JPS63150785A - Character recognition device - Google Patents

Abstract

PURPOSE:To identify and recognize a similar character with different size and position by providing a recognizing means recognizing a normalized pattern as a part of a recognizing parameter for the power variation ratio decided by a power variation means. CONSTITUTION:In receiving a document image data from an input section 11, a CPU 12 uses a processing program stored in a ROM 13 to segment a character while using a RAM 14 as a storage and an auxiliary storage of image data and the data is subjected to power variation by either an average power variation ratio or a power variation ratio corresponding to each pattern based on the power variation ratio of each pattern and the average power variation ratio of plural patterns. Then the normalized pattern is recognized by a dictionary corresponding to the power variation ratio in use in a recognition dictionary 15 for special characters or a recognition dictionary 16 for other character and the result is outputted from an output section 17 such as a display device or a storage device.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application in Industry] The present invention relates to a character recognition device, and particularly to a character recognition device that normalizes the size of characters before handing them over to the recognition process as preprocessing.

[Prior Art] Conventionally, in this type of device, the position of written characters is specified, and the size of the characters is limited to a certain size within a predetermined writing frame. . on the other hand,
When recognizing characters written at arbitrary positions in a free format, it is necessary to extract each character and to normalize the size of the extracted characters because the size of the extracted characters is undetermined. Generally, in order to normalize the size, the procedure is to calculate the scaling factor from the circumscribed frame of the character pattern cut out for each character to the recognition pattern, and apply the pattern. Further, in many cases, this magnification ratio is independent of the X direction of the character pattern and the X direction. In this case, for example, a character pattern such as "1" is normalized to -"mouth", or a pattern such as "O" or "0" where the X direction is slightly different from the X direction is completely ignored. Normalized to the same recognition pattern. It has also become clear that in such a case, the problem can be easily solved by changing the magnification in the X direction and in the X direction, that is, by performing normalization based on the smaller magnification.

However, the recognition of special characters, that is, (”, “・“,
”・“. When recognizing characters such as ``...'', the above normalization means may not be able to distinguish them, or the normalization may result in different characters.

[Problems to be Solved by the Invention] The present invention aims to eliminate the drawbacks of the above-mentioned conventional examples and provide a character recognition device that identifies and recognizes similar characters with different sizes and similar characters with different pattern positions. purpose.

[Means for Solving the Problems] In order to achieve the above object, the character recognition device of the present invention has the following configuration.

That is, in a character recognition device that recognizes arranged characters, a separating means separates patterns arranged in a predetermined direction, and the patterns separated by the separating means are calculated by calculating the scaling ratio of each pattern and the average of a plurality of patterns. and a scaling means for normalizing to a predetermined size by scaling at a scaling ratio determined based on the scaling ratio; and and recognition means for recognizing the pattern normalized by the multiplication means.

[Operations] In the above configuration, in the character recognition device that recognizes arranged characters, there is a separating means for separating patterns arranged in a predetermined direction, and a scaling ratio of each pattern for the patterns separated by the separating means. and a scaling means for normalizing to a predetermined size by scaling at a scaling ratio determined based on the average scaling ratio of the plurality of patterns; The scaling means operates to recognize a normalized pattern based on the dictionary.

FIG. 2(a) is a block diagram showing the configuration of a conventional character recognition device, and FIG. 2(b) is a processing flowchart of the conventional character recognition device.

A conventional character recognition device has an input section 21, a CPU 22, and an RO.
It is equipped with an M23, a RAM 24, a recognition dictionary 25, and an output section 26, and when a document is manually inputted from the human power section 21 using image data, the CPU 22 uses the RAM 24 as a memory for image data and as an auxiliary memory according to a processing program stored in a ROM 23. Character cutting, normalization, and recognition are performed, and the results are output from an output unit 26 such as a display device or a storage device.

The operation will be explained according to the flowchart shown in FIG. 2(b). In step 5200, the input document data from the human power section 21 is input manually. Here, input using a scanner. This scanner converts human text into image data, but generally when a sensor such as a CCD is used, the image data represents an analog value. In step 5201, this analog value is binarized (into a pattern of 0 and 1). The binarized image data is stored in the RAM 24 of the image memory, and characters are cut out in step 5202.

Cutting out characters will be explained according to the input document example shown in FIG. Image data is scanned in the X direction to obtain a density histogram HY in the X direction, and character strings are determined for each pulse waveform generated for each line. In Figure 3, ■~
■ is considered a character string. Furthermore, separate each string,
A density histogram Hx in the X direction is obtained for each.

