JP2000076378A - Character recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recognize a character with high precision even from a character image which is of small size or has illuminance spots or deformation and to precisely recognize characters from character images of many kinds of type without using many dictionaries corresponding to the types. SOLUTION: A recognizing processor 2 to which this method is applied generates outline data of image data inputted from an image input device 1 such as a digital still camera and matches the data obtained by normalizing the outline data (including enlargement, reduction, etc., of data) against dictionary data prepared by extracting outline data of standard characters registered in a storage device 4 to perform character recognition. At this time, the outline data of the image data are deformed into a type to be recognized with the dictionary data or deformed so as to have distortion corrected depending upon the type, distortion, etc., of the input image data. Consequently, even an image which is of small size or has illuminance spots or deformation can be recognized with high precision and precise character recognition of character images of many kinds of type can be performed only with e.g. a dictionary corresponding to one kind of type.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a digital camera,
The present invention relates to a character recognition method for recognizing characters included in an image input by a video camera, a scanner, or the like.

[0002]

2. Description of the Related Art Conventionally, there are many methods for character recognition. For example, there is a method of performing pattern matching between input data and image data of a dictionary to recognize characters of the input data, or performing character recognition of the input data using background characteristics and stroke characteristics of the characters. However, in any of the character recognition methods, a relatively high-quality and high-resolution image obtained by inputting a document with a scanner or the like is targeted for character recognition.

Heretofore, the method of inputting documents with a scanner has been the mainstream, but recently, with the spread of digital still cameras and video cameras, the use of these cameras as image input devices is rapidly increasing. As an application, character recognition has been performed on images input by these image input devices.

[0004]

Since the conventional character recognition method is intended for high-quality and high-resolution character images, this method can be directly applied to a low-resolution character image input by a digital camera or a video camera or the like. When applied to a character image having luminance unevenness or distortion, the character recognition rate is significantly reduced.

[0005] The causes include the inability to obtain a character size necessary for recognition due to low resolution, luminance unevenness and noise of an image unique to a camera, and distortion due to a shooting position. Further, in the conventional character recognition method, small characters cannot be recognized even when a document is input by a scanner in many cases.

Further, since the conventional character recognition application is assumed to be used on a personal computer or the like, if the character to be recognized is a plurality of typefaces, a dictionary corresponding to each typeface is created. This is stored in a large-capacity storage device such as a hard disk and used.

[0007] However, for example, when character recognition is performed by a device having no large-capacity storage device such as a portable terminal, the dictionary corresponding to each typeface described above cannot be held.
There is a problem that the typefaces that can be recognized are limited.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned conventional problems, and has as its object to improve the size of a character image input with a digital still camera or the like or a character image having a luminance unevenness or distortion. It is an object of the present invention to provide a character recognition method capable of accurately recognizing characters and accurately recognizing character images of various types of fonts without using a large number of dictionaries corresponding to various types of fonts.

[0009]

According to a first aspect of the present invention, there is provided a character recognition method for character recognition of character image data, comprising extracting a luminance contour line of the character image data to obtain a character image. Obtaining contour data of
Collating the obtained contour data with the dictionary data to obtain a character recognition result.

According to the first invention, for example, the contour lines of the character image are obtained by extracting the luminance contours of the character image data input from a digital still camera or the like. Usually, the contour line data is enlarged or Perform normalization for reduction to obtain data of an appropriate size. After that, as the dictionary data, for example, the contour data obtained by extracting the luminance contours of the standard character data and the normalized data are collated. At this time, even if the character image data has a low resolution, even if there is a luminance unevenness, etc., the obtained contour data is clear, so that the collation with the dictionary data can be severely performed,
A highly accurate character recognition result can be obtained. Even when the size of the input character image is small, the normalization converts the outline data to an appropriate size and performs the collation, so that a highly accurate character recognition result can be obtained.

According to a second aspect of the present invention, in the character recognition method according to the first aspect, when the contour data is compared with dictionary data, the contour data is deformed.

