JP2008077621A - Method for making up for learning character sample of character recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method for automatically forming a character group having no unreadable characters for each character quality which is highly deformed in various manners as compared with an original character sample group when the learning character sample group is not enough for learning in character recognition by statistic pattern recognition. <P>SOLUTION: Based on a main component analysis result of characteristic vectors including coordinates of characteristic points of the centerlines of collected handwritten characters or the contour lines of print characters of other kinds of fonts, a range that a person can recognize a character of a type or a range permitted as the character type is determined by visual observation or according to a mathematical condition to determine a readable range or allowable range represented by a sectional super-ellipse, and sufficient characters having various deformation degrees and qualities are automatically generated using a psychological distance showing the distance from the center of the sectional super-ellipse, and added to the learning character sample group collected first. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

Detailed Description of the Invention

  The present invention relates to enhancement of a learning character sample group of a character recognition system.

  In the field of handwritten OCR, a recognition rate of over 99.5% for unlearned characters was obtained with the ETL9 (B), a regular handwritten hiragana / kanji database created by the Institute of Electronics Technology Research Institute of the Ministry of International Trade and Industry and Fujitsu Limited. However, in spite of many efforts to improve the feature amount and the identification unit, in recent years, the recognition rate has been increasing.

  ETL9 (B) was written on a prescribed form with a focus on writing, but in our daily work, each person is given more attention to the work. The quality is low. In order for handwritten OCR to be truly practical, it is necessary to be able to read characters written on a daily basis with practical accuracy, and at the current technical level, there are no conditions for popularization of handwritten OCR.

  Although significant progress has been made in both handwritten zip codes and reading classification rates, it is necessary to reduce the correct reading classification rate by rejecting the first three digits in order to avoid misclassification. If accuracy is realized, it will be possible to further reduce labor costs. The reading accuracy of handwritten postal addresses and numbers presented by the United States Postal Service to discover the best recognition technology from all over the world is a recognition rate of 50% and a misreading rate of 2% or less. Still far from human reading accuracy.

  Therefore, considerable technical progress is still necessary for handwritten OCR to occupy a position as truly practical.

  On the other hand, the recognition of printed characters is generally considered easier than the recognition of handwritten characters, but the printed characters can also be recognized with high accuracy for various typefaces and innumerable character designs. It is not easy to do. In Europe and the United States, it is important to recognize not only the typeface, but also the design, and this function is realized as an important function of English OCR, but this function can be used in documents in any country. Then, the recognition rate of the whole document printed with the same character shape can be greatly improved.

  In addition, mathematical characters appearing in many publications in the science and engineering field represent differences in the mathematical meaning of differences in typefaces such as Roman, Italic, Gothic, and bold, and the difference between uppercase and lowercase letters in the alphabet. Accurate identification between them is indispensable, and if you want to read mathematical formulas such as books and papers from ancient and modern east and west, there is an infinite variety of glyph designs for each typeface, and capital letters Are similar to each other (C and c, S and s, etc.), and compared to general sentences, character recognition errors in mathematical formulas are difficult to correct by context. Therefore, a technique for obtaining a sufficiently high recognition rate practically in terms of mathematical formulas has not been realized.

  Well, in both handwritten characters and printed characters, the recognition character group of the recognition system is to collect a large number of handwritten characters that have been written so far in handwritten OCR, and has been used so far in OCR for printed characters. It is ideal to collect as many different font types as possible, but it is physically and economically difficult to implement. As shown in FIG. 1, the learning character sample group that is actually collected usually has an insufficient distribution and is unevenly distributed.

  In statistics, the existence of a group of data that can give a correct probability distribution is a precondition, but in a high-dimensional feature space that requires high recognition accuracy, this condition must be satisfied in a group of samples gathered by normal efforts. Experts have pointed out that it is impossible, and that increasing the number of samples (within the range of human abilities) is extremely ineffective.

Moreover, even if the problem of learning character group collection is solved, there are too many unclear points about correspondence to characters to be written and fonts to be designed in the future.
There appear to be two main approaches to solving this problem.

