時間は、 パーソナルコ ンピュー夕を使用した場合、 約 1 5 1秒、 認識率は 9 8 . 7 5 %で あった。 この場合、 大文字と小文字とが殆ど類似のパターンを有す ることに起因する誤認識が生じたが、 文字の大きさの情報を入力す ることにより、 認識率を 9 9 . 7 5 %とすることができた。
また、 J I S第 1水準の漢字 2 9 6 5文字種をレーザプリ ン夕で プリ ン トァゥ ト し、 前述の 3 0 0 d p i のスキャナにより読み取つ て辞書作成用の手本パターン及び認識率測定用の入力パターンとし た。 その場合、.惑識率は、 9 9 . 9 7 %であった。 この場合の 1文 字当たりの認識所要時間は約 0 , 0 1 9秒であった。 また、 0〜 9 及び Xの手書き文字について辞書作成及び認識率測定を行った。 そ の場合、 全文字数 1 0 9 9 4 に対して、 認識率は 9 8 . 8 6 %であ つた。 すなわち、 実用上充分な認識率を得ることができた。
また、 パターン認識用の辞書は、 手本パターンより N個の特徴量 を抽出して特徵ベク トルとし、 その特徴ベク トルより N— 1個の特 徵集合を形成し、 それらを手本パターンのカテゴリ一名と対応させ てメモリに格納するものであり、 一つの特徴集合に複数のカテゴリ —名を対応させることができる。 そのため、 手本パターン及びカテ ゴリ一名が多数の場合でも、 辞書作成に要する時間は短くて済み、 経済的に辞書を作成することができる。 さらに、 手本パターンを追 加する場合でも、 辞書の一部を修正して追加すれば良いから、 簡単 に追加することができるという利点がある。 産業上の利用可能性
本発明は、 特徴べク トル作成方法を適切に選ぶことによってあら ゆるパターン認識に応用することができる。 すなわち、 本発明は、 コ ンピュータの入力装置としての光学式文字読み取り装置 (O C R ) を初めとして、 医療診断システムや音声認識システム (音声波形を パターンとみなして利用する) にも適用可能であろう。 なぜならば、 辞書作成時間の短縮及び認識所要時間の短縮という効果は、 これま で説明したように、 特徴べク トル作成方法がどんなものであろう と も得られるからである。 また、 高い認識率が得られることは、 辞書 作成に使用されたパターンは必ず正解となること、 及び文字認識の 各種の事例 (手書き文字、 印刷文字、 数字、 英数字、 漢字等) にお いていずれも良好な認識率が実際に得られたことによって証明され ている

