JPH11224309A - Character recognition method and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform classification at a high speed while keeping a broad classification rate at a fixed value by successively changing an initial value so as to minimize an operation amount required for hierarchical broad classification. SOLUTION: Relating to a broad classification part 4, a first stage classification circuit 41 calculates a distance between the n1 pieces (n0 >n1 ) of features specified in a broad classification constitution decision part 9 in the n0 pieces of the features of an input character pattern obtained from a feature extraction part 3 and the n1 pieces of the features of respective categories stored in a broad classification dictionary 5 and narrows the categories to c. pieces in the ascending order of the distance. A second stage classification circuit 42 calculates the distance between the n2 pieces (n0 >n2 >n1 ) pieces of the features specified in the broad classification constitution decision part 9 from the n0 pieces of the features and the n2 pieces of the features of the c1 pieces of the categories outputted from the first stage classification circuit 41 and narrows the categories to c2 pieces (c2 <c1 ) in the ascending order of the distance. The hierarchical broad classification of reducing a candidate number while increasing a feature number through the number of stages is performed similarly thereafter.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method and apparatus, and more particularly, to a character recognition method and apparatus capable of performing large-scale classification at high speed.

[0002]

2. Description of the Related Art In a conventional character recognition method and apparatus, a method of recognizing an input character pattern by comparing the input character pattern with a dictionary created for each category is adopted in many cases. If the character type to be recognized is of the order of several tens, such as numbers and symbols, the calculation time is relatively short even if the collation operation is performed between the input character pattern and all categories. If the number of character types is in the order of thousands, such as kanji, the calculation time is more than one hundred times that of numbers and symbols, even if it is simply estimated.

[0003]

Therefore, in order to reduce the calculation time, first, a large classification is performed to narrow down the candidates to a relatively small number.
Next, a processing method for identifying the narrowed candidates is adopted.

[0004] As a well-known method in the large classification, a logic operation is performed between a predetermined feature and a threshold value among features extracted from an input character pattern as seen in an identification tree or combinational logic. There is a method in which a category satisfying is used as a classification result. Although this method has the advantage that the processing is fast, since the major classification is performed by the threshold logic of a specific feature, if the feature value of the feature fluctuates due to deformation or the like and takes a value outside the threshold, the logic is reduced. Is no longer satisfied, resulting in misclassification.

As a solution to the above problem, there are a method of increasing the number of logics by relaxing the threshold value, and a method of leaving the category as a candidate until the number of unsatisfied logics reaches a predetermined value. However, since the solution has a large logic, the processing time is increased, and the solution is not based on a principle capable of coping with a sudden change in a feature value, and stable classification is still difficult.

As another known method, there is a method of performing a large classification by calculating a distance, a similarity, and the like between a relatively large number of features obtained from an input pattern and a standard pattern. In this method, since comprehensive judgment using a plurality of features is performed, even if a small number of feature values fluctuate, if the variation of other many feature values is small, the sum of the two values is within a normal variation range. And has the advantage of stable classification.
However, there is a problem in that vector operations such as distance and similarity require a longer calculation time than threshold logic.

As a method of reducing the calculation time, a plurality of classification circuits are connected in series to adopt a hierarchical structure. First, the first hierarchical level roughly narrows down candidates using a small number of features, and the next hierarchical level reduces the number of features. There is a hierarchical large classification method that obtains a desired number of candidates in the last hierarchical level by repeating the above processing to further narrow down candidates by increasing the number of candidates. However, in this method, if the number of features in each layer is increased and the number of candidates is increased, the large classification rate is increased, but the calculation time is increased. However, there is a contradictory property that the large classification rate decreases. In addition, the classification rates of the respective layers are also related to each other, and it is not easy to appropriately arrange the load on the classification of the respective layers.