  One of them is to try to make up for the lack of information in the learning character group by incorporating general knowledge that people seem to use for character recognition into the system with the ingenuity and idea of the designer (various distance / discriminant functions). And the introduction of flexible matching methods such as relaxation methods), and many attempts have been made for many years, but the effect seems to be saturated in recent years.

Another approach is to apply image processing deformation to the character pattern or to generate a large number of various deformed characters using a dynamic model of human writing movement.
In recent years, [Non-Patent Document 2] has published a method for compensating for the lack of learned characters by adding deformation by affine transformation to each stroke during online input and a result of improving recognition accuracy.
However, in this type of method, as the deformation is increased, characters that are not recognized as characters of that character type occur, and unless there is a means for preventing this, the recognition rate is saturated.
H. Miyano, M .; Maruyama, Y .; Nakano, and T.K. Hananoi, “Off-Line Handwritten Character Recognition by SVM based on Virtual Examples Synthesized from On-line Characters” Proc. of Eighth International Conference on Document Analysis and Recognition (Soul), pp. 494-498, Aug. 29-Sept. 1,2005.

  On the other hand, humans do not see most handwritten characters in the world, but they show sufficient recognition accuracy. This is probably due to the excellent abstraction and generalization ability of human beings, but as long as statistical recognition methods are used, abstraction that can achieve sufficient recognition accuracy with realistically collected character samples Development of alternatives to capacity / generalization ability is essential.

  In [Non-Patent Document 2], the applicant has important points such as end points, bending points, and branch points of the thinning pattern as feature points, and subfeatures that divide strokes existing between these feature points at equal intervals. A model with points (also referred to as auxiliary points) is set for each character type (order numbers are set for feature points and sub-feature points), and on the learning characters thinned for each character type, Points corresponding to the feature points / sub-feature points are obtained, and their feature coordinates are obtained by arranging their x and y coordinates in numerical order.

  The principal component analysis is performed for each character type in the feature space determined in this way, and the character core line pattern corresponding to the point on the principal component axis (eigenvector) is given a thickness so that the subject can read the character up to which point. Measurements were made to answer this question. By this method, the range that the subject recognizes as a character belonging to the character type in the feature space was captured as a piecewise hyperelliptic surface. The point on the piecewise hyperelliptic surface is assumed to be at a distance of 1 from the origin, and the point in the piecewise hyperelliptic surface is defined to decrease proportionally, and this is named psychological distance.

  In the recognition process, unknown character patterns (characters to be recognized) are thinned, and matching with each character type is performed after the points corresponding to the feature points defined in the standard pattern of the character type are obtained. A vector is required. From this vector, the psychological distance to the standard pattern of each character type is calculated, and the one that gives the minimum psychological distance is the recognition result.

With this method, a method to compensate for some deficiencies in the contents of the learning character sample group has been obtained, but as is well known, the conventional thinning technique distorts the character core wire near branch points, intersections, and inflection points. As a result, it was found that the true character core line could not be handled, which made it more difficult to access human recognition results.
For this problem, it is necessary to develop a thinning algorithm with almost no core distortion, but a sufficient algorithm has not been realized yet.

  In character recognition using statistical pattern recognition, when sufficient learning is not possible with just the collected learning character sample group, the deformation intensity and mode of deformation are far more diverse and wide than those of the learning character sample group, and can be read by humans. Automatically create character groups that do not contain missing characters according to character quality.

Means to solve the problem

  The present invention relates to a method mainly using generation of a character sample pattern for learning handwritten characters (method using a character core line) and a method focusing on generation of a character sample pattern for learning printed characters (a method using a character outline). ).

(1) Method using character core line For handwritten characters, the psychological distance known from [Non-Patent Document 1] is considered as the strength or quality of deformation of the character, and the interval [0, 1) of the psychological distance is defined. Divide into equal intervals. For example, if it is divided into 10 sections, a large number of characters in sections 0 to 0.1, 0.1 to 0.2,... Is automatically created.

The original learning characters can be entered on a pen tablet, or collected off-line, displayed on the display with characters such as a handwritten character database, and traced with a pen. When inputting an enormous number of character samples such as handwritten kanji, it is practical to use a thinning process with little distortion, such as the thinning algorithm currently under patent application, and calibrate on the display. I will.
FIG. 2 conceptually illustrates a method for extracting feature points defined on a standard pattern from characters input with a pen tablet.