According to the second aspect, for example, when an input character image is distorted due to a photographing position of a digital still camera or the like, a deformation such as coordinate transformation for correcting the distortion is performed. By applying the data to the contour data of the character image data, contour data in which distortion has been corrected is obtained. By comparing the contour data in which the distortion has been corrected with the dictionary data, a highly accurate character recognition result can be obtained even for a distorted character image. Also,
When the typeface of the input character image is different from the standard typeface of the dictionary data, the contour data of the input character image is subjected to deformation by, for example, coordinate conversion, and the typeface is adjusted to the standard typeface and then collated with the dictionary data. As a result, a highly accurate character recognition result can be obtained. Therefore, it is not necessary to have many dictionary data corresponding to many types of typefaces.

According to a third aspect of the present invention, in the character recognition method according to the second aspect, the degree of deformation of the outline data is changed according to positions of a plurality of character images corresponding to the outline data. It is in.

According to the third aspect, when the deformation is applied to the contour data having distortion, when the distortion of a plurality of character images corresponding to the contour data differs depending on the position, the deformation is performed according to the position of the character image. By changing the degree, the distortion of all character images can be properly corrected, and in this case, a highly accurate character recognition result can be obtained.

According to a fourth aspect of the present invention, there is provided the above second or third aspect.
In the character recognition method described in (1), the transformation is to perform coordinate transformation on the outline data.

According to a fifth aspect of the present invention, there is provided the first to fourth aspects of the present invention.
In any one of the character recognition methods described above, the collation includes:
After normalizing the contour data, a distance between corresponding points of the normalized contour data and the dictionary data is obtained, and a character recognition result is obtained by summing the distances.

According to the fifth aspect, the distance between the corresponding points of the normalized contour data and the dictionary data is obtained, and the character of the dictionary data having the smallest sum of the distances is determined as the recognized character. Become the first candidate.

According to a sixth aspect of the present invention, there is provided the first to fourth aspects.
In any one of the character recognition methods described above, the collation includes:
After normalizing the contour data, the normalized contour data and the dictionary data are respectively developed into bitmap data, and a character recognition result is obtained based on the degree of coincidence between the two bitmap data.

According to the sixth aspect, a character to be recognized is selected based on the degree of coincidence between the bitmap data of the contour data of the character image and the bitmap data of the dictionary data. Even if there is a contact or interruption, if this is a part, it does not significantly affect the matching degree, so that a high character recognition rate can be obtained.

A seventh feature of the present invention is the above-mentioned claims 1 to 6.
In any one of the character recognition methods described above, the dictionary data is contour data created by extracting luminance contours from standard characters.

According to an eighth aspect of the present invention, there is provided the above-described first to sixth aspects.
In any of the character recognition methods described above, the dictionary data is an outline font of standard characters.

According to the eighth aspect, dictionary data can be easily obtained by using an existing outline font.

A ninth aspect of the present invention is the character recognition method according to the first aspect, wherein the character recognition result is further collated with a word dictionary to obtain a final character recognition result.

According to the ninth aspect, for example, when a character image to be recognized is an English word or the like, one word cannot be recognized for the entire character, and when the recognition is performed in units of one character. If the correct answer is, for example, up to the third candidate, even if it is not the first candidate, the correct word can be obtained by collating these character strings (words) with the word dictionary. Further, the character recognition rate can be improved.

[0025]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a character recognition device to which the character recognition method of the present invention is applied. An image input device 1 such as a scanner, a digital still camera, a video camera and a combination of capture devices inputs image data of a subject to a recognition processing device 2. The recognition processing device 2 is also connected to an input device 3 such as a mouse or a keyboard for inputting an instruction or the like from a user.
Character recognition is performed using the storage device 4 that stores dictionary data and programs, and the results are displayed on the image display device 5 or output to an output device 6 such as a printer.

The above-described character recognition device can be constituted by, for example, a personal computer and its peripheral devices.

Next, the operation of this embodiment will be described. First, the dictionary data stored in the storage device 4 is created in accordance with the flowchart shown in FIG. That is, in step 201, contour lines are created by extracting brightness contour lines from the image data of the standard characters, and the contour data is registered in the storage device 4 in step 202 as dictionary data. Note that outline font data already created as character font data may be used as dictionary data.