As described above, since there is a complicated relationship between hierarchies in and between hierarchies of the above method, the use of each stage for maintaining the large classification rate at a constant value and minimizing the calculation time is used. There is a problem that a method of designing a large classification unit for finding an optimal solution of the number of features and the number of candidates to be narrowed down has not been established.
Also, depending on the application of the character recognition device, there are various requests from the case where the recognition speed is as low as possible, even if the recognition speed is slow, to the case where at most high speed recognition is required for the misreading. There is also a problem that it is difficult to respond to the above request by adjusting the above.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to perform high-speed classification while maintaining a large classification rate at a constant value, and perform high-precision recognition or high-speed recognition. An object of the present invention is to provide a character recognition method and apparatus that can respond to various requests.

[0010]

The present invention achieves the above object by calculating a classification measure between all or some of the features of an input character pattern and all or some of the features of a standard dictionary. A classification circuit for reducing the number of candidates to be classified is configured to form a hierarchical structure, and the number of features used for classification of the classification circuit is increased while the processing proceeds from a preceding layer to a subsequent layer. Hierarchical large classification means for reducing the number of candidates to be classified and setting of the number of features and the number of candidates of each layer in the hierarchical large classification method are performed by first setting the initial number of features and the number of candidates in each layer. Value is set,
Next, a large classification configuration determining unit that sequentially changes the initial value so as to minimize the amount of computation required for the hierarchical large classification under the constraint that the large classification rate is held at a predetermined constant value. The setting of the number of features and the number of candidates for each layer in the hierarchical large classification method is such that when a target value of the large classification rate is specified, the target value is replaced with the fixed value and the sequential change is performed. When the target value of the large classification processing speed is designated, the constant value is obtained from the relationship between the previously obtained large classification processing speed and the large classification rate, and the value is sequentially changed. And a trade-off function setting means capable of selecting a type designation.

The principle of the large classification configuration determining means of the present invention is based on the principle of finding the sub-optimal feature number and candidate number by sequentially changing the initially set number of features and number of candidates in each hierarchy. Even when there is a complicated relationship between hierarchies or between hierarchies, a large classification unit can be easily constructed. Further, by selecting and specifying a desired large classification rate or processing speed by the trade-off function setting means, a large classification unit satisfying the above performance can be constructed, so that tuning according to the application can be performed.

[0012]

Next, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of a character recognition apparatus according to an embodiment of the present invention.

The character recognition device of this embodiment includes a character pattern input unit 1, a preprocessing unit 2, a feature extraction unit 3, a large classification unit 4, a large classification dictionary 5, a classification unit 6, a classification dictionary, 7, a recognition result memory unit 8, a large classification configuration determining unit 9, and a trade-off function setting unit 10.

A character pattern input unit 1 captures a character pattern by an input device such as a scanner or a television camera, a preprocessing unit 2 performs normalization and noise removal, and a feature extraction unit 3 stores features to be used for recognition. Extract. The major classification unit 4 calculates a distance for each category between the features of the character pattern and the features stored in the major classification dictionary 5, and outputs the obtained candidates. The identification unit 6 calculates a distance between the features of the character pattern and the features stored in the identification dictionary 7 for the candidates output from the large classification unit 4 and uses the obtained candidates as the final result as the recognition result. Output to the memory unit 8.

Next, the operation of the large classification unit 4 of the present invention will be described with reference to FIG. FIG. 2 is a functional block diagram of the large classifying unit 4 showing one embodiment of the present invention.
, A k-th stage classification circuit 43,..., A K-th stage classification circuit 44.

The first-stage classification circuit 41 selects n ₁ (n ₀ > n ₁ ) specified by the large classification configuration determination unit 9 out of the n ₀ features of the input character pattern obtained from the feature extraction unit 3. And the n ₁ features of each category stored in the large classification dictionary 5 are calculated, and the categories are narrowed down to c _{1 in} ascending order of the distance. The second-stage classification circuit 42 outputs the n ₂ features (n ₀ > n ₂ > n ₁ ) designated by the large-class configuration determining unit 9 from the n ₀ features, and outputs them from the first-stage classification circuit 41. A distance is calculated between the n ₂ features of the c ₁ categories and the categories are c ₂ (c ₂ <c) in ascending order of the distance.
₁ ).