Subsequently, as shown in FIG. 3, sub-feature points that equally divide the character lines connecting the feature points are extracted.
According to [Non-Patent Document 1], when a standard pattern of a character type (defined by one or more curves) has N feature points / sub-feature points, the original feature vector F of a character belonging to the character type is ,
F = (x ₁ , y ₁ , x ₂ , y ₂ ,..., X _N , y _N ) (1)
It is represented by Here, x _i and y _i are the x-coordinate and y-coordinate of the feature point i, and N is the total number of feature points and sub-feature points combined. Further, in order to facilitate handling, the original feature vector F is normalized by Expression (2), and is represented by x.
x = (x ₁ , x ₂ ,..., x _n ) = F / | F | (n = 2N) (2)

ここで、μは当該字種の平均、φ_ｉ，Ｕ_ｉはそれぞれ第ｉ主成分軸（固有ベクトル）と人間の可読限界であり、これらは字種ごとに異なる。For each character type, a feature vector x for a standard character shape is determined in advance, and corresponding feature points are extracted by the DP method or the like on the input learning character (line figure), and their x and y coordinates are determined. F and x are obtained in the order of feature point numbers.
Subsequently, a distance (difference) from each character type is obtained by a psychological distance formula of each character type having the form of Expression (3).

Here, μ is the average of the character type, and φ _i and U _i are the i-th principal component axis (eigenvector) and the human readable limit, respectively, which are different for each character type.

ただし、Ｕ_ｉ ^＋，Ｕ_ｉ ⁻は、それぞれ第ｉ主成分軸の正側及び負側の可読限界値を表す。Further, (x−μ, φ _i ) represents an inner product of vectors, and U _i is given by Expression (4).

However, U _i ⁺ and U _i ⁻ represent readable limit values on the positive side and the negative side of the i-th principal component axis, respectively.

  The above is the contents described in [Non-Patent Document 1], but in the present invention, character type identification is performed by applying Equations (3) and (4) to feature vectors extracted from thinned characters. Rather than selecting a large number of points in a piecewise hyperellipsoid (0 ≦ d (x) ≦ 1) measured as a human-readable range on the feature space, and covering the human-readable range A pattern group is obtained and added to the original learning character sample group.

  For this reason, an n-dimensional vector (vector direction is randomly selected) having a uniform random number as a component is created as x−μ, and the value of d (x) obtained by the equations (3) and (4) is represented by D Then, (x−μ) / D gives a psychological distance of 1. When a parameter k taking a value in the interval [0, 1] is introduced, k (x−μ) / D gives a psychological distance k.

  If k (x−μ) / D + μ is created by adding the average vector μ of the character type to this, this becomes a feature vector having a psychological distance k in the original feature space coordinates, and the feature points / sub-feature points determined by the components If a line segment is connected by a line segment according to the definition regarding the feature points of the standard pattern, a character core line having a psychological distance k is restored.

For example, as described above, if the interval [0, 1] is divided into 10 intervals of 0.1, and k is randomly generated in each interval, a 10-level character group according to quality is obtained.
By adding this to the original learning character group, a character group having a far more diverse and unbiased distribution in the feature space than the original learning character group that can be normally collected is realized.

In addition, a distance value (representing the thickness of the portion of the character line) z _i (i: feature point number) at the feature point is added to the original feature vector of Expression (1), and the original feature vector is
F = (x ₁ , y ₁ , z ₁ , x ₂ , y ₂ , z ₃ ,..., X _N , y _N , z _N ) (5)
This method is also included in the scope of this patent application. This method can detect local variations in the thickness of the character line, and is effective for handling brushstroke characters.

(2) Method using a character outline In the method using a character core described in (1), even if information on the thickness of a character is added, a character with a complicated decoration, such as a printed character, is added. The application of is often inefficient.

  As shown in Fig. 4, select existing standard fonts (generally more than one may be required, but if you make an appropriate selection, you probably only need one), and each character type Set feature points (refractive points, curvature discontinuities, etc.) in the pattern outline pixel array.