Claims

請 求 の 範 囲
1. (a) 入力パターンの認識対象領域を N個の領域に分割し、 そ れぞれに対応する分割領域番号を付与して、 N個の分割領域を作成 するステップと、
(b) 該 N個の分割領域の各々に対して、 所定の基準に従って特徴 量を算出し、 該 N個の特徴量を要素とする特徴べク トルを作成する ステップと、
(c) 該特徵べク トルの N個の要素の中の最大のもの又は最小のも のを検索し、 該検索された要素に対応する 1個の分割領域番号から なる第 1 の特徴集合を作成し、 次いで、 最大のもの又は最小のもの から順に 2個を検索し、 該検索された 2個の要素に対応する 2個の 分割領域番号の組み合わせからなる第 2の特徴集合を作成し、 以下 同様にして (N— 1 ) 個の分割領域番号の組み合わせからなる第
( N - 1 ) の特徴集合まで作成することにより、 合計 (N— 1 ) 個 の特徴集合を作成するステップと、
(d) 前記ステップ(a) から(c) までを予め各種の手本パターンに 対して実施して得られた特徴集合を該手本パターンのカテゴリ一名 とともに記録した辞書を参照することにより、 該入力パターンの特 徴集合と該辞書内の特徴集合との類似度をカテゴリー名対応に求め. 該類似度が最大となるカテゴリ一名を該入力パターンのカテゴリ一 名と判定するステップと、
を具備するパターン認識方法。
2. ステップ(a) における分割は、 手本パターンの認識対象領域 を N個の等面積の領域に分割するものであり、 かつ、 ステップ(b) における特徴量は、 分割領域内の ドッ トの値の合計に基づく もので ある、 請求の範囲第 1項に記載のパターン認識方法。
3. ステップ(a) における分割は、 手本パターンの認識対象領域 をそれぞれドッ トの値の合計が同一となるように N個の領域に分割 するものであり、 かつ、 ステップ(b) における特徴量は、 分割領域 の面積に基づく ものである、 請求の範囲第 1項に記載のパターン認
6 法。
4. 前記分割領域内の ドッ トの値は、 手本パターンの黒白境界線 からの距離に応じて重みづけされている、 請求の範囲第 2項に記載 のパターン認識方法。
5. 認識されるべきパターンを入力するパターン入力手段と、 入力されたパターンの認識対象領域を N個の領域に分割し、 それ ぞれに対応する分割領域番号を付与して、 N個の分割領域を作成す る手段と、
該 N個の分割領域の各々に対して、 所定の基準に従って特徴量を 算出し、 該 N個の特徴量を要素とする特徴べク トルを作成する手段 と、
該特徴べク トルの N個の要素の中の最大のもの又は最小のものを 検索し、 該検索された要素に対応する 1個の分割領域番号からなる 第 1 の特徴集合を作成し、 次いで、 最大のもの又は最小のものから 順に 2個を検索し、 該検索された 2個の要素に対応する 2個の分割 領域番号の組み合わせからなる第 2の特徴集合を作成し、 以下同様 にして (N— 1 ) 個の分割領域番号の組み合わせからなる第 (N— 1 ) の特徴集合まで作成することにより、 合計 (N— 1 ) 個の特徴 集合を作成する手段と、
前記パターン入力手段、 分割領域作成手段、 特徴べク トル作成手 段、 及び特徴集合作成手段を予め各種の手本パターンに対して適用 して得られた特徵集合を該手本パターンのカテゴリー名とともに記 録した辞書を記憶する手段と、 該入力パターンの特徴集合と該辞書内の特徴集合との類似度を力 テゴリ一名対応に求め、 該類似度が最大となるカテゴリ一名を該入 力パターンのカテゴリ一名と判定する手段と、
を具備するパターン認識装置。
6. ステップ(a) における分割は、 手本パターンの認識対象領域 を N個の等面積の領域に分割するものであり、 かつ、 ステップ(b) における特徴量は、 分割領域内のドッ トの値の合計に基づく もので ある、 請求の範囲第 5項に記載のパターン認識装置。
7. ステップ(a) における分割は、 手本パターンの認識対象領域 をそれぞれドッ トの値の合計が同一となるように N個の領域に分割 するものであり、 かつ、 ステップ(b) における特徴量は、 分割領域 の面積に基づく ものである、 請求の範囲第 5項に記載のパターン認
8. 前記分割領域内の ドッ トの値は、 手本パターンの黑白境界線 からの距離に応じて重みづけされている、 請求の範囲第 6項に記載 のパターン認識装置。
9. (a) 入力された手本パターンの認識対象領域を N個の領域に 分割し、 それぞれに対応する分割領域番号を付与して、 N個の分割 領域を作成するステップと、
(b) 該 N個の分割領域の各々に対して、 所定の基準に従って特徴 量を算出し、 該 N個の特徴量を要素とする特徴べク トルを作成する ステップと、
(c) 該特徴べク トルの N個の要素の中の最大のもの又は最小のも のを検索し、 該検索された要素に対応する 1個の分割領域番号から なる第 1 の特徵集合を作成し、 次いで、 最大のもの又は最小のもの から順に 2個を検索し、 該検索された 2個の要素に対応する 2個の 分割領域番号の組み合わせからなる第 2の特徴集合を作成し、 以下 同様にして (N— 1 ) 個の分割領域番号の組み合わせからなる第 ( N - 1 ) の特徴集合まで作成することにより、 合計 (N— 1 ) 個 の特徴集合を作成するステツプと、
(d) 該特徴集合の各々 と該手本パターンのカテゴリー名とを対応 させて所定のメモリに格納するステツプと、
を具備するパターン認識用辞書作成方法。
10. ステップ(a) における分割は、 手本パターンの認識対象領域 を N個の等面積の領域に分割するものであり、 かつ、 ステップ(b) における特徴量は、 分割領域内の ドッ トの値の合計に基づく もので ある、 請求の範囲第 9項に記載のパターン認識用辞書作成方法。
11. ステップ(a) における分割は、 手本パターンの認識対象領域 をそれぞれドッ トの値の合計が同一となるように N個の領域に分割 するものであり、 かつ、 ステップ(b) における特徴量は、 分割領域 の面積に基づく ものである、 請求の範囲第 9項に記載のパターン認 識用辞書作成方法。
12. 前記分割領域内の ドッ トの値は、 手本パターンの黒白境界線 からの距離に応じて重みづけされている、 請求の範囲第 1 0項に記 載のパターン認識用辞書作成方法。
PCT/JP1994/002160 1993-12-21 1994-12-20 Procede et appareil destines a la reconnaissance de structures et procede de compilation d'un dictionnaire en vue de la reconnaissance de structures WO1995017734A1 (fr)

