JPH05303663A - Character segmenting device - Google Patents

Abstract

PURPOSE:To provide a character segmenting device which is highly resistant to the pitch variance, the noises and the blurs of characters and can segment the characters at a high speed. CONSTITUTION:The function is minimized when the character segmenting position is optimum in a total character string with the parallel arithmetic carried out by a density smoothing network part 4 where the operator of the multi-input/single-output is connected to a network and a segmenting position estimating network part 5. Meanwhile the initial output value is decided by an initial value deciding part 7 so that a parameter control part 6 can process the processing parameter for each operator at a high speed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character slicing device for separating characters from a character string image in a document image character by character as a preprocessing of character recognition technology.

[0002]

2. Description of the Related Art As a conventional character slicing device, for example, "Proceedings of the 1990 Autumn National Conference of the Institute of Electronics, Information and Communication Engineers,
6-355, Masato Suda et al., Line / Character detection cut-out for automatic mail address classification machine.

FIG. 11 is a block diagram of this conventional character slicing device, in which 1 is a character string image input unit, 2 is a memory, 3 is a density histogram calculation unit, 111 is a character pitch estimation unit, and 112 is Gauss. A filter unit, a threshold processing unit 113, a section memory 114, a character start point detection unit 115,
Reference numeral 116 is a character end point detection unit, and 11 is a character pattern output unit.

The operation of the conventional character cutting device will be described with reference to the drawings. The memory 2 stores and holds the character string image read by the character string image input unit 1.

The density histogram calculation unit 3 counts the number of black pixels in the input character string direction and the vertical direction in the input character string image, and estimates the character pitch using the values obtained at each coordinate position in the character string direction as the density histogram. Part 111 and
It is output to the Gaussian filter unit 112. The character pitch estimation unit 111 sets the average value of the values in the input density histogram within a certain range with the maximum value as a reference, and outputs the average value to the character end point detection unit 116.

The Gaussian filter unit 112 suppresses the high frequency component of the density histogram by convolving the Gaussian function with respect to the inputted density histogram, and emphasizes the valley portion of the histogram. The output of the Gaussian filter unit 112 is input to the threshold value processing unit 113, and the start point and the end point of the section where the input value is a certain value or more are obtained and output to the section memory 114. The set of the start point and the end point obtained by the threshold processing unit 113 is sequentially stored and held in the section memory 114 as section information.

In the conventional apparatus, the character position is sequentially determined based on the section held in the section memory 114. Next, the determination of the character position will be described.

First, the character start point detection unit 115 reads out the one closest to the coordinate origin among the sections held in the section memory 114, and uses that start point as the start point of the first character and the character end point detection unit 116. It is output to the character pattern output unit 9. Letting Xstart be the input point just entered and L the value of the character pitch obtained by the character pitch estimation unit 4,
Next, the character end point detection unit 116 determines the end point of the section that satisfies (Equation 1) among the sections held in the section memory 114.
Xend is output to the character pattern output unit 11 and the character start point detection unit 115 as the end point of the first character.

[0009]

[Equation 1]

The detection by the character end point detection unit 116 is performed by changing the value of α until the end point can be detected.

When the end point of the first character is input from the character end point detecting section 116, the character starting point detecting section 115 reads out the starting point which has a value larger than the ending point and is closest to the coordinate origin point. Similar to the case of a character, the start point is output to the character end point detection unit 116 and the character pattern output unit 11 as the start point of the second character. In the same manner, the end point of the second character, the start point of the third character, and the end point of the third character are sequentially serialized until all the sections stored in the section memory 114 correspond to the character positions. The character position is determined, and the information on the character position is output to the character pattern output unit 11.

The character pattern output unit 11 reads an image in the range from the start point to the end point of the input character from the input character string image held in the memory 2 and sequentially outputs it as a character pattern.

[0013]

However, in the above-mentioned configuration, since it is necessary to sequentially proceed with the processing, the processing time is correspondingly increased. Further, there is a problem in that the cutout result cannot be evaluated as a whole, and once a cutout error occurs, the processing accuracy is deteriorated because it greatly affects the cutout results thereafter.

In view of the above point, the present invention has an object to provide a character slicing device having a short processing time due to parallel processing and a high processing accuracy as a whole.

