JPH0338631B2 - - Google Patents

Description

[Detailed Description of the Invention] (a) Field of Industrial Application The present invention relates to a character recognition system for recognizing handwritten kanji, and relates to a recognition ranking determination method for candidate character type categories.

(b) Conventional technology First, in a character recognition system, the degree of similarity between a feature pattern extracted from an input character pattern and a standard feature pattern for each character type category registered in advance in a dictionary section is calculated, and the degree of similarity is calculated. n candidate character type categories with a large value are selected. Then, in addition to outputting the candidate character type category with the highest degree of similarity as a recognition result, in order to correct misrecognition, the selected n candidate character type categories are ranked from 1st to nth in order of the degree of similarity. Decide on the recognition ranking up to the first rank.

However, as mentioned above, using only similarity to determine the recognition ranking often results in false recognition, so after selecting multiple candidate character categories based on similarity, some post-processing is performed to determine the recognition ranking. Now I can think of a way to do that.

Conventionally, as post-processing, JP-A-59-
As disclosed in Japanese Patent No. 32082, there are methods that perform grammatical processing, and methods that determine whether the characters before and after the recognized character are kanji, katakana, or hiragana, as in Japanese Patent Application Laid-open No. 59-27381. Something was proposed.

(c) Problems to be solved by the invention In the conventional technology, grammatical processing is performed as post-processing, so the knowledge section such as a grammatical dictionary becomes enormous, and furthermore, the processing content becomes extremely complicated. Furthermore, with the method of determining whether the preceding and succeeding characters are kanji or katakana, if the selected candidate character category is only kanji or hiragana, it cannot be expected to improve the recognition rate. .

Therefore, in order to improve the recognition rate in a short processing time without requiring a huge knowledge department, the applicant set priorities for each character type category according to the learning stage or frequency of school education. I tried to determine the recognition order of multiple candidate character type categories using this priority and similarity, but
Since the recognition rate changes depending on the balance between similarity and priority, the problem has been how to properly set this balance.

(d) Means for solving the problem The present invention calculates the difference in similarity and the difference in priority between candidate character type categories, and if the difference in similarity is larger than a predetermined value, the difference in priority is calculated. The recognition order of the candidate character type category is changed by swapping the recognition order based on the priority only when the difference in priority is greater than a predetermined value, and not replacing the recognition order based on the priority when the difference in priority is smaller than the predetermined value. It was decided to do so.

(e) Effect In the present invention, in candidate character type categories whose similarities are separated by a certain degree, when their priorities are close, the priorities are ignored and the recognition order is determined only by the similarity. The recognition order based on the priority is determined only when the degrees are significantly different. Therefore, the recognition order is determined by effectively utilizing both the similarity and the priority, and the recognition rate is improved.

(F) Embodiment Figure 1 is a block diagram of a character recognition system to which the present invention is applied, and 1 is a character observation system that reads characters written on an input manuscript and outputs the reading results as a binary character pattern. 2 is a feature extraction unit that extracts feature patterns from input character patterns; 3 is a dictionary unit that stores standard feature patterns for each character type category; and 4 is a feature extraction unit that matches extracted feature patterns with standard feature patterns; This is a pattern matching unit that calculates the similarity of patterns.

The character type categories in the dictionary section 3 are divided into categories according to frequency or learning stage of school education, and a priority is determined for each category set. For example, as shown in Figure 2, the glyph categories learned in the first to third years of elementary school are set to category sets 1 and 3a, the glyph categories learned in the fourth to sixth years of elementary school are set to category sets 2 and 3b, and the glyph categories learned in junior high school and above. All character categories are divided into three category sets, such as category sets 3 and 3c, and priorities 0 to 2 are assigned to category sets 1 to 3 in order.

The pattern matching unit 4 includes three calculation units 4a to 4c corresponding to category sets 1 to 3, respectively, and each calculation unit selects n candidate character type categories from each category set in descending order of similarity. is selected, and its character type code and the degree of similarity as a calculation result are stored in the candidate memory 5. On this occasion,
The calculation unit adds the priority of the corresponding category set to the character type code and similarity, and these three pieces of information are stored in the candidate memory 5 as information for each candidate character type category. In this way, candidate memory 5 stores n candidate character type categories from each category set, for a total of 3n candidate character type categories.

Further, in FIG. 1, reference numeral 6 denotes a knowledge section that stores the relationship between similarity and priority, and 7 refers to the contents of this knowledge section 6 to determine the 3n candidate character type categories stored in the candidate memory 5. It is a clustering control processing unit that determines the recognition ranking of the top n characters among them and stores the character type codes in the result memory 8 in order of recognition ranking, and the answer output control unit 9 determines the character type with the first recognition ranking. The code is output as a recognition result to a character display device such as a word processor or a personal computer, and the type of character is displayed. If the operator gives an instruction of incorrect recognition, the answer output control section 9 sequentially outputs the second and lower character type codes and controls the output so that the correct recognition result is displayed.