  Input the collected characters of various fonts, and find the points corresponding to the feature points of the standard pattern in the outline pixel row, as in the method (1) using the character core line, and further, the characters between the feature points The line is equally divided into the same number of line segments as the standard pattern to obtain points corresponding to the sub-feature points, and an original feature vector F having these coordinate values as elements is created.

  Subsequently, principal component analysis is performed for each character type using the normalized feature vector x, points are provided on each principal component axis at regular intervals with the standard deviation as a unit, and the points are converted into character data.

  For the character created by the above means, the variation allowable width is determined by obtaining a limit area that satisfies the condition that the contour line does not intersect with another contour line (see FIG. 5). In this way, feature vectors are randomly generated within the piecewise hyperellipse. The feature vector is converted into character outline data, and a process of filling a region surrounded by the outline is performed to generate character image data having shape information.

FIG. 6 shows an example of a pseudo font character generated by randomly selecting a point within the allowable area of the character type “F” in the feature space.
The method for determining the region limit described above is different from the method using the character core wire of (1), but depending on the purpose of use, it may be more reasonable to use the visual method similar to (1). possible.
FIG. 7 exemplifies a pseudo font character corresponding to a point in a section that falls within the allowable region of the character type “F” on a certain principal component axis.

The invention's effect

  In character recognition, a high-dimensional feature space is used to achieve practical recognition accuracy. Therefore, when applying the statistical pattern recognition method, the character sample population that can be normally collected is required by statistical theory. It is difficult to say that the conditions for satisfying this condition are satisfied, and it can be said that this type of recognition system has been a barrier to improving the correct reading rate.

  However, in the present invention, the character sample group that can be normally collected is not used as it is for learning (design) of the recognition system as it is, but the readable area of each character type obtained by the statistical analysis and the psychological measurement based on it or the permissible area. A sufficient number of characters of various qualities are automatically generated in the entire readable range or allowable range using a region (denoted as a piecewise hyperelliptic surface in the feature space) and a psychological distance formula, and added to the learning character group. Therefore, the amount of information necessary for character recognition (particularly human character type discrimination criteria) is much larger than the character sample group collected first, and if this is used for the learning character sample group, it is close to humans. A recognition result or a high recognition rate can be obtained.

The learning character sample group obtained here can be expressed as a binary image, so it can be used for learning of any recognition method, and the learning problem that can be said to be the biggest challenge in the construction of a character recognition system for a new character group. It is expected to make a significant contribution to the solution.
It can also contribute to clarifying problems when applying statistical discriminant analysis methods that have been widely used in the past.

Even when the method of this application is used, the better results can be expected as the sample of characters is collected as much as possible and the distribution is not largely biased.
(1) Method using character core wire In the case of the method using character core wire, the collected characters must be represented by the core wire. Therefore, there are many advantages to the online input method using a pen tablet.

  In particular, in the number recognition, there is little problem of fluctuation of the stroke order, and it is easy to extract points corresponding to the feature points defined in the standard pattern by using the writing order (the appearance order of the feature points). For detection of the inflection point, an almost natural position is detected by using a well-known DP matching method. However, if strictness is desired, the position of the inflection point can be detected by a man-machine interface using a pen tablet system with a display. Adjustments can be made.

  However, when sufficient character samples cannot be obtained by online input alone, the character pattern of a database collected offline (for example, the handwritten numeric database IPTP CD-ROM1 created by the Postal Service Institute of the Ministry of Posts and Telecommunications) is displayed. If the pen is moved, or if it is a pen tablet with a display, the input character data can be increased by tracing the core portion of the display character with the pen.

  If a thinning algorithm with little distortion (patent pending) is used, data can be collected by thinning the character images collected off-line and confirming / calibrating them with a pen tablet with a display.

This method may be the most effective method especially when there are many types of characters, such as kanji recognition. In this case, since information such as the stroke order cannot be used for feature point extraction, as shown in [Non-Patent Document 3], a line segment (two-dimensional) connecting feature points defined to represent the relative positional relationship between feature points is used. Vector) matching is used.
Fumitaka Kimura, Mitsuru Yoshimura, Koji Miyake, Masato Ichikawa, “Free Handwritten Katakana Character Recognition by Stroke Structure Analysis,” IEICE Transactions, Vol. 62-D, no. 1, pp. 16-23, Jan. 1979.