FIG. 3 shows the density histogram HX in the 0 column,
As a result, characters such as ■...■ are cut out at one character position.

Figure 4 shows an example of characters cut out to the size of one character. As shown in FIG. 4, it can be seen that the widths in the X direction are all at the same interval. On the other hand, the width in the X direction changes depending on the horizontal size of the character. In other words, ", 'ki" is cut out almost exactly as the outer shape of the character, and "-" has almost the same outer shape as "bi" but is offset only to the center of the width in the X direction.

For "1", ".", etc., elongated data is cut out.

In step 5203, the characters cut out in step 5202 are smoothed. In particular, the focus here is on removing isolated images (noise removal). Step 5204
Then, as shown in FIG. 5, rectangular extraction of the outer shape of the character is performed. In step 5205, a scaling factor g for normalizing the size is determined from the magnification of the vertical and horizontal widths of the rectangular figure f (i) extracted in step 5204 and the size F of the image pattern to be characterized. calculated. For example, the rectangular figure f(i) in FIG. 5 obtained in step 5204
is made up of 32×32 pixels, and when the image pattern F in which features are depicted is 64×64, a magnification ratio g=2 is obtained.

By scaling f(i) by this scaling factor g, a normalized pattern F(i) (the final image pattern to be normalized) is created in step 5206. Step 5
In step 207, the characteristics of the normalized pattern F(i) are drawn, and in step 3208, identification is performed by comparison with parameters stored in the recognition dictionary 25 in advance, and in step 5209, the recognition result is output to the output unit 26. Ru. In this case, a plurality of candidate characters are often output.

FIG. 1(a) is a block diagram of a character recognition device according to the first embodiment, and FIGS. 1(b) and 1(c) are processing flowcharts of the first embodiment.

The character recognition device of the first embodiment includes an input section 11 and a CPU 1.
2, a ROM 13, a RAM 14, a recognition dictionary 15, a recognition dictionary 16, and an output section 17.When a document is input as image data from the human power section 11, the CPU 12 stores the image data in the RAM 14 according to the processing program stored in the ROM 13. Characters are extracted and normalized while being used as auxiliary memory. The recognition dictionaries 15 and 16 are selected based on the normalized pattern, character recognition is performed, and the results are output from an output unit 17 such as a display device or a storage device.

The operation will be explained according to the processing flowcharts shown in FIGS. 1(b) and 1(C). Step 5100 - Step 510
Step 3 performs the same processing as steps 3200 to 5203 described above. In step 5104, rectangular data f (1), f (2), . . . as shown in FIG. 5 are obtained from the cutout pattern data of n consecutive characters.
, obtain f(n).

In step 5105, the average value f'X of the X component and the average value t"y of the X component of the obtained rectangular data are calculated using the following equations. Note that fxHc) and fy(k) are f (k ).Hereinafter, the X component and
Since the processing is the same as that of the components, in FIG. 1(b), X and Y are written in the same way without distinction. In step 5106, the scaling factor g is calculated from the average value f' and the size F of the recognized character.
=F/f' is calculated.

In step 5107, the magnification of each character g' (k)=
Calculate F/f (k). Let k be the order of each character.

In step 5toa, a comparison is made between g and g'(k). This means that if the scaling factor g'(k) of each character is larger than the average scaling factor g plus a predetermined value A, an abnormally large scaling factor is executed by taking the average scaling factor in step 5109. It will not be done. For example, the cut out characters are given in the case of ".", ",", etc. in FIG. Note that processing may be performed by calculating the overall average value/magnification ratio without distinguishing between X and Y as in this example.

On the other hand, in step 5108, the scaling factor g'(k
) is less than the value obtained by adding the predetermined value A to the average scaling factor g, then the scaling factor g(k) of each character is used as in the conventional method.

The characters scaled in steps 5109 and 110 are used as a normalized pattern F (k) in step 5111, features are extracted in steps 3112 to 115, character recognition is performed, and the characters are output to the output unit 17.

In this case, in the block diagram of FIG. 1(a), if the recognition dictionary 15 for special characters is separately provided, the recognition dictionaries 15 and 16 are selected in step 5113, and the process proceeds to step 5109, the recognition dictionary is used after feature extraction. 15, the recognition time can be significantly reduced. That is, in the first embodiment, the recognition dictionary 15 is a special character recognition dictionary, and the recognition dictionary 16 is for other characters. Furthermore, since the special character recognition dictionary 15 is provided, poor recognition of character data can be avoided in special character recognition processing.