Priority Applications (3)

Application Number Priority Date Filing Date Title
KR1019950703453A KR100242458B1 (ko) 1993-12-21 1994-12-20 패턴인식 방법 및 장치와 패턴인식 사전을 작성하는 방법
US08/500,995 US5689584A (en) 1993-12-21 1994-12-20 Method of and apparatus for pattern recognition and method of creating pattern recognition dictionary
EP95903008A EP0686933A4 (en) 1993-12-21 1994-12-20 METHOD RECOGNIZING METHOD AND DEVICE AND METHOD FOR LEXICON COMPILATION FOR PATTERN RECOGNITION

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP5344690A JP2937729B2 (ja) 1993-12-21 1993-12-21 パターン認識方法及び装置及び辞書作成方法
JP5/344690 1993-12-21

Publications (1)

Publication Number Publication Date
WO1995017734A1 true WO1995017734A1 (fr) 1995-06-29

Family

ID=18371229

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP1994/002160 WO1995017734A1 (fr) 1993-12-21 1994-12-20 Procede et appareil destines a la reconnaissance de structures et procede de compilation d'un dictionnaire en vue de la reconnaissance de structures

Country Status (6)

Country Link
US (1) US5689584A (ja)
EP (1) EP0686933A4 (ja)
JP (1) JP2937729B2 (ja)
KR (1) KR100242458B1 (ja)
CA (1) CA2156521A1 (ja)
WO (1) WO1995017734A1 (ja)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH10143613A (ja) * 1996-10-30 1998-05-29 Hewlett Packard Co <Hp> パタン認識方法
US6539115B2 (en) * 1997-02-12 2003-03-25 Fujitsu Limited Pattern recognition device for performing classification using a candidate table and method thereof
JP3639126B2 (ja) * 1998-01-22 2005-04-20 富士通株式会社 住所認識装置及び住所認識方法
US6320985B1 (en) * 1998-07-31 2001-11-20 International Business Machines Corporation Apparatus and method for augmenting data in handwriting recognition system
JP2001160067A (ja) * 1999-09-22 2001-06-12 Ddi Corp 類似文書検索方法および該類似文書検索方法を利用した推薦記事通知サービスシステム
US7095894B2 (en) * 2002-09-04 2006-08-22 Lockheed Martin Corporation Method and computer program product for recognizing italicized text
US20050004915A1 (en) * 2003-05-21 2005-01-06 Jean-Philippe Hermand Method for generating names
EP1574988B1 (en) * 2004-03-08 2014-06-18 Siemens Product Lifecycle Management Software Inc. Determining and using geometric feature data
KR100923935B1 (ko) * 2007-11-28 2009-10-29 엔에이치엔(주) Ocr을 위한 문서 영상의 자동 평가 방법 및 시스템
JP5789128B2 (ja) * 2011-05-26 2015-10-07 キヤノン株式会社 画像処理装置、画像データの処理方法およびプログラム
US9348808B2 (en) * 2011-12-12 2016-05-24 Empire Technology Development Llc Content-based automatic input protocol selection
US9881354B2 (en) * 2012-03-15 2018-01-30 Microsoft Technology Licensing, Llc Image completion including automatic cropping
US11423206B2 (en) * 2020-11-05 2022-08-23 Adobe Inc. Text style and emphasis suggestions

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPS6375989A (ja) * 1986-09-19 1988-04-06 Alps Electric Co Ltd 画像認識装置
JPS63106088A (ja) * 1986-10-22 1988-05-11 Ricoh Co Ltd 領域分割方法
JPS63234372A (ja) * 1987-03-24 1988-09-29 Oki Electric Ind Co Ltd 特徴抽出方式