[0015]

By means of a character string image input section for inputting a character string image, a memory for storing and holding an input character string, and counting the number of black pixels in the input character string direction and in the vertical direction. A parameter that determines the optimum parameter value by matching the density histogram calculation unit that calculates the obtained density histogram and the input from the density histogram calculation unit and the parameter dictionary that uses the density histogram distribution as an index. A control unit, which receives inputs from the density histogram calculation unit and the parameter control unit, and is configured by a network in which multi-input one-output operators corresponding to pixel positions in the input character string direction are mutually coupled, By minimizing the function that takes the minimum value when the values of the density histogram of the character part are almost the same, A density smoothing network unit that smoothes the density histogram and reduces the influence of noise in the input character string, and a pixel position in the input character string direction that receives inputs from the density smoothing network unit and the parameter control unit. It consists of a network in which the corresponding multi-input and single-output operators are connected to each other, and the function that takes the minimum value when the cutting position in the entire character string is optimum minimizes the character cutting position. A cutout position estimation network unit to estimate, a cutout position determination unit that determines a character position based on an input value from the cutout position estimation network unit, and the cutout position determination unit to output a correct output. The user refers to the output results of the learning unit that rewrites the contents of the parameter dictionary in the parameter control unit and the cutout position determination unit, and it is incorrect. Case comprises correct positive clipping position presenting unit which outputs input to the learning section, a character pattern output unit for character pattern output of each character from the memory based on input from the cutting position determination unit.

[0016]

According to the present invention, with the above-described configuration, the evaluation function that takes the minimum value when the optimum cutting result is obtained as a whole is operated in parallel by the network in which the operators having a large number of inputs and a single output are connected in parallel. The processing speed and processing accuracy can be improved. In addition, the learning unit that rewrites the parameter dictionary in the parameter control unit by obtaining the processing parameter value that matches the individual habits and fluctuations of the user,
Further, the processing accuracy for each user can be improved.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram of a character slicing device according to an embodiment of the present invention. In FIG. 1, 1 is a character string image input unit, 2 is a memory, 3 is a density histogram calculation unit, 4 is a density smoothing network unit, 5 is a cutout position estimation network unit, 6 is a parameter control unit, and 7 is an initial value. A determination unit, 8 is a cutout position determination unit, 9 is a learning unit, 10 is a normal cutout position presentation unit, and 11 is a character pattern output unit.

The operation of the character slicing device of the present embodiment constructed as described above will be described below.

The memory 2 stores and holds the character string image read by the character string image input unit 1.

The density histogram calculator 3 counts the number of black pixels in the input character string image in the input character string direction and in the vertical direction, and smoothes the density obtained by using the values obtained at each coordinate position in the character string direction as a density histogram. Network part 4,
It is output to the parameter control unit 6 and the initial value determination unit 7.

FIG. 2 is a specific configuration diagram of the density smoothing network unit 4 and the cut-out position estimation network unit 5, and 2
Reference numeral 1 is a density value calculation unit, and 22 is a cutout position calculation unit. In FIG. 2, the density value calculation unit 2 is shown for simplicity.
1, the output state of only one each of the cutout position calculation unit 22 is shown.

As shown in FIG. 2, the concentration value calculation unit 21
The cutout position calculation unit 22 is associated with the coordinates in the character string direction, and the output of the density histogram calculation unit 3 becomes an initial input to the corresponding density value calculation unit 21.

The output of each density value calculation unit 21 and the cut-out position calculation unit 22 is output by another density value calculation unit 21,
It serves as an input to the cutout position calculation unit 22 and a feedback input to itself. The cutout position calculation unit 2
The output of 2 is input to the cutout position determination unit 8.

In order to explain the operations of the density value calculation unit 21 and the cutout position calculation unit 22, first, the principle of the character cutout process in the present invention will be explained.

In general, the following items can be assumed as knowledge about the character cutting position based on the density histogram. (1) There is a high possibility that the value of the density histogram is small and the cutout position is located in a portion where the histogram distribution is convex downward. (2) The cutout positions are distributed with a substantially constant width. (3) The number of characters to be cut out is limited by the height of the character string.

A function that takes the minimum value when these assumptions are satisfied is expressed as (Equation 2), (Equation 3), (Equation 4) for each assumption.

For the sake of explanation, the output value of the i-th density value calculation unit 21 on the coordinate in the character string direction is F
(i) Similarly, let the output value of the i-th cut-out position calculation unit 22 be L (i). However, it is assumed that the value of L (i) is closer to 1 as the possibility of a character cutout position is higher, and closer to 0 as it is lower.

[0028]

[Equation 2]

[0029]

[Equation 3]

[0030]

[Equation 4]

The meaning of the parameters in each function will be described below. A in (Equation 2) represents the ratio of how small the value of F (i) is and how close the value of L (i) is to 1.