When the feature vector F of Equation (1) is obtained by the above method, the feature vector x normalized by Equation (2) is obtained. This corresponds to one point in the n-dimensional feature space, and a principal component analysis is performed on a group of points for each character type.
The following follows the method described in [Non-Patent Document 1].
That is, in the principal component analysis, the principal component axis is determined as an eigenvector and the variance related to the axis is determined as an eigenvalue by eigenvalue analysis of the variance-covariance matrix for each character type.

Numbers i (i = 1, 2, 3,..., N) are assigned to principal component axes (eigenvectors) and eigenvalues in descending order of eigenvalues. When a line figure is generated from x given by Equation (2) for the i-th principal component axis, a line figure as shown in FIG. 8 is obtained when i = 1 with the numeral “9”.
x = μ + mσ _i φ _i (5)
Here, μ is the average of the character type, φ _i is the i-th eigenvector, and σ _i is the square root (standard deviation) of the corresponding eigenvalue.

Increase m to the positive side and the negative side, and visually check the limit that a person can recognize as a character of that character type. Assuming that the positive and negative limit values are m ⁺ and m ⁻ , the coordinates U _i ⁺ and U _i ⁻ on the principal component axis are as shown in Equation (6).
U _i ⁺ = m ⁺ σ _i (6)
U _i ^- = m ^- σ _i

で与えられる。ｄ（ｘ）＝１が可読域（区分的超楕円体）の表面，すなわち可読限界面を与える．The psychological distance d (x) from the average of the character type of the vector x is

Given in. d (x) = 1 gives the surface of the readable region (piecewise hyperellipsoid), ie the readable limit surface.

Here, _{H i} is the i principal component axis on the coordinate value of x (= (x-μ, φ i)), U when _{U i} is _H i> 0 ⁱ _+, when _H i <0 _{U i} ^- . K determines how much higher-order principal component axes are taken, and it is better to focus on capturing meaningful fluctuations. Taking a very high order axis only picks up the cause of the noise.
FIG. 9 shows a measured value of U _i ⁺ and U _i ⁻ of each principal component axis and an example of a letter shape at the readability limit.

Then, assuming that the component value of n components of the n-dimensional vector x−μ is a uniform random number and a vector whose direction is random is substituted into the equation (7) is D, (x−μ ) / D gives psychological distance 1, so the character pattern with psychological distance k is (x-μ) k / D.
Since the beginning of this vector coincides with the end of μ, the vector starting from the origin O of the feature space is μ + (x−μ) k / D, and this is restored to a character pattern expressed by a curve. Later, thickness is given to make a character pattern.

If the value of the vector k is divided into several intervals as described above and is made a uniform random number, the character group divided by the strength of the character deformation or the grade by quality is obtained. Figure 10 shows an example of characters generated in this way.
Further, it is desirable that a plurality of prototypes (standard patterns) for each character type be provided as necessary, and FIG. 11 shows a prototype for each character type (each prototype is shown as one example of character shape).

  In the recognition experiment using the handwritten New Year's postcard postal code database IPTP CD-ROM1 created by the Postal Service Institute of the Ministry of Posts and Telecommunications, the weighted direction index histogram method developed by the applicants ([Non-Patent Document 4]) is shown in Table 1. As shown in Fig. 4, 99.17% of untrained characters ([Non-Patent Document 4] is experimental data, etc.) with the conventional method (the system that won the highest award in the character recognition contest sponsored by the Postal Service). The recognition rate is not the same because of slightly different), but by adding 2560 characters (256 characters for each character type) generated by the method of the present invention, it is improved to 99.42% is doing.

However, there are few characters that have a large rotation in the actual written characters, and there are few rotated characters in the character group generated only by this method, so some rotation conversion to characters as generated line segments. (Primary transformation) is added.
Wakabayashi Satoshi, Tsuruoka Shinji, Kimura Fumitaka, Miyake Koji, “High accuracy of handwritten digit recognition by increasing the number of dimensions of features,” IEICE Transactions (DII), Vol. J77-D-II, no. 10, pp. 2046-2053, Oct. 1994.