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE3815869A1 (de) * 1987-05-08 1988-11-17 Ricoh Kk Verfahren zum extrahieren von merkmalsmengen eines zeichens
NL8802078A (nl) * 1988-08-23 1990-03-16 Philips Nv Werkwijze voor het herkennen van een patroon in een veld met een meerwaardige amplitude, inrichting voor het uitvoeren van een dergelijke werkwijze.
US5255342A (en) * 1988-12-20 1993-10-19 Kabushiki Kaisha Toshiba Pattern recognition system and method using neural network
US5293429A (en) * 1991-08-06 1994-03-08 Ricoh Company, Ltd. System and method for automatically classifying heterogeneous business forms
US5315668A (en) * 1991-11-27 1994-05-24 The United States Of America As Represented By The Secretary Of The Air Force Offline text recognition without intraword character segmentation based on two-dimensional low frequency discrete Fourier transforms

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPS6375989A (ja) * 1986-09-19 1988-04-06 Alps Electric Co Ltd 画像認識装置
JPS63106088A (ja) * 1986-10-22 1988-05-11 Ricoh Co Ltd 領域分割方法
JPS63234372A (ja) * 1987-03-24 1988-09-29 Oki Electric Ind Co Ltd 特徴抽出方式

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP0686933A4 *

Also Published As

Publication number Publication date
JPH07182453A (ja) 1995-07-21
EP0686933A4 (en) 1997-07-02
CA2156521A1 (en) 1995-06-29
EP0686933A1 (en) 1995-12-13
US5689584A (en) 1997-11-18
JP2937729B2 (ja) 1999-08-23
KR960701413A (ko) 1996-02-24
KR100242458B1 (ko) 2000-02-01

Similar Documents

Publication Publication Date Title
US5313527A (en) Method and apparatus for recognizing cursive writing from sequential input information
US5267332A (en) Image recognition system
US5029223A (en) Constraint driven-on line recognition of handwritten characters and symbols
CN113591457B (zh) 文本纠错方法、装置、设备及存储介质
WO1995017734A1 (fr) Procede et appareil destines a la reconnaissance de structures et procede de compilation d&#39;un dictionnaire en vue de la reconnaissance de structures
JPH05217024A (ja) 自動手書き文字認識に対する統計的混合手法
EP0689153A2 (en) Character recognition
JP2003524258A (ja) 電子ドキュメントを処理する方法および装置
Shirali-Shahreza et al. Recognition of handwritten Persian/Arabic numerals by shadow coding and an edited probabilistic neural network
Miller On-line recognition of hand-generated symbols
JPH06325212A (ja) 手書き文字認識方法および装置
US6859556B2 (en) Word recognizing apparatus for dynamically generating feature amount of word and method thereof
JP3299875B2 (ja) 文字処理装置
Wilkinson et al. Machine-assisted human classification of segmented characters for OCR testing and training
JP3925011B2 (ja) パターン認識装置及び認識方法
EP0519737A2 (en) Image recognition system
JP3128357B2 (ja) 文字認識処理装置
JP3109476B2 (ja) 候補数制御機能付き文字認識装置及び方法
JPH05233581A (ja) ニューラルネットワーク応用システム
JP2639314B2 (ja) 文字認識方式
JP3352467B2 (ja) オンライン文字認識装置
JPH0253193A (ja) 文字認識方法および文字認識装置
JPH0737037A (ja) 最近傍分類を認識する比較関数に基づくocr方法
JPH0338631B2 (ja)
JPH01156889A (ja) オンライン手書き文字認識方式

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): CA KR US

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): AT BE CH DE DK ES FR GB GR IE IT LU MC NL PT SE

WWE Wipo information: entry into national phase

Ref document number: 08500995

Country of ref document: US

WWE Wipo information: entry into national phase

Ref document number: 1995903008

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 2156521

Country of ref document: CA

121 Ep: the epo has been informed by wipo that ep was designated in this application
WWP Wipo information: published in national office

Ref document number: 1995903008

Country of ref document: EP

WWW Wipo information: withdrawn in national office

Ref document number: 1995903008

Country of ref document: EP