Na in (Equation 2) represents a range to be considered when calculating the degree to which the density histogram value centered on the i-th is small.

Nf in (Equation 2) represents a range to be considered when calculating the ratio in which the distribution of the density histogram value centered on the i-th is convex downward.

The function C of (Equation 3) is a monotonically decreasing function having a positive value. This means that the possibility that another cutout position will be closer to a certain cutout position is smaller as the position is closer, and the value taken by the function C represents the ratio.

Ni in (Equation 3) represents a range in which the assumption represented by the function C is to be considered, with the i-th centered.

Nn in (Equation 4) represents a limited number of cutout positions. The values of these parameters are set by the parameter control unit 6
Is determined based on the output of the density histogram calculation unit 3, which will be described later in the operation of the parameter control unit 6.

The character cutting position can be determined by minimizing the function represented by the linear sum of the above functions, but in the present invention, the value of L (i) is in the range of 0 to 1. Furthermore, the functions expressed by (Equation 5) and (Equation 6) are added as the restriction of.

[0038]

[Equation 5]

[0039]

[Equation 6]

However, the function S of (Equation 6) satisfies (Equation 7), and m (i) is input to the i-th cutout position calculation unit 22 from the density value calculation unit 21 and the cutout position calculation unit 22. Is the sum of the input values. Further, λ is assumed to be set to an appropriate value in advance.

[0041]

[Equation 7]

Further, in order to eliminate the influence of the blur of the character string image and the noise of isolated points, it is necessary to proceed with the clipping process while smoothing the density histogram. If this is expressed by a function similar to the above, it can be written as (Equation 8).

[0043]

[Equation 8]

Next, the principle of minimizing these functions will be described. For the purpose of explanation, let E (x) be a differentiable function with respect to a variable x.
In the case of 0, (Equation 9) is established.

[0045]

[Equation 9]

Assuming (Equation 10), (Equation 11) is established. However, ε is a sufficiently small positive number.

[0047]

[Equation 10]

[0048]

[Equation 11]

As is clear from (Equation 11) (Equation 1)
By changing the variable x according to 0), x that minimizes E (x) can be obtained. Further, even if the variable x is a vector, the same thing can be said, and it is obvious that it may be changed completely independently for each vector dimension.

Based on such a principle, the density value calculating unit 21 independently minimizes the function shown in (Equation 12) as smoothing of the density histogram at each coordinate point in the character string direction. Similarly, the cutout position calculation unit 22 independently minimizes the function shown in (Equation 13) as the estimation of the character cutout position at each coordinate point in the character string direction. However, the coefficients in (Equation 12) and (Equation 13) represent the ratios in which each function is considered, and are assumed to be set to appropriate values in advance.

[0051]

[Equation 12]

[0052]

[Equation 13]

By minimizing these functions (Equation 2),
As is clear from (Equation 3) and (Equation 4), the output of the cutout position calculation unit 22 is determined by the interaction between the density value calculation unit 21 and the outputs of other cutout position calculation units 22. Therefore, the optimum cutout position can be estimated for the entire character string, and the character cutout can be achieved by parallel processing that is resistant to pitch variation, noise, and blurring of characters.

Next, specific operations of the density value calculation unit 21 and the cutout position calculation unit 22 will be described with reference to the drawings.

FIG. 3 is a concrete configuration diagram of the density value calculation unit 21, in which 31 is a density partial differentiation unit, 32 is a density multiplication unit, 33 is an addition unit, 34 is an output value memory, and 35 is an output unit. ..

For the sake of explanation, i on the coordinate in the character string direction
It is assumed that it is the th density value calculation unit 21.

First, at the beginning of the process, the contents of the output value memory are initialized to the input values from the density histogram calculating section 3.
After that, the stored contents are rewritten according to the output value of the adding unit 33 for each time.

The density partial differentiation unit 31 includes a parameter control unit 6, a density calculation unit 21, and a cutout position calculation unit 22.
DF (i) expressed by (Equation 14) based on the input value from
Is output to the density multiplication unit 32.

[0059]

[Equation 14]

The density multiplication unit 32 multiplies the input value from the density partial differentiation unit 31 by a sufficiently small positive number εf, and outputs it to the addition unit 33.

The adding unit 33 includes a density multiplying unit 32 and an output unit 3.
5, the sum of each input value is calculated, and the output unit 3
After sending the trigger signal to 5, when the trigger signal from the output unit 35 is received, the contents of the output value memory 34 are rewritten.