Next, the processing of the clustering control processing section 7 and the contents of the knowledge section 6 will be explained in more detail.

In this embodiment, the city block distance d is used as the degree of similarity, and it is assumed that the smaller this distance is, the greater the degree of similarity is. Threshold values D〓, D〓, D〓 (D〓<D〓<D〓) are stored. As shown in the flowchart of FIG. 3, the clustering control processing unit 7 first selects n candidate character type categories G from the 3n candidate character type categories stored in the candidate memory 5 in descending order of distance. ₁ , G ₂ . . . G _o are selected and stored in order in the working memory 7a. _Then , by comparing _distances d ₁ , d _{2 .} A, B, as shown in Figure 4,
Classification is performed into four classes, C and D. That is, the clustering control processing unit 7 is provided with class pointers 10a to 10d for each class, which indicate to which address in the working memory 7a candidate character type categories belonging to that class are stored. As shown in the flowchart, the contents of each pointer 10a, 10b, 10c, 10d
TA, TB, TC, and TD are incremented when comparing distances. For example, class A is assigned to addresses 0 to 1 of the working memory 7a, and class B is assigned to addresses 2 to 4 of the working memory 7a.
Class, address 5-6, C class, address 7
If the D class candidate character type category is stored in , the content of each pointer will be
TA=2, TB=5, TC=7, TD=8.

By the way, as shown in FIG. 4, the knowledge section 6 stores in advance whether or not the order of priority can be changed and the conditions for each class, and the clustering control processing section 7 refers to the contents of this knowledge section 6. Then, the recognition ranking of the plurality of candidate character type categories stored in the candidate memory 5 is determined.

Here, the character type code is M _I , the city block distance is d _I , and the priority is P〓(I=A, B,...H, K, L,
...R, S, T, ...Z, P _I = 0, 1, 2), the candidate character type category G _I is expressed as (M _I , d _I , P _I ). As shown in FIG.
is selected, and the relationship between each city block distance d _I is
Assuming that d _S < d _K < d _L < d _T < d _A < d _U < d _M < d _B <..., the working memory 7a of the clustering control processing section 7 has the following information as shown in FIG. , eight candidate character type categories G _S to G _B are selected and stored in descending order of distance d _I .

Therefore, the relationship between each city block distance d _I of the candidate character type category and the threshold values D〓, D〓, D〓 is, for example,
If d _K <D〓<d _L , d _A <D〓<d _U , d _B <D〓, the clustering processing unit 7 performs classification, and each class pointer 10a to 10d
The contents are TA=2, TB=5, TC8, TD=0. That is, the candidate character categories G _S , G _K are in the A class, the candidate character categories _GL , G _T , and G _A are in the B class, and the candidate character categories G _U , _GM ,
G _B is classified into C class. Next, the clustering control processing section 7 refers to the knowledge section 6 and rearranges the ranking of candidate character type categories in each class. At this time, the character type codes of candidate character type categories _G1 to _G8 at addresses 0 to 7 in the working memory 7a are set to M(1) to M(8), the city block distance is set to d(1) to d(8), Set the priority to P(1) to P(8),
The character type code of the character type category whose final recognition rank is i is defined as m(i).

Then, first of all, since the character type category of class A is not replaced by priority, character type codes M _S and M _K are determined as the first m (1) and the second m (2), and their ranks are The result is written to memory 8. The following three character type codes for B class M _L ,
Regarding M _T and M _A , the processing shown in the flowchart of FIG. 6 is performed. That is, in the case of B class, the initial values are m(i)=M(TA+1), i=TA+1,
Set l=TB and τo=a.

In the above example, since TA=2 and TB=5, first the character type code of m(3) is determined. Here, d _T −d _L >a, d _A −d _T >a, and therefore d _A −d _l >a.

First, the distance difference τ=d _T −d _L between G _L and G _T is calculated and compared with a predetermined value a. In this case, since τ>a, next the priority difference t=P _L -P _T is calculated and compared with the predetermined value 1. In this case, P _L has a higher priority than P _T , so it becomes -1, and m(3), that is, M _L and M _T are not swapped. Next, the difference in distance between G _L and G _A , γ = d _A - d _L , is calculated and compared with a predetermined value a,
Since τ>a, the priority difference t=P _L -P _A is calculated and compared to see if τ is larger than a predetermined value 1. In this case, P _A has higher priority than P _L , but t=1
Therefore, priority swapping between M _L and M _A is not performed. Therefore, at this point, the third recognition rank m(3) is determined to be M _L , and M(3) to M(5) are
M _L , M _T , and M _A remain the same. And then m
(4) The character type code is determined. In this case too,
The distance difference τ=d _A −d _T between G _T and G _A is calculated and compared with a predetermined value a. Since τ>a, the difference in priority t=
P _T -P _A is calculated and compared to see if it is larger than a predetermined value 1. In this case, the difference in priority is t=2
Therefore, the order of M _T and M _A is exchanged based on the priority, and the character type code m(4) which ranks fourth in the recognition order is determined to be M _A. When the processing for m(4) is completed, i=5 and l=TB=5
Therefore, the ranking replacement process based on the priority in the B class is completed, and the recognition rankings of the third to fifth places are determined as M _L , _MA , and M _T . still,
If the order is simply determined based on priority within class B, it becomes M _A , M _L , and M _T .