Next, an operation in which the recognition processing device 2 shown in FIG. 1 performs character recognition using the dictionary data created as described above will be described with reference to a flowchart shown in FIG. First, in step 301, character image data input by the image input device 1 such as a digital still camera or a scanner is stored in the storage device 4. Next, in step 302, the recognition processing device 2 extracts contour lines of brightness from the character image data to obtain contour data of the character (sometimes referred to as character data). In this case, since one character may be composed of a plurality of outlines, in such a case, integration is performed based on the positional relationship of the outline data to obtain data for one character.

Here, a method for extracting the above-mentioned luminance contour lines will be described. The extraction of the luminance contour is performed by a method that extends the contour extraction of the binary image, and is extracted as a luminance distribution as shown in FIG. Such a luminance distribution is calculated by a certain threshold value of
The contour in the case of the value is obtained by tracing the boundary between the white pixel and the black pixel as shown in FIG. As shown by the thick solid line P shown in FIG. 5, the contour always forms a closed loop. Since the given threshold value is between the pixel values on both sides of the boundary line, the luminance contour should just move the position of this boundary line according to the pixel values on both sides.

In the luminance distribution as shown in FIG.
If 0, a boundary line P is obtained as shown in FIG. 5 described above. If this boundary line P is moved by performing a proportional calculation using the pixel values on both sides, a smoother line as shown by Q in FIG. 5 is obtained. A high brightness contour line is obtained. FIG. 6 shows A shown in FIG.
FIG. 8 is a diagram showing how to obtain a contour line by the proportional calculation of the section.

However, when the boundary line Q bends in the traveling direction and two points after the movement are close to each other, the middle point between the two points is taken as the contour point. Further, an appropriate value may be designated as the threshold value used in the above processing, or a method such as a discriminant analysis method may be used to automatically obtain the threshold value from the density histogram of the image. Since the contour line extracted in this way has irregularities and adversely affects the collation, smoothing is performed next. In the smoothing, the coordinate value of each point is replaced with the average value of the coordinate values of several points before and after.

Here, an example of the extracted character data is shown in FIG. 7 (A) shows the result of extracting the contours of the input character image data using the luminance contours, and FIG. 7 (C) shows the result of extracting the contour lines shown in FIG. 7 (B). Character data (contour line data) obtained by the conversion process.

The recognition processing device 2 sequentially compares the character data obtained as described above with dictionary data in the storage device 4. Before that, in step 303, the input character data is normalized. This normalization is, as shown in FIG. 8B, a process of performing vertical and horizontal enlargement / reduction according to the size of the circumscribed rectangle of each character to make it the same size as the dictionary data shown in FIG. 8A. .

Thereafter, it is determined whether or not the character data normalized in step 304 needs to be further transformed by using coordinate transformation or the like. The necessity of this deformation is determined in step 304 based on the typeface of input character image data described later and the presence or absence of distortion. If it is not necessary, the process proceeds directly to step 306, where the character image data is collated with the dictionary data. In step 307, the recognition processing device 4 outputs the character recognition result obtained by the collation to the image display device 5.

The above collation methods can be roughly divided into two types. One is a method of calculating the distance between the coordinate values of the contour data as character data, and the other is a method of developing the contour data into a bitmap and calculating the degree of coincidence between the bitmaps.

First, a method of calculating the distance between coordinate values will be described. In this method, the contour data is made to correspond according to the arrangement order, and the distance between corresponding points is obtained.
First, it is necessary to match the start point of the contour data. For this purpose, a zigzag scan as shown in FIG.
The lower left point in the character can always be the starting point. still,
FIG. 9B shows an example of a normal scan.

For example, when the character m shown in FIG. 9A is normally scanned in the horizontal direction from the lower left, the starting point is A
It is not possible to specify which of the BCs, but if zigzag scanning is performed, it will always be A, eliminating the need for starting point alignment.