Similarly, in the k-th classifying circuit 43, n _k
The operation of obtaining _ck candidates using the above features is repeated.
In the process _{n 1 <n 2 <... <} n k <... <n K,
c _1> c _2> ... has _{a> c k>...> c K} , has a hierarchical major classification to reduce the number of candidates with increasing number of features in accordance undergo stages. C _K candidates are output from the K-th classifying circuit 44 at the last stage.

Assuming that the number of categories to be recognized is c ₀ and the time required to calculate the distance between one feature obtained from the input character pattern and one feature stored in the dictionary is Δt, Distance calculation time T when no major classification is performed
₀ is T ₀ = n ₀ · c ₀ · Δt (1). Now, n _k-number of features to be used in the k-th stage classification circuit 43 is assumed to include the front of the (k-1) stage classification circuit n _k-1 one feature. At this time, the distance calculation time T required for the hierarchical large classification is

[0019]

(Equation 1)

## EQU1 ## Second term is not the _{n k · c k-1 (} n k -n k-1) · c k-1 and turned by the n _k-1 or distance calculation feature of What is already completed in the preceding stage It depends. Distance speed ratio computation time η is expressed by the ratio of T ₀ to T, which is defined by the following equation.

Η = T ₀ / T (3) In order to see the effect of the processing amount reduction by the hierarchical large classification, as an example, K = 3, n ₀ = 500, c ₀ = 3,500, n
_{1 = 200, c 1 = 500} , n 2 = 300, c 2 = 300, n 3
= 500, η = 500 by the calculation of equations (1) to (3).
2.5, which means that the speed of distance calculation can be increased by 2.5 times.

Next, the operation of the large classification configuration determining unit 9 will be described with reference to FIG.
This will be described with reference to FIGS. FIG. 3 is a functional block diagram of the large classification configuration determining unit 9 showing one embodiment of the present invention.
It comprises an initial value setting circuit 91, a calculation amount calculation circuit 92, a large classification parameter setting circuit 93, and a classification rate table 94.

FIG. 4 is a processing flowchart of the large classification configuration determining section 9. FIG. 5 shows an embodiment of the classification rate table 94, in which n _i (i = 1, 2,..., I) candidates are selected from c ₀ to c _j (j = 1, 2,. .., J) is stored as the classification rate r (i, j) (0 ≦ r (i, j) ≦ 1). The r (i, j) is a value obtained in advance by a recognition experiment using a large amount of learning data. In the figure, I is the number of types of the number of features that can be taken in the classification of each layer, and J is the number of types of the number of candidates that can be taken, all of which are predetermined numbers.

The initial value setting circuit 91 determines the initial value of the number of features used in each stage and the number of candidates to be narrowed down from the value r ₀ of the large classification rate to be held provided by the trade-off function setting unit 10. First obtained by determining the K (step 21), the target value r _t of classification rate in each stage ^{_{r t = (r 0) 1}} / K (4) ( step 22). In the present embodiment, the load sharing of the classification rate of each stage is equal, but the load may be variable, such as reducing the weight of the front stage and increasing the weight toward the rear stage. The operation amount calculation circuit 92 constructs the classification rate table 94 by r (i, j).
Are selected from among K elements taking a value close to r _t ,
R _k ⁽¹⁾ (k = 1, k ^{) in} ascending order of j in (i, j)
2, ..., K) (step 23). FIG. 6 shows K = 3,
It is a figure which shows the example of selection at the time of I = 5, J = 5, and r (4,1), r (3,3), r (1,5) enclosed with the (circle) mark selected r ( i, j), and r ₁ ⁽¹⁾ = r (4,
1), r ₂ ^{(1 )} = r (3, 3), r ₃ ⁽¹⁾ = r (1,
5) is replaced. The large classification rate r _p of the time that has passed through the K stages

[0025]

(Equation 2)

The distance calculation time T _{p is obtained} by the equation (2) using the number of features and the number of candidates specified by r _k ⁽¹⁾ (step 24). Next, the values selected at random from among the four elements that are vertically and horizontally adjacent to each to r _k ⁽¹⁾ and r _k ⁽²⁾ in the classification rate table 94.