However, for example, d _T −d _L ≦a, d _A −d _T ≦a,
If d _A − d _L ≦ a and the distance between each candidate character type category is closer than in the above example, the recognition in this case The ranking is M _A , M _L ,
It is determined that M _T.

In this way, the determined character type code M _L ,
M _A and M _T are sent to the result memory 8 and stored in this order following M _S and M _K.

Next, the clustering control processing section 7 moves on to C class processing.

In the case of C class, as an initial value, m(i)=M
(TB+1), i=TB+1, l=TC, and τ _p =b, and the process shown in the flowchart of FIG. 6 is performed as in the case of the B class. For example, d _M
−d _U >b, d _B −d _M >b, so d _B −
If d _U > b, then τ = d _M − d _U > b,
The priority of M _M is higher than the priority of M _U , but since the difference in priority is t = P _U − P _M = 1≦1, M _U and M _M are not swapped based on priority. . Next, τ=d _B −d _U is calculated such that τ>b and
Since P _U − P _B = 2 > 1, M _U and M _B are exchanged according to priority, becoming M _B , M _M , M _U , and the 6th character type code m (6) is determined as M _B. be done. and,
Since M _M and _{M U} have a higher priority, their rankings remain unchanged. Therefore, the 6th to 8th character type codes are determined as M _B , M _M , M _U ,
The results are stored in the result memory (8) in this order.

When the rankings are determined as described above, eight character type codes are stored in the result memory 8 in the rankings shown in FIG. 5C.

In this way, in the present invention, when the difference in the distance between two candidate character type categories is greater than a predetermined value, and the priorities are close, the rankings are not swapped based on the priorities, and the priorities are far apart. Reordering based on priority is performed only when there is a problem.

In this embodiment, the priority is 0, 1, and 2.
Although the steps are described above, the present invention is applicable even if the steps are increased. Furthermore, in this embodiment, the dictionary section 3 is divided into category sets according to priorities, but if priority information is added to each character type category,
It is not necessarily necessary to divide the dictionary section 3. Furthermore, in a more specific process, contrary to the flowchart of FIG. 6, the distance difference may be checked after the priority difference is checked.

Note that the candidate character type categories selectively stored in the working memory 7a do not necessarily include all of the items in each class, and there are cases where all the candidate character type categories are only B class or C class, and in this case, these All recognition rankings are determined only within the same class.

According to experiments, when ranking all candidate character categories in the same class based on priority, the recognition rate for the first place was approximately 90.0%, but when the present invention is applied, the recognition rate increases. This improved to approximately 92.5%.

(G) Effects of the Invention According to the present invention, it is possible to determine the recognition order of candidate character type categories by effectively utilizing both similarity and priority, and therefore, the recognition rate is improved.
Furthermore, the knowledge section does not require a huge capacity, and furthermore, ranking can be determined in a short processing time.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a character recognition system to which the present invention is applied, FIG. 2 is an explanatory diagram showing the contents of a category set, and FIG. 3 is a flowchart showing the contents of the classification process of the clustering control processing section.
FIG. 4 is an explanatory diagram showing the contents of the knowledge section, FIG. 5 is an explanatory diagram showing a specific example of recognition ranking determination, and FIG. 6 is a flowchart showing the contents of the ranking replacement process of the clustering control processing section. Explanation of main figure numbers, 3...Dictionary section, 4...Pattern matching section, 5...Candidate memory, 6...Knowledge section, 7...
Clustering control processing unit, 8...Result memory, 9
...Answer output control section.

Claims (1)

[Claims] 1 Calculate the degree of similarity between the feature pattern extracted from the input character pattern and the standard feature pattern for each character type cadecory registered in advance in the dictionary section, select a plurality of candidate character type categories, and select the character type category from the character type category. In a method in which a priority is determined for each of the candidate character type categories and a recognition order is determined based on the similarity and priority, the difference in similarity and the difference in priority between the candidate character type categories are calculated, and the If the difference in priorities is larger than a predetermined value, the recognition order based on priority is changed only when the difference in priority is greater than a predetermined value, and when the difference in priority is smaller than a predetermined value, the recognition order based on priority is changed. A recognition ranking determination method characterized by not replacing .