In addition, adding 8000 handwritten postal codes (average of 800 characters for each character type) to the learning character group using the conventional method only shows an improvement of 99.18% and 0.01%. This suggests that it is difficult to obtain a high recognition rate.
Table 2 shows the misreading rate, rejection rate, and recognition performance evaluation scale S used in the contest, and it can be seen that misreading can be greatly reduced without reducing the correct reading rate.

  Figure 12 shows the newly read characters by applying this method. These were human-readable characters with large deformations. FIG. 13 shows typical examples of characters that are misread even if this method is used, but these are often difficult for humans to interpret, and the results will be easily understood by the user of the recognition system. .

  Fig. 14 shows other misreading examples, the main cause of which is due to severe contact between noise and character lines (such as eyelids). If this problem is solved, the correct reading rate will reach 99.66%. It has been confirmed that misreading due to noise is solved by a non-deterministic removal method because the noise has a smaller area than the character part. In this case, the recognition rate reaches 99.47%.

  The ratio of the modified characters created by this method to be added to the original learning character data group by quality or thickness is determined by experiment. It is most effective to add a character group of 0.9-1.0 (occurrence frequency is very low in nature). When reading characters with many contact between character lines, it is better to add thick characters It has been confirmed that the recognition rate increases. In some recognition experiments, it has been confirmed that the correct reading rate exceeds 99.50%.

  In FIG. 15, handwritten numerals are divided into characters to be read, characters that are difficult to recognize, and characters that cannot be (correctly) read, and correspond to the recognition rate obtained in the recognition experiment, but 99.17 obtained by the conventional method. At the correct reading rate of%, things that would be readable by humans were conspicuous, whereas numbers such as 99.42%, 99.47%, and 99.66% were quite close to human recognition rates. Recognize that it is.

(2) Method of using the outline of characters By analyzing the collected outline information of the printed character pattern statistically, extracting the factors governing the difference between typefaces and fonts, and controlling those factors, A large number of pseudo typefaces and fonts are generated and used as a learning character group of the recognition system. By applying the OCR technology applied to scribes published in [Non-Patent Document 5], it is expected to be able to recognize typefaces and fonts and improve the character recognition rate of documents printed in the same font. Is done.
Tsuruoka Shinji, Morita Hiroyuki, Kimura Fumitaka, Miyake Koji, “Free Handwritten Character Recognition with Adaptive Function for Writers”, IEICE Transactions D, Vol. J70-D No. 10, pp. 1953-1960, Oct. 1987.

The pseudo font is generated by the following procedure.
(I) Select an existing standard font by the method described above, and set feature points (refractive points, curvature discontinuities, etc.) and sub-feature points on the outline pixel string of each character type character pattern. (Fig. 4).

(Ii) Input the collected characters of the existing font and find the points corresponding to each feature point on the outline. Next, an original feature vector F having the coordinate values of feature points and sub-feature points as elements is created.
(Iii) The original feature vector F is normalized and set as a feature vector x.

(Iv) Principal component analysis is performed for each character type using the feature vector x, points are provided on each principal component axis at regular intervals in units of standard deviation, and the points are converted into character data.
(V) For the character created by the above means, the fluctuation tolerance is determined by obtaining a limit on the principal component axis that satisfies the condition that the contour lines do not intersect (FIG. 5).

Similar to the case of the character core line, according to the distance formula according to the equation (7), a feature vector is randomly generated in the allowable region for each distance value, and this is converted into character outline data and surrounded by the outline. The area is filled and character image data having shape information is generated.

(Vi) By adding the character image data obtained in (v) to the learning character group of the recognition system in addition to the initially collected character data, multi-font or omni-font compatible OCR is realized.

  If the method of using the letter cord is applied to the classification of letters, the labor cost can be further reduced. Also, considering that it is difficult to correct the numerical values of handwritten numbers in general documents by the context, it will contribute to the spread of handwritten OCR in the field of office work.

  In the field of handwritten OCR, the character recognition of the handwritten Hiragana / Kanji database ETL9 (B) has resulted in a recognition rate exceeding 99.5% for unknown characters, but it is of lower quality than what we write during daily work. The present invention, which opens the way to OCR that can recognize the characters of Japanese characters with practical accuracy, will contribute to further expanding the application range of OCR.