When determining the correspondence between the contour data as shown in FIG. 10B and the dictionary data as shown in FIG. 10A, the correspondence is not changed by the width or height of the character. First, feature points in the contour line, in this case, a corner point are set as corresponding points (points indicated by the same number), and corresponding points of other points are obtained. The turning point is determined by judging a point where the coordinate value becomes a maximum or a point at which the direction of the outline changes greatly from the coordinate value.

The recognition processing device 2 obtains the distance between the corresponding points determined in this way as shown in FIG. 10 (C), and the sum is used as the distance between the dictionary data and the character data.
In step 307, the recognition result is output starting from the one with the shortest distance. When one character is composed of a plurality of outlines, the distance between the outline data is obtained for each of the outlines, and the distance is summed up.

Next, a method of performing bit map development from contour data will be described. In the above-described method, if the number of contour lines or the order of contour points is different, a correct point may not be obtained because a corresponding point with dictionary data cannot be obtained. For example, as shown in FIG.
In the case of “planarity”, a contour extraction result as shown in FIG. 11B is obtained, but a part of “a” may come into contact. In such a case, the dictionary data has two contour lines as shown in FIG. 11D, whereas the character data has three contour lines as shown in FIG. 11C. Since the route is also different, a recognition result of a cannot be obtained. In particular, in the case where there is luminance unevenness in the image, lines tend to touch or conversely break, and this method reduces the character recognition rate.

For example, in Japanese Patent Publication No. Hei 6-162264, a character pattern is deformed and matched in order to cope with a break. However, when there is a local luminance unevenness or noise such as an input from a digital camera. Does not provide sufficient recognition results.

Therefore, a method of developing a bitmap of a predetermined size from the contour data and calculating the degree of coincidence between the bitmaps is adopted. According to this method, the shape of the character itself can be collated without being largely affected by contact or interruption of the line. Since the input image data is converted into vector data by contour line extraction, a bitmap image of an arbitrary size can be created by coordinate value conversion. For this reason, a high-quality bitmap image can be created as compared with a general method of directly enlarging or reducing an input image, so that the character recognition rate can be increased.

FIG. 12A shows an example of the bitmap development described above. This is an example developed to a size of 32 × 32. As shown in FIG. 12B, the dictionary data is similarly subjected to bitmap development, the number of bitmaps that do not match white / black is defined as the distance, and the smaller distance is output as a recognition result in step 307.

The size of the bitmap image to be expanded is arbitrary, and an appropriate value can be designated according to the capacity of the memory or the storage device 4, the performance of the CPU, and the like.

The dictionary data is not developed in a bitmap manner for each input character, but is developed in a bitmap manner when the outline of the dictionary data is read.
Calculation time can be reduced.

However, according to this method, the letters l and i
For example, vertically long characters such as... Become black in the entire bitmap due to normalization, and characters with much black are erroneously recognized. In order to avoid this, a method is adopted in which vertically long characters and other characters are grouped in advance, and character recognition is performed by selecting a group according to the aspect ratio of the input characters.

In addition, it is also possible to set a group for the position of the character. For example, the letters q, j. p
And the like are located below the reference line. Therefore, by recognizing these from the coordinate values and narrowing down the groups, the recognition rate can be further increased.

Next, a description will be given of a method of dealing with various types of character images. Here, English characters will be described as an example. There are various differences depending on the typeface, and for example, differences in line thickness, vertical and horizontal line widths, decorative portions, inclinations, and the like are given depending on the typeface. Generally, as in Japanese Patent Application Laid-Open No. 5-159107, dictionaries for fonts of characters to be recognized are prepared to handle various types of fonts.
Due to U-performance constraints, large amounts of data may not be available. Therefore, there is a need for a method of recognizing various types of fonts by only having a dictionary of one standard type and at most two to three types of typefaces.

Therefore, in this example, the outline data obtained from the input character image data is subjected to coordinate transformation in step 305 and deformed, so that one type of characters of various types is obtained, Recognition is possible with dictionaries for three types of fonts.