When the output section 35 receives the trigger signal from the addition section 33, the output section 35 reads out the content stored and held in the output value memory 34, outputs it as the concentration calculation unit 21, and sends the trigger signal to the addition section 33. ..

FIG. 4 is a concrete configuration diagram of the cutout position calculating unit 22. 41 is a position partial differentiation unit, 42 is a position multiplication unit, 43 is a non-linear function unit, 33 is an addition unit, and 34 is an output value memory. , 35 are output units.

For the sake of explanation, i on the coordinate in the character string direction
It is assumed that the cutout position calculation unit 22 is the th.

First, at the beginning of the process, the contents of the output value memory 34 are initialized by the initial value determination unit 7 based on the output of the density histogram calculation unit. This will be described later in the operation of the initial value determination unit 7. explain. After that, the stored contents are rewritten according to the output value of the adding unit 33 for each time.

The position partial differentiation unit 41 includes a parameter control unit 6, a density calculation unit 21, and a cutout position calculation unit 22.
Dm (i) expressed by (Equation 15) based on the input value from
Is output to the position multiplication unit 42.

[0067]

[Equation 15]

The density multiplication unit 32 multiplies the input value from the density partial differentiation unit 31 by a sufficiently small positive number εl, and outputs it to the addition unit 33.

The adder 33 includes a position multiplier 32 and an output unit 3.
5, the sum of each input value is calculated, and the output unit 3
After sending the trigger signal to 5, when the trigger signal from the output unit 35 is received, the contents of the output value memory 34 are rewritten.

Upon receiving the trigger signal from the adder 33, the output unit 35 reads the contents stored and held in the output value memory 34, outputs the contents to the non-linear function unit 43, and sends the trigger signal to the adder 33.

The non-linear function section 43 transforms the input value by the non-linear function S represented by (Equation 7) and outputs it as the output of the cutout position calculating unit 22.

By repeating the above processing, the estimation of the cut-out position by minimizing the function described above is achieved.

Next, the operation of the parameter controller 6 will be described with reference to the drawings. FIG. 5 is a specific configuration diagram of the parameter control unit 6, 51 is a parameter standard value calculation unit, 52 is a parameter dictionary, 53 is a matching unit, and 54 is a parameter value output unit.

The parameters controlled by the parameter control unit 6 are (Equation 2), (Equation 3), (Equation 4), (Equation 12), (Equation 1).
3) A, Na, Nf, Ni, C (x), Nn, Cf, Ci, Cn,
Cv and Cg. The parameter standard value calculation unit 51 first obtains the standard value of the value of each parameter as follows.

FIG. 6 is a concrete configuration diagram of the parameter standard value calculating unit 51. 61 is a maximum value detecting unit, 62 is a character height estimating unit, 63 is a squareness estimating unit, and 64 is a standard value calculating unit. is there.

The maximum value detection unit 61 receives the input from the density histogram calculation unit 3, detects the maximum value of the input value for each constant section, and outputs it to the standard value calculation unit 64. The character height estimation unit 62 uses the input value from the density histogram calculation unit 3 for each constant section in the character string direction as an average value of values within a certain range based on the maximum value as a character height in that section. It is output to the standard value calculation unit 64.

From the character height estimating unit 62, the squareness estimating unit 63 obtains the average value of the length of the portion where the input value from the density histogram calculating unit 3 is larger than the constant value for each constant section in the character string direction. The value divided by the input value is output to the standard value calculation unit 64 as the squareness of the character.

It should be noted that instead of "the average value of the length of the portion where the input value from the density histogram calculation unit 3 is larger than a fixed value", "the input value from the density histogram calculation unit 3 is larger than the fixed value" The length at which the length histogram is the largest "may be used.

For the purpose of explanation, the input value from the maximum value detecting unit 61 is M, the input value from the character height estimating unit 62 is T, the input value from the squareness estimating unit 63 is H, and the character string direction is constant. Let the length of the section be D.

The standard value calculation unit 64 determines the values of the parameters A, Nf, Ni, Nn and the function C for each constant section in the character string direction as follows.

The parameter A is determined according to (Equation 16) with M being the maximum value of F (i) as a criterion so that the processing is not affected by the range of values taken by F (i). However, β is a positive number and is set to an appropriate value in advance.

[0082]

[Equation 16]

It is sufficient for the value of Na to have a range of one character at the maximum, and the estimated character height T is used as a criterion (Equation 17).
Determined according to. However, γ1 is a positive number less than or equal to 1 and is set to an appropriate value in advance.