  In addition, the method using the character outline not only opens the way to the omnifont print character OCR, but also mathematical expressions that require accurate typeface recognition in situations where an infinite font exists, such as mathematical expression recognition. It will also play a big role in the field of cognitive understanding.

  It is the figure explaining the lack and bias | inclination of the character form diversity seen in the learning character group which can be collected by the normal method on the feature space.   This is a method for finding a point corresponding to a feature point set in a standard pattern on an input pattern (collected original learning character pattern).

  It is a figure explaining the method of setting the sub feature point which equally divides the character line which exists between feature points, and producing the feature vector F which makes the coordinate of a feature point and a sub feature point an element.   In the method using the outline of a character, the example of the feature point defined on the outline is shown in figure.

  This is a diagram illustrating a phenomenon in which contour lines intersect with each other in a character contour line corresponding to a point sufficiently away from the origin on the principal component axis.   An example of a pseudo font character generated by randomly selecting a point in the existence area of the character type “F” in the feature space is shown.

  This is an example of a pseudo font character corresponding to a point (within an allowable region) on a certain principal component axis in the character type “F”.   This shows an example in which a character is generated at a point moved on the first principal component axis by an integral multiple (moving system number) to the origin and the positive and negative sides with the standard deviation as a unit.

  It shows the range of characters that can be read correctly by humans on each principal component axis and the shape of characters at the limit of legibility.   The example of the character (number) produced | generated at random in the range of the psychological distance 0.5-0.6 and the psychological distance 0.9-1.0 is shown.

  It shows the number of sub-character types (subcategories) set for each character type and typical character shapes.   It is the figure which showed the character newly read correctly by the method of this invention.   A typical example of characters that could not be read correctly even with this method is shown.

  This shows the correct reading rate when there is no noise or blur.   It is a figure for demonstrating the recognition performance achieved by the method of this invention, and the reachable recognition rate of a postal code recognition system.

Table 1

  This is a comparison of the correct reading rate between the method of increasing the size of the character group collected by the normal method and the method of this method.

Table 2

  This is a comparison between the conventional method using the IPTP-CDROM1 created by the Postal Service Institute of the Ministry of Posts and Telecommunications directly as a learning character group and the recognition rate by this method.

Claims (2)

In response to the practical problem of not being able to collect handwritten character samples with sufficient content for actual learning of a handwritten character recognition system, the feature points and sub-feature points (sequence numbers) of typical character-shaped character cores are provided for each character type. And a feature vector formed by arranging the values of the x and y coordinates in the order of the order number, and a point corresponding to the feature point / sub-feature point is selected from the pixel string representing the core line of the learning character. Extracted and based on the principal component analysis result for the feature vector obtained from those x and y coordinates, the human recognizable area in the feature space (the range that humans can recognize as the character type) is visually piecewise hyperelliptic. A method for obtaining a body and a character recognition method performed using a psychological distance defined therein are disclosed in [Non-Patent Document 1]. However, the present invention further includes a distance as an attribute of a feature point as necessary. Value (statement In addition to extending the method to add z), the piecewise hyperellipsoid is divided into several shells according to psychological distance, and the character pattern is reversely generated from the randomly generated points. Thus, the present invention relates to a method of automatically generating a sufficient number of character patterns for each deformation strength or quality grade and enriching the content of learning characters in addition to the learning character samples collected first.
[Non-Patent Document 1]
Tsuruoka Shinji, Murase Akihiko, Kimura Fumitaka, Yokoi Shigeki, Miyake Koji: “Free Handwritten Katakana Character Recognition Using Human Character Type Identification Criteria”, IEICE Transactions, Vol, J68-D, No. 4, pp. 781-788, 1985. In order to enhance the contents of the learning character group, mainly multi-font or omnifont print character recognition system, important feature points (corner points, curvature discontinuities, etc.) Define and perform principal component analysis in the same way as in [Non-Patent Document 1], determine the allowable range of characters of the character type by mathematical constraints or visual observation on each principal component axis, Using a distance value in the same format as the psychological distance for a piecewise hyperellipsoid with a long axis, [This is related to a method for generating a variety of fonts in the same way as claim 1 and enriching the contents of the learning sample. is there.

Images