In addition, lowercase alphabets and consonants (ya, yu,...
·) etc. can be distinguished, so if these characters are stored in the recognition dictionary 15, discrimination between uppercase letters and similar lowercase letters can be achieved.

FIG. 6(a) is a block diagram of the character recognition device of the second embodiment, and FIGS. 6(b) and ('C) are processing flowcharts of the second embodiment.

The character recognition device of the second embodiment includes an input section 61 and a CPU.
62, ROM63, RAM64, and recognition dictionary 15a, 15
b, 65c, - and an output section 67, and when document image data is input from the human power section 61, the CPU 62 outputs the data from the ROM.
According to the processing program stored in the RAM 64
Character segmentation and normalization are performed while using the image data storage and auxiliary storage. Based on the normalized result, that is, the normalized procedure, the recognition dictionaries 65 a, 65 b, 65 c, . Output from

The operation will be explained according to the processing flowcharts in FIGS. 6(b) and 6(c). Step 3600~Step 56
03 performs the same processing as 5200 to 5203 of the conventional example. In step 5604, the center of gravity of the figure of the cut-out character pattern is calculated. The center of gravity is A (0, O) based on point A, if the cut out character is a figure like the one shown in Figure 7, for example.

B (W+, O), C (0, h+), D (W+
.

h2), - as an example, f (x.

When y) is the density value of the pixel at the coordinates (x, y), (X++'jl) can be found as the coordinates of the center of gravity by calculating the following. In step 5605, the obtained X+
, '/r, the position of the center of gravity g(
i) Make the assignment. i is an integer from 1 to 9, and g(5) is assigned to the pattern in FIG. Patterns such as "," will be assigned g(3).

In step 5606, the external shape of the pattern is determined. The outline figure is the starting and ending positions of the pattern arranged in the X direction and the X direction. X at this time
Let the width in the direction be f (x) and the width in the X direction be f (y). In step 5607, the predetermined values A and f (x),
Compare f (y) and both t (x),
If f (y) is smaller, that is, if it is a consonant character or a special character, the process advances to step 5608. In the case of other characters, the process advances to step 5609.

In step 5609, g' = Min (Fx/f(x), Fy/f(y
)), the smaller value of the magnification ratio of the width in the X direction and the width in the X direction is selected. Now, for a character like "-", the scaling factor in the X direction is taken, and for a character like "1", it is
The magnification ratio in the direction will be selected. Note that Fx and Fy are Fx=64°Fy=64 if the recognition pattern is a (64°64) pattern.

On the other hand, in step 5608, the above calculated g(i)
The flow differs depending on the situation. The end of g(1), that is, “′
”, proceed to step 5610a;
When g(2), that is, a character like "*", the process proceeds to step 5610b, and when g(3), that is, a character like ","
”, the process of step 5610c is performed.

Although only the cases g(1) to g(3) are shown in FIG. 6(b), the respective processes are also performed in the cases g(4)...g(9). Also, g(1).

Processing in which g (4) and g (7) are represented by g (1) can also be considered. Steps 5610a, 610b.

610 c, ・- processing is F/f details, Mi
The magnification ratio is determined by n(Fx/f (x), Fy/f (y)), and further 1/4.1/3.1/2.・・・
By multiplying by , the normalized magnification factor is made different. This makes the patterns of the feature vector space after normalization between "" and "," different, thereby improving the recognition rate.

Step S61], step 5610a.

Normalization of the manual data is performed using the scaling factor of g′ obtained in steps 610b, 610c, . . . F (x) ,
A normalized pattern of F (y) will be found.

Feature extraction is performed on the normalized pattern in step 5612. In step 5613, the feature extracted vector f(a)...where a=1 to 64 is added to g' obtained in the previous step.

By adding g(i), the final vector space for performing the identification process is obtained.

In the second embodiment, two values, g' and g (i), are added to the vector space, but either one can have a large effect. In step 5614, the recognition dictionary 65a is determined based on the values of g' and g(i).

Selection of 65b, 65c, . . . is executed. g′ and g
If a recognition dictionary is provided for each of the combinations in (i), even the flow shown in Figure 6(b) will yield 4X
9=36 types of recognition dictionaries can be created. The number of recognition dictionaries may be provided according to cost and system necessity, and the division for determining the center of gravity position is not limited to nine.

Further, as in the first embodiment, processing may be performed without separating the X component and the X component. Thereafter, characters are identified based on the recognition dictionary selected in step 5615, and in step 561
6, candidate characters are output.

As described above, this embodiment provides an effective means for identifying similar characters "ya", "ya", etc. due to differences in size. For that purpose, the rectangular shape of the character was cut out.

By calculating the length in the X direction, the average value in multiple X directions and the average value in the X direction of the pattern can be calculated, and the individual lengths of the pattern data can be compared based on the average value. The purpose of the present invention is to have means for determining the magnification of the pattern data according to the comparison result. Furthermore, the scaling factor described above is set independently for each pattern data, and the scaling factor of the average value or the scaling factor corresponding to each pattern is selectively determined depending on the scaling factor for normalization of each pattern. It is characterized by being

Furthermore, according to this embodiment, punctuation marks and "eyes" are added to sentences written in free text.

It is possible to improve the character recognition rate itself by providing a means to normalize while maintaining the fidelity of reading pattern data for vertically long sentences such as ``日'' and ``1'' as well as horizontally long characters. Become.

Furthermore, in this embodiment, the position of the center of gravity of the image pattern of the extracted character is calculated, and the scaling means for normalization is changed depending on the position of the center of gravity and the size of the image pattern. This allows for similar characters that vary depending on the positional relationship of the pattern.
, ", ", etc. is provided.

In other words, in this embodiment, by focusing on the fact that the above-mentioned similar characters have different centers of gravity, by varying the scaling ratio from the character outer shape pattern to the recognized shape pattern depending on the size. The spatial properties of the feature vectors are changed.

Furthermore, this embodiment is characterized in that features are extracted from the recognized figure pattern normalized by the above means, and the scaling factor found by the above means is added as a feature parameter to the feature vector space found by the extraction. shall be. Further, the present invention is characterized in that a parameter of the address position of the center of gravity determined after character segmentation is added to the feature vector space normalized and determined in the above-mentioned means. Of course, the above-mentioned scaling parameter and barycenter address position parameter can be added as two independent vector spaces.

[Effects of the Invention] According to the present invention, it is possible to provide a character recognition device that identifies and recognizes similar characters with different sizes and similar characters with different pattern positions.

[Brief explanation of the drawing]

Fig. 1(a) is a block diagram of the character 1B recognition device of the first embodiment, Fig. 1(b) and (C) are the processing flowchart of the first embodiment, and Fig. 2(a) is the conventional A block diagram of the character recognition device, Figure 2(b) is a conventional processing flowchart, Figure 3 is a diagram showing an example of human-powered characters, Figure 4 is a diagram showing an example of extracted characters, and Figure 5 is the same as in Figure 4. A diagram showing a rectangular figure corresponding to a character, FIG. 6(a) is a block diagram of the character recognition device of the second embodiment, FIGS. 6(b) and (C) are a processing flowchart of the second embodiment, FIG. 7 is a diagram for explaining the calculation of the center of gravity in the second embodiment, and FIG. 8 is a diagram for explaining the assigned position of the center of gravity in the second embodiment. In the figure, 11...human resources department, 12...-CPU, 13...
ROM, 14...RAM, 15.16...Recognition dictionary, 17...Output section, 61...Input section, 62...
CPU. 83-ROM, 64-RAM, 65a, 6
5b. 65c... Recognition dictionary, 66... Output unit. Patent applicant: Canon Co., Ltd. Figure 1 (0) Figure 1 (C) Figure 2 (0) Figure 6 (0) Figure 6 (C)

Claims (4)

[Claims] (1) In a character recognition device that recognizes arranged characters, a separating means separates patterns arranged in a predetermined direction, and the patterns separated by the separating means are divided into a plurality of patterns according to a scaling ratio of each pattern. A scaling means for normalizing to a predetermined size by scaling at a scaling ratio determined based on the average scaling ratio, and a dictionary corresponding to the scaling ratio used by the scaling means. and recognition means for recognizing the pattern normalized by the scaling means. (2) The character recognition device according to claim 1, wherein the magnification changing means changes the magnification separately in the vertical direction and the horizontal direction. (3) The scaling means includes a comparison means for comparing the scaling ratio of each pattern with an average scaling ratio of a plurality of patterns, and a scaling ratio of each pattern is determined from a comparison result of the comparison means for a plurality of patterns. If the scaling ratio of each pattern is smaller than the average scaling ratio of multiple patterns, the scaling ratio of each individual pattern is normalized. 2. The character recognition device according to claim 1, further comprising: scaling ratio determining means for determining a scaling ratio of . (4) The character recognition device according to claim 1, wherein the recognition means uses the scaling ratio determined by the scaling means as part of the recognition parameter.