JP2749692B2 - Standard pattern creation method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for creating a standard pattern used for character recognition processing by pattern matching.

[Conventional technology]

FIG. 2 shows a general character recognition processing procedure using pattern matching. In the figure, step S1 shows input of a character image, step S2 shows feature extraction, and step S3 shows character type determination.

Here, the character type determination unit (step S3) calculates a similarity between a standard pattern prepared in advance for each character to be recognized and a feature pattern extracted from the input image, and calculates the character type of the input character. Is determined.

The similarity G is obtained by using an n-dimensional vector S (n), P (n) as a standard pattern and a feature pattern of an input image, respectively.
It is obtained from the following equation.

Here, (S, P) = Σ _i S (i) P (i), which represents the inner product of the vectors. (S, S) and (P, P)
These are called autocorrelation values of standard patterns and feature patterns.

FIG. 3 shows a general procedure for creating a standard pattern used in this manner.

As can be seen from the figure, feature extraction from the multiple samples obtained in step S1 is performed in step S2, and the result is added for each element in step S3, and the necessary sum is obtained for the total pattern obtained in step S4. Is performed in step S5 to obtain a standard pattern. The density here refers to the size of the characteristic element. The same applies hereinafter.

Generally, as the density conversion, a process of uniformly changing the size of other elements in proportion to the predetermined value F of the largest element in each standard pattern is performed. Sum pattern H _j of the j-th character type among a plurality of character types
If the maximum element of (n) is M _j , the standard pattern S
The i-th element of _j (n) is obtained as _Sj (i) = _Hj (i) .F / _Mj (1).

This is because a limited bit width is effectively used due to the number of significant digits of the device that performs the recognition process and restrictions on the storage area of the standard pattern. Alternatively, in order to match the order of the density of the characteristic pattern of the input image with the density of the standard pattern, a process of dividing each element of the total pattern by the number of samples is performed.

[Problems to be Solved by the Invention] By the way, the standard pattern created in this way may have a small number of feature elements of prominent values as a result, in which case, these prominent elements It has been found that this has a great effect on the discrimination of the character type. In other words, the character is less susceptible to fluctuations and the recognition rate is lower.

FIG. 4 is a graph showing the density distribution in the standard pattern created by the conventional method.
It can be seen that there is one pixel at each of 255, 208 and 207, which is a feature element with a small number of prominent values.

An object of the present invention is to improve such a problem, and in the case of a standard pattern having a small number of feature elements having a prominent value, by processing those feature elements well so as not to reduce the recognition rate. Another object of the present invention is to provide a method for creating a standard pattern.

[Means for solving the problem]

When creating a standard pattern, if the sum pattern obtained by adding the feature extraction results of each sample has a small number of prominent feature elements, an operation that takes the a-th root (a> 1.0) for each feature element of the sum pattern Or a function f (x) similar to a logarithm
, A nonlinear conversion of density as shown in FIG. 6 is performed, and then a linear conversion of density is performed as necessary.

f (x) is a convex function that satisfies the following condition for an arbitrary x (≧ 0).

f (x) −f (x−Δx) ≧ f (x + Δx) −f (x)
≧ 0 (2) Further, the following condition is satisfied.

Here, Δx is an arbitrary positive real number and represents an increment of a function.

[Action]

By performing the non-linear conversion as described above, it becomes possible to further press down the prominent feature elements more than the other feature elements, so that it is possible to obtain a stable standard pattern that is strong against character fluctuations, The recognition rate can be improved in the recognition processing using the same processing as the related art.

〔Example〕

FIG. 1 is a flowchart showing a standard pattern creating method as one embodiment of the present invention. As can be seen from the figure, feature extraction processing is sequentially performed on each sample image and added. For this sum pattern, as shown in step S5, the a-th root of each feature element (a is a real number larger than 1.0) is obtained, and a new sum pattern is obtained as a result. In other words, the calculation that takes the a-th root means performing a non-linear density conversion by a calculation using a non-linear function f (x) as described below.

f (x) = x ^{1 / a} (4) Thereafter, the above-mentioned equation (1) is used as density conversion (step S6 in FIG. 1 or step S5 in FIG. 3) which has been conventionally performed. Density conversion is performed. Here, uniform density conversion is performed with F = 255 in order to compress the bit length of each feature element to 8 bits. The difference from the conventional method is that nonlinear transformation is performed by taking the a-th root of each feature element.