FIG. 6 shows that r (3, 3) is adjacent to the upper, lower, left and right sides.
Examples of the four elements are shown in squares. The r _k ⁽²⁾Using K
Classification rate r when passing_uIs calculated by equation (5), and
The r_k ⁽²⁾Using the number of features and the number of candidates specified in
Distance calculation time T_uIs obtained by Expression (2) (Step 2)
5).

The major classification parameter setting circuit 93 is set to the allowable value determined in advance [delta], the following three conditions _{r 0 -δ ≦ r u ≦ r} 0 + δ (6) r u> r p (7) T u <T p If satisfied (8) (step 26), the r _k corresponding features number and the number of candidates in ⁽²⁾ and temporarily registered as a large classification parameters, it substitutes the value of r _u in r _p (step 27) .

When the above processing is completed, the value selected by random numbers from the four elements adjacent to r _k ⁽²⁾
The processing from step 5 to step 27 is performed. The above process is repeated a predetermined number of times (L times) (step 28). By such a process, a large classification parameter that minimizes the distance calculation time T while having a large classification rate close to the value r ₀ of the large classification rate to be held is selected. After the L trials, the number of features and the number of candidates left in the large classification parameter setting circuit 93 are output from the large classification configuration determination unit 9 (step 29), and are provided to the processing of the large classification unit 4.

Next, the operation of the trade-off function setting unit 10
Will be described with reference to FIGS. FIG. 7 shows the processing speed / Oita
It is an example of a trade-off table of similarity, and is an M-stage.
Processing speed s_mLarge classification for (m = 1,2, ..., M)
Rate r_mConsists of s_m, R_mUses a lot of learning data
This is a value that is obtained in advance. Generally s₁<S _Two<
_...<S_M, R₁> R_Two>_...> R_MAnd processing speed
Of the trade-off that the large classification rate decreases with the speeding up
have.

FIG. 8 is a processing flowchart of the trade-off function setting section 10. In FIG. 8, the user inputs either the desired large classification rate r _d or the processing speed s _d (step 31).

When r _d is input, r closest to r _d among the M r _m stored in the trade-off table
Select _m (step 32) and sends the r _m in the major classification configuration determination section 9 (step 33). Major classification configuration decision unit 9
Is replaced by the value r ₀ of the major classification rate should retain the r _m, determines a large classification parameters by treatment of the large classification configuration determination section 9 (step 34).

[0033] s _{if d} is input, closest to s _d in the M stored in the tradeoff table s _m s
Select _m (step 35), r _m was removed (step 36) corresponding to the s _m, and sends the major classification configuration determination section 9 (step 33). Major classification configuration determination section 9 is replaced by the value r ₀ of the major classification rate should retain the r _m, determines a large classification parameters by treatment of the large classification configuration determination section 9 (step 34).

[0034] The user selects a r _d when the precision-oriented,
By selecting _sd when speed is important, it is possible to realize a large classification unit that operates at a large classification rate or processing speed according to the purpose. This is a trade-off between classification performance and processing speed.

In this way, a large classification unit for realizing a target large classification rate is automatically configured. This makes it possible to respond to various requests of the user in terms of recognition rate and processing speed.

In the present embodiment, rk ⁽²⁾ selected from the classification rate table 94 of the large classification configuration determining unit 9 is selected from four elements adjacent to rk ^{( 1) in} the upper, lower, left and right directions. Is not limited to the four elements adjacent to, but may be eight elements surrounding rk ⁽¹⁾ or the like.

In the present embodiment, the system for directly inputting the large classification rate or the processing speed in the trade-off function setting section 10 has been described. Even in this case, the same processing as in the present embodiment can be performed.

As described above, the present invention has been specifically described based on the embodiments. However, it is needless to say that the present invention is not limited to the above-described embodiments and can be variously modified without departing from the gist thereof.

