JPH03260888A - Standard pattern preparing method - Google Patents

Abstract

PURPOSE:To improve the rate of recognition by carrying out a non-linear transformation upon each feature element and preparing a standard pattern so as to suppress the outstanding feature elements in the total sum pattern of adding the feature elements extracted from each sample character picture when preparing the standard pattern. CONSTITUTION:When the standard pattern is prepared and the total sum pattern of adding the feature extracted results of respective samples contains a few outstanding feature elements, the operation of extracting (1/a)-th root [(a)>1.0] or an operation that uses a function f(x) resembling to the logarithm function is applied to each feature element of this total sum pattern. The non-linear transformation of density is executed and afterwards, a linear transformation of the density is executed as needed. Here, the f(x) is a convex function which satisfies the condition given by equation I for an arbitrary (x) (>=0) and further satisfies the condition given by equation II. In the inequality I, DELTAx is an arbitrary positive real number and expresses the increment of the function. Thus, the stable standard pattern immune to the fluctuation of the character can be obtained and the rate of recognition is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a method for creating a standard pattern used in 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 character image input, step S2 shows feature extraction, and step S3 shows character type discrimination.

Here, the character type discrimination unit (step 33) calculates the degree of similarity between a standard pattern prepared in advance for each character to be recognized and the feature pattern extracted from the input image, and calculates the degree of similarity between the character type of the input character. is determined.

The degree of similarity G is obtained from the following equation, where the standard pattern and the characteristic pattern of the human image are respectively n-dimensional vectors S (n) and P (n).

Here, (s, p) = Σ = S (i) P (i), and the inner product of the vectors is calculated. (S, S), (P,
P) are called autocorrelation values of the standard pattern and the characteristic pattern, respectively.

A general procedure for creating a standard pattern used in this manner is shown in FIG.

As seen in the figure, features are extracted from the multiple samples obtained in step Sl in step S2, the results are added for each element in step S3, and the necessary A standard pattern is obtained by performing density conversion according to step S5. The density here refers to the size of the feature element. The same shall apply thereafter.

Generally, density conversion is performed by uniformly converting the size of other elements in proportion to the largest element in each standard pattern so as to equalize it to a predetermined value F. Total pattern J(n
), the standard pattern S, (n)
The i-th element of is 5J(i)=Hj(i)・F/MJ
......obtained as (1).

This is to effectively use the limited bit width due to the number of effective digits of the device that performs the recognition process and the storage area of the standard pattern. Alternatively, in order to match the density order of the characteristic pattern of the input image and the standard pattern, a process is performed in which each element of the total pattern is divided by the number of samples.

[Problem to be solved by the invention]

However, the standard pattern created in this way may end up having a small number of characteristic elements with outstanding values.
In that case, it was found that these prominent elements had a significant influence on character type discrimination. In other words, it becomes vulnerable to character fluctuations and the recognition rate becomes low.

Figure 4 is a graph showing the density distribution in a standard pattern created using the conventional method.
It can be seen that there is one pixel at each of 55.208 and 207, which is a feature element with a small number of outstanding values.

The purpose of the present invention is to improve this problem and, in the case of a standard pattern having feature elements with a small number of outstanding values, to prevent the recognition rate from decreasing by effectively processing those feature elements. The objective is to provide a method for creating standard patterns.

[Means to solve the problem]

When creating a standard pattern, if the summation pattern obtained by adding up the feature extraction results of each sample has a small number of prominent feature elements, the a-th root (a
>1.0) and functions similar to logarithms f (x)
By performing calculations using , non-linear conversion of density as shown in FIG. 6 is performed, and then linear conversion of density is performed as required.

f (x) is a convex function that satisfies the following conditions for any X (≧0).

f(x) −f(x−711x)≧f(x+,dx)−
f(x)≧0 (2) Furthermore, the following conditions are satisfied.

Here, Ax is any positive real number and represents the increment of the function.

[Effect]

By applying the non-linear transformation as described above, it is possible to suppress the prominent feature elements to a greater extent than the non-protruding feature elements, so it is possible to obtain a stable standard pattern that is resistant to character fluctuations. The recognition rate can be improved in recognition processing using processing similar to conventional methods.

[Embodiment] FIG. 1 is a flowchart showing a standard pattern creation method as an embodiment of the present invention. As seen in the figure, feature extraction processing is sequentially performed and added to each sample image. As shown in step S5, the a-th root of each characteristic element (a is a real number greater than 1.0) is taken for this summation pattern, and a new summation pattern is obtained as a result. That is, the operation of taking the a-th root means performing nonlinear concentration conversion by the operation using the nonlinear function f (x) as shown below.

f (x)=χI/11...
...(4) After that, density conversion, which has been conventionally performed (step S6 in FIG. 1 or step S5 in FIG. 3)
Then, density conversion is performed using equation (1) above. Here, in order to compress the bit length of each feature element to 8 points, uniform density conversion is performed with F=255. The difference from the conventional method is that nonlinear transformation is performed by taking the 8th root of each feature element.

An example of a standard pattern when using the conventional method and the present method (a = 4) is an image normalized to a (16x16) mesh for the character "a" with 80 learning samples and a feature extraction method. When creating and comparing contour pixels using a method (Patent Application No. 71757 of 1999) in which the contour pixels of In this case, the prominent feature elements are kept low, and then the density is converted to increase the bulk of the whole image, so you can see that the overall density is higher.

When a standard pattern is created for 75 hiragana characters using the method according to the present invention, and the same recognition processing as before using similarity as a discriminant is performed on 80 learning samples and 75 hiragana character types, the conventional method The recognition rate was 95.5% compared to 90.5% when using the standard pattern created by
% of the recognition rate. Please refer to FIG. 5, which shows the relationship between the variable a (a of the 8th root) and the recognition rate.

The present invention does not depend on the feature extraction method described above,
The effectiveness of background features using propagation stop processing (Japanese Patent Laid-Open No. 59-792) has also been confirmed.

Here, in equation (4), a=4, but the optimal value of a is considered to vary depending on the field to be recognized, the feature extraction method, and the type of discriminant. Furthermore, nonlinear transformation that takes the logarithm of each characteristic element of the summation pattern is also effective, as is nonlinear transformation that takes the 8th root. That is, f
The function shown below is used as (x).

f(x)=log+o(x+1)−
- (5) However, as a in equation (4) is increased, the feature pattern approaches a binary image and the recognition rate decreases.Therefore, each element of the feature pattern extracted from the sample image is raised to the a power. By adding them together to obtain the summation pattern, it is possible to keep the values of prominent feature elements low while preserving the density distribution in the summation pattern to some extent.

〔Effect of the invention〕

According to the present invention, when creating a standard pattern, a nonlinear transformation is applied to each feature element to suppress any prominent feature elements that may occur in the summation pattern of feature elements extracted from each sample character image. By applying this method to create a standard pattern, there is an advantage that the recognition rate is improved. When using the fourth root as a nonlinear transformation, the recognition rate improves from 90.5% to 95.5%, and the recognition rate is 53%.
% improvement was seen.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing one embodiment of the present invention, and FIG.
The figure is a flowchart showing a general character recognition procedure using pattern matching. FIG. 5 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 characteristics of the nonlinear transformation according to the present invention. Explanation of symbols S1 to S6...Step 1 Figure 2 Figure 5 Figure 5

Claims (1)

[Claims] 1) In a standard pattern creation method in which a standard pattern used for character recognition processing by pattern matching is created from multiple sample character images, the feature extraction results from each sample character image are added for each feature element to create a summation pattern. After obtaining the summation pattern, the a-th root of its size (where a>
1.0) or a logarithm, which performs an operation that performs nonlinear transformation, and obtains a summation pattern consisting of the result of the operation.