An example of a standard pattern using the conventional method and the present method (a = 4) is as follows. The character “A” is normalized to a (16 × 16) mesh as a feature extraction method using 80 learning samples. Created using a method (Japanese Patent Application No. 71757) that assigns the contour pixels of the image to a plurality of feature planes (for example, 15 planes) based on the relative distance between the contour pixels sandwiching the white pixels. In the latter, since the prominent characteristic elements are suppressed low and thereafter the density is converted so that the whole is raised, it can be seen that the density is increased as a whole.

When a standard pattern is created for 75 characters of Hiragana by the method according to the present invention, and the same recognition processing as in the past using similarity as a discriminant is performed for 80 characters of a learning sample and 75 characters of Hiragana, the conventional method is used. The recognition rate was improved to 95.5%, compared to the recognition rate of 90.5% when using the standard pattern created in. The relationship between the variable a (a of the a-th root) and the recognition rate is the fifth.
Please refer to the figure.

The present invention does not depend on the above-described feature extraction method, and its effectiveness has been recognized even for background features using a propagation stop process (Japanese Patent Laid-Open No. 59-792).

Here, in the above equation (4), a = 4.
The optimal value of is the field to be recognized, the feature extraction method,
It is thought that it depends on the type of discriminant. Further, the non-linear transformation taking the logarithm of each feature element of the sum pattern is also effective as the non-linear transformation taking the a-th root. That is, f
The following function is used as (x).

f (x) = log ₁₀ (x + 1) (5) However, as a in the above equation (4) increases, the feature pattern approaches a binary image, and the recognition rate decreases.
Therefore, by adding the respective elements of the feature pattern extracted from the sample image to the power a and adding the sum to obtain the sum pattern, the value of the prominent feature element can be kept low while the density distribution in the sum pattern is preserved to some extent. Can be.

〔The invention's effect〕

According to the present invention, at the time of creating a standard pattern, in the sum total pattern of the characteristic elements extracted from each sample character image, a non-linear transformation is performed on each characteristic element such that a prominent characteristic element that may occur is suppressed. There is an advantage that the recognition rate is improved by creating the standard pattern. When the fourth root is used as the nonlinear transformation, the recognition rate is improved from 90.5% to 95.5%, and the recognition rate is improved by 53%.

[Brief description of the drawings]

FIG. 1 is a flowchart showing one embodiment of the present invention, and FIG.
FIG. 3 is a flowchart showing a general character recognition procedure using pattern matching, FIG. 3 is a flowchart showing a conventional standard pattern creation method, FIG. 4 is a graph showing a standard pattern density distribution obtained by a conventional method, and FIG. FIG. 6 is a characteristic diagram showing the relationship between the variable a and the recognition rate, and FIG. 6 is a characteristic diagram showing an example of the characteristic of the non-linear conversion according to the present invention. Explanation of reference symbols S1 to S6 ... Step

──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Tetsuo Kiuchi 1 Fujimachi, Hino-shi, Tokyo Fujifacom Control Co., Ltd. (72) Akiko Konno 1 Fujimachi, Hino-shi, Tokyo Fujifacom Control Ltd. (56) References JP-A-1-116892 (JP, A) JP-A-2-242391 (JP, A) JP-A-60-160488 (JP, A) JP-A-49-103536 (JP, A) JP-A-63-313283 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G06K 9/68

Claims (1)

(57) [Claims] In a standard pattern creation method for creating a standard pattern used for character recognition processing by pattern matching from a plurality of sample character images, a feature extraction result from each sample character image is added for each feature element. After obtaining the sum pattern, a non-linear operation such as an operation for taking the a-th root (a> 1.0) of the size or an operation for taking a logarithm is performed on each of the characteristic elements constituting the obtained sum pattern. And performing a density conversion to obtain a new total pattern from the calculation result, thereby suppressing the value of the prominent feature element.