[0039]

As described above, according to the present invention, since the hierarchical large classification means is employed and the function of automatically constructing the large classification part is provided, the specified recognition rate is maintained. There is an advantage that a system that can perform high-speed recognition while performing, or a system that achieves a specified processing speed can be easily constructed. Also,
Since the setting of the classification rate table and the trade-off table can be changed according to the quality of the input characters, there is an advantage that it can be applied to various uses.

[Brief description of the drawings]

FIG. 1 is a block diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is a block diagram of a large classification unit 4;

FIG. 3 is a block diagram illustrating a configuration of a large classification configuration determination unit 9;

FIG. 4 is a processing flowchart of a large classification configuration determining unit 9;

FIG. 5 is a diagram showing an example of a classification rate table 94.

FIG. 6 shows r (i, j) selected in the classification rate table 94
It is a figure showing an example of.

FIG. 7 is a diagram illustrating an example of a trade-off table.

FIG. 8 is a processing flowchart of the trade-off function setting unit 10;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Character pattern input part 2 Preprocessing part 3 Feature extraction part 4 Large classification part 5 Large classification dictionary 6 Identification part 7 Identification dictionary 8 Recognition result memory part 9 Large classification configuration decision part 10 Trade-off function setting part

Continued on the front page (72) Inventor Toru Wakahara 3-19-2 Nishi Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Corporation (72) Inventor Nobuo Miyamoto 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Nippon Telegraph and Telephone Within Telephone Co., Ltd. (72) Inventor Kenji Ogura 3-19-2 Nishi Shinjuku, Shinjuku-ku, Tokyo Japan Telegraph and Telephone Co., Ltd.

Claims (2)

[Claims] 1. A character recognition method for comparing a feature of an input character pattern with a feature of a standard dictionary to output a category to which the input character pattern belongs as a candidate, wherein: A classification circuit that reduces the number of candidates to be classified by calculating a classification scale between a part and all or a part of the features of the standard dictionary is configured to form a hierarchical structure, and a hierarchy is arranged from a previous stage to a subsequent stage. Hierarchical large classification method in which the number of candidates to be classified is reduced while increasing the number of features used for classification of the classification circuit as the processing proceeds to the hierarchy of The setting of the number of features and the number of candidates in each layer is performed by first setting the initial values of the number of features and the number of candidates in each layer, and then maintaining the large classification rate at a predetermined constant value. of Is performed by a method of sequentially changing the initial value so as to minimize the amount of computation required for the hierarchical large classification. The setting of the number of features and the number of candidates of each hierarchy in the hierarchical large classification method is performed by setting a large classification rate. When the target value is specified, the target value is replaced by the constant value, and the method of sequentially changing the target value is performed. When the target value of the large classification processing speed is specified, the target value is obtained in advance. The method is performed by applying the method of obtaining the constant value from the relationship between the large classification processing speed and the large classification rate, and sequentially changing the value, and performs a trade-off in which the two types of designation can be selected. Character recognition method to be characterized. 2. A character recognition apparatus for comparing a feature of an input character pattern with a feature of a standard dictionary to output a category to which the input character pattern belongs as a candidate, A classification circuit that reduces the number of candidates to be classified by calculating a classification scale between a part and all or a part of the features of the standard dictionary is configured to form a hierarchical structure, and a hierarchy is arranged from a previous stage to a subsequent stage. Hierarchical large classification means for decreasing the number of candidates to be classified while increasing the number of features used for the classification of the classification circuit as the processing proceeds to the hierarchy of The setting of the number of features and the number of candidates in the hierarchy is performed by first setting the initial values of the number of features and the number of candidates in each layer, and then maintaining the large classification rate at a predetermined constant value. so A large classification configuration determining means for sequentially changing the initial value so as to minimize the amount of computation required for the hierarchical large classification, and setting of the number of features and the number of candidates of each layer in the hierarchical large classification method, When the target value of the large classification rate is specified, the target value is replaced with the constant value and the value is sequentially changed. When the target value of the large classification processing speed is specified, the target value is obtained in advance. Trade-off function setting means for obtaining the constant value from the relationship between the large classification processing speed and the large classification rate and changing the constant value sequentially, and which can select the two types of designations. Character recognition device.