For example, it is assumed that contour data of a standard typeface as shown in FIG. 13A is registered in the storage device 4 as dictionary data. This example is a typeface that is simple, has no decoration, and has the same vertical and horizontal line widths that represent the shape characteristics of English characters. Hereinafter, the operation when a typeface as shown in FIG. 13B is given as an input character image from the image input apparatus 1 will be described. First, since the input character image is in italic, the recognition processing device 2 generates, as shown in FIG.
A coordinate transformation for erecting this is performed, and further a transformation for narrowing the vertical and horizontal line widths in accordance with the dictionary data is performed as shown in FIG.

Here, the erect of the input character image is shown in FIG.
The coordinate conversion and the line width correction as shown in FIG.
This is realized by the movement of the coordinate points as shown in FIG.
If the vertical and horizontal line widths are different, by changing the vertical and horizontal movement ratios, it is possible to perform conversion such as thickening only the horizontal lines. Although it is necessary to delete the decoration part of the input image, it is necessary to perform a lot of processing to delete it based on the shape.

Therefore, attention is paid to the end portion of the character image which adversely affects the recognition. If normalization is performed as it is when there is a decoration at the end of the character image, the position of the vertical line will be shifted, and many parts will not match the dictionary. Therefore, FIG.
As shown in FIG. 4 (C), a special coordinate transformation for increasing the movement amount at the end of the input image is performed, and as shown in FIG. Perform a conversion to eliminate. In all of the above processes, since coordinate conversion is performed on the coordinate data of the outline, there is an advantage that a shorter processing time is required than performing the same process on the bit map.

The recognition processing device 2 determines the type of coordinate transformation to be performed by obtaining shape characteristics such as the maximum length angle of the input image, the projection distribution of vertical and horizontal luminance, and the luminance distribution of the periphery of the character. It is determined by judging which of the assumed typefaces corresponds.

Next, the deformation (step 305) for correcting the distortion when the input image has distortion will be described. Conventional character recognition is often designed based on the assumption that a paper is read with a scanner or an image at a fixed position is input with a fixed camera, and inherently has inherent image distortion like a digital camera. The case where the input image is distorted due to the input image being distorted or photographed from an oblique position with respect to the photographing target has not been considered. Therefore, in such a case,
Although it is necessary to correct the distortion of the input image, this is also performed by the coordinate conversion as in the case of the typeface described above.

For example, when a character image as shown in FIG. 15A is input from the image input device 1 to the recognition processing device 2, the contour extraction result of the input character image is obtained as shown in FIG.
As shown in FIG. 3B, since the dial is photographed from above, the lower part is distorted as if it were shrunken, and the distortion is different between right and left. In the case where the distortion is small, a correct answer can often be obtained even when the dictionary data is collated as it is, but when the distortion is large, it is necessary to correct the distortion. Here, by performing coordinate conversion while changing the correction width of the inclination depending on the position in the character image, it is possible to correct the shape to a shape close to a standard character as shown in FIG.

In character recognition, in general, there is known a method of correcting a character recognition result based on information of a word or an idiom and obtaining a correct answer in addition to a recognition result of a single character. For example, in the recognition of English words, even if the recognition result by itself is wrong, it is possible to obtain a correct word from surrounding correct characters by using a word dictionary. for that purpose,
Instead of outputting only one recognition result, it is necessary to output a plurality of recognition results and their certainty.

Also in this example, for example, 3
Two characters may be output and compared with a word dictionary to obtain a correct answer. By this combination, for example, if a correct character is included in the third place, a correct word can be obtained even if another character is erroneously recognized. At the time of word search, candidate words can be taken out of the dictionary, and evaluation results can be obtained by using the above-described certainty factors to obtain recognition results.

Finally, FIG. 16 shows an overall processing flow in which the above-described basic character recognition processing of the present example and deformation processing relating to various types of fonts and distortion of input character data are summarized.