[0084]

[Equation 17]

The same applies to the value of Nf, as long as there is a range for one character at the maximum, it is determined according to (Equation 18) using the estimated height T of the character as a criterion. However, γ2 is a positive number less than or equal to 1 and is set to an appropriate value in advance.

[0086]

[Equation 18]

The value of Ni needs to be set to a small value if the squareness of the characters included in the section is small. Based on the estimated squareness H of the character, the character height T is used as a criterion (Equation 19).
Determined according to. However, the function I is a monotonically increasing function with a range of 0 to 1.

[0088]

[Formula 19]

The value of Nn is a value determined by the number of characters included in the section, and is determined according to (Equation 20) with the value obtained by dividing the length of the section by the character width as a criterion. However, δ is a positive number and is set to an appropriate value in advance.

[0090]

[Equation 20]

The value of the function C (x) is 0 when x = Ni and has the maximum value when x = 1, and from the meaning thereof, the value of A is determined according to (Equation 21). However, x ≠ 0, θ
Is a positive number and is set to an appropriate value in advance.

[0092]

[Equation 21]

The standard value calculation unit 64 outputs the standard value of each parameter obtained as described above to the parameter dictionary 52.

FIG. 7 is a specific configuration diagram of the parameter dictionary 52. Reference numeral 71 is a parameter standard value memory, 72 is a density histogram distribution memory, and 73 is a parameter value memory. Standard value calculation unit 6 in parameter standard value calculation unit 51
The input from 4 is stored and held in the parameter standard value memory 71.

1 for each density histogram distribution memory 72
One parameter value memory corresponds.

The density histogram distribution memory 72 stores and holds the density histogram distribution input from the learning section 9. The parameter value memory stores and holds the value of each parameter determined by the learning unit 9 in the corresponding density histogram distribution memory 72.

The contents of the density histogram distribution memory 72 and the parameter value memory 73 are determined by the output of the learning unit 9 as just described, but the specific contents will be described later in the operation of the learning unit 9. I will tell you when.

The matching unit 53 performs matching between the input from the density histogram calculation unit 3 and the density histogram distribution stored and held in each density histogram distribution memory 72 for each section of a predetermined fixed length according to (Equation 22). Then, the density histogram distribution memory 72 having the maximum similarity R of (Equation 22) is obtained, and the value of the similarity R and the address of the density histogram distribution memory 72 are output to the parameter value output unit 54. However, the input from the density histogram calculation unit 3 is {x1, x2, ..., Xn}, and the content of the density histogram distribution memory 72 is {v1, v2, ..., Vn}.

[0099]

[Equation 22]

When the similarity R input from the matching unit 53 satisfies (Equation 23), the parameter value output unit 54 refers to the address of the density histogram distribution memory 72, and stores the corresponding parameter value memory 73. The content is output to the density smoothing network unit 4 and the cutout position estimation network unit 5 as an output of the parameter control unit 6.
If (Equation 23) is not satisfied, the parameter dictionary 52
The content of the parameter standard value memory 71 therein is output from the parameter control unit 6. However, μ is set to an appropriate value in advance.

[0101]

[Equation 23]

As described above, the parameter control unit 6
Determines the parameter value for each section and outputs it to the corresponding density calculation unit 21 in the density smoothing network unit 4 and the corresponding cutout position calculation unit 22 in the cutout position estimation network unit 5.

Next, the cut-out position determining unit 8 determines a certain fixed value among the output values of the cut-out position calculating unit 22 in the cut-out position estimating network unit 5 corresponding to each coordinate in the character string direction after a certain time. Cutting position calculation unit 22 having a large value
The coordinate corresponding to is set as the character cutout position and is output to the normal cutout position presentation unit 10 and the character pattern output unit 11. It should be noted that this constant value is set to a value close to 1 in advance.

Next, specific operations of the learning unit 9 and the normal cutout position presenting unit 10 will be described with reference to the drawings. FIG. 8 is a configuration diagram of the learning unit 9, where 81 is a parameter changing unit, 82 is an activity calculating unit, and 83 is a parameter correcting unit.

The parameter changing unit 81 receives an input from the normal cut-out position presenting unit 10 or the parameter correcting unit 83, and receives the parameter standard value memory 7 in the parameter control unit 6.
A, Na, Nf, Ni, C (x), Nn, C stored in 1
The standard values of f, Ci, Cn, Cv, and Cg are read in order, and the value of each parameter is increased or decreased by a value that is predetermined for each parameter. