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 PDFInfo
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/18—Extraction of features or characteristics of the image
- G06V30/18086—Extraction of features or characteristics of the image by performing operations within image blocks or by using histograms
- G06V30/18095—Summing image-intensity values; Projection and histogram analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/191—Design or setup of recognition systems or techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
- G06V30/1912—Selecting the most significant subset of features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Character Discrimination (AREA)
- Image Analysis (AREA)
Description
Claims
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 |
EP95903008A EP0686933A4 (en) | 1993-12-21 | 1994-12-20 | METHOD RECOGNIZING METHOD AND DEVICE AND METHOD FOR LEXICON COMPILATION FOR PATTERN RECOGNITION |
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 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP5/344690 | 1993-12-21 | ||
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 |
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)
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 |
JP6033326B2 (ja) * | 2011-12-12 | 2016-11-30 | エンパイア テクノロジー ディベロップメント エルエルシー | コンテンツベースの自動的な入力プロトコルの選択 |
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)
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)
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 |
-
1993
- 1993-12-21 JP JP5344690A patent/JP2937729B2/ja not_active Expired - Fee Related
-
1994
- 1994-12-20 EP EP95903008A patent/EP0686933A4/en not_active Withdrawn
- 1994-12-20 KR KR1019950703453A patent/KR100242458B1/ko not_active IP Right Cessation
- 1994-12-20 CA CA002156521A patent/CA2156521A1/en not_active Abandoned
- 1994-12-20 US US08/500,995 patent/US5689584A/en not_active Expired - Fee Related
- 1994-12-20 WO PCT/JP1994/002160 patent/WO1995017734A1/ja not_active Application Discontinuation
Patent Citations (3)
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)
Title |
---|
See also references of EP0686933A4 * |
Also Published As
Publication number | Publication date |
---|---|
JPH07182453A (ja) | 1995-07-21 |
US5689584A (en) | 1997-11-18 |
EP0686933A1 (en) | 1995-12-13 |
JP2937729B2 (ja) | 1999-08-23 |
KR960701413A (ko) | 1996-02-24 |
EP0686933A4 (en) | 1997-07-02 |
KR100242458B1 (ko) | 2000-02-01 |
CA2156521A1 (en) | 1995-06-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US5029223A (en) | Constraint driven-on line recognition of handwritten characters and symbols | |
KR100297482B1 (ko) | 수기입력의문자인식방법및장치 | |
US20050074169A1 (en) | Holistic-analytical recognition of handwritten text | |
WO1995017734A1 (fr) | Procede et appareil destines a la reconnaissance de structures et procede de compilation d'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 | |
JPH0634185B2 (ja) | 入力語認識装置 | |
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 | |
JP3925011B2 (ja) | パターン認識装置及び認識方法 | |
JPH06103402A (ja) | 名刺認識装置 | |
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 | |
JP2976990B2 (ja) | 文字認識装置 | |
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 | |
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 | |
CA2497586C (en) | Method and apparatus for recognizing cursive writing from sequential input information |
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 |