According to this embodiment, the contour data obtained by extracting the contour of character image data such as image data input from a digital still camera or the like is normalized (the size of the input character image is enlarged or reduced). By comparing this with dictionary data obtained by extracting the outline of the standard character, erroneous low-resolution and small-size character images with luminance unevenness and the like can be mistaken. Highly accurate character recognition can be performed with less recognition. Further, when the input character image is distorted, the character is recognized after being deformed to correct the distortion, so that highly accurate character recognition can be performed even on distorted character image data. .

Further, in recognizing character images of various types of fonts, the above-described character recognition is performed from a character image of a typeface which can be recognized by transforming the character image. By simply preparing a dictionary, character images of various types of fonts can be recognized with high accuracy. Therefore, even in a device such as a portable terminal in which a large-capacity storage device cannot be mounted, various types of character images can be recognized with high accuracy.

[0061]

As described above in detail, according to the character recognition method of the present invention, even if the size of a character input by a digital still camera or the like is small, or a character image having uneven brightness or distortion, character recognition can be performed with high accuracy. In addition to this, it is possible to accurately recognize character images of various types of fonts without using many dictionaries corresponding to various types of fonts.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a character recognition device to which a character recognition method of the present invention is applied.

FIG. 2 is a flowchart showing a process of creating dictionary data registered in the storage device shown in FIG.

FIG. 3 is a flowchart illustrating a character recognition processing procedure performed by the recognition processing device illustrated in FIG. 1;

FIG. 4 is a diagram showing an example of a luminance distribution by extracting luminance contour lines of input image data.

FIG. 5 is a diagram showing a method of creating a contour from the luminance distribution shown in FIG.

FIG. 6 is a diagram for explaining details of the contour line creation method shown in FIG. 5;

FIG. 7 is a diagram illustrating an example of an input image, its contour line extraction result, and the result of smoothing.

FIG. 8 is a diagram illustrating a method of normalizing input data.

FIG. 9 is a diagram illustrating how to scan input data.

FIG. 10 is a diagram illustrating a method of obtaining a distance between corresponding points between input image data (contour line data) and dictionary data.

FIG. 11 is a diagram illustrating input data suitable for performing bitmap development, a contour line extraction result thereof, and problems.

FIG. 12 is a diagram showing an example of bitmap expansion of input image data (contour line data) and dictionary data.

FIG. 13 is a diagram illustrating a method of transforming input image data (contour line data) according to a typeface.

FIG. 14 is a diagram for explaining a method of deformation used in FIG.

FIG. 15 shows distorted input image data (contour line data).
It is a figure explaining the method of transforming.

FIG. 16 is a flowchart showing an integrated flow of character recognition and various processes related to the character recognition by the apparatus shown in FIG. 1;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input device 2 Recognition processing device 3 Input device 4 Storage device 5 Image display device 6 Output device

Claims (9)

[Claims] 1. A character recognition method for character recognition of character image data, comprising the steps of: extracting luminance contours of the character image data to obtain contour data of the character image; And obtaining a character recognition result by collating the characters. 2. The character recognition method according to claim 1, wherein the contour data is deformed when the contour data is compared with dictionary data. 3. The character recognition method according to claim 2, wherein the degree of deformation of the outline data is changed according to positions of a plurality of character images corresponding to the outline data. Recognition method. 4. The character recognition method according to claim 2, wherein the transformation is performed by performing a coordinate transformation on the outline data. 5. The character recognition method according to claim 1, wherein the collation is performed by normalizing the contour data, and then determining a corresponding point between the normalized contour data and the dictionary data. A character recognition method comprising: obtaining a distance between each other; and obtaining a character recognition result based on a sum of the distances. 6. The character recognition method according to claim 1, wherein in the collation, the contour data is normalized, and then the normalized contour data and the dictionary data are bit-mapped. A character recognition method which develops data and obtains a character recognition result based on the degree of coincidence between both bitmap data. 7. The character recognition method according to claim 1, wherein the dictionary data is contour data created by extracting luminance contours from standard characters. 8. The character recognition method according to claim 1, wherein the dictionary data is an outline font of a standard character. 9. The character recognition method according to claim 1, wherein the character recognition result is further collated with a word dictionary to obtain a final character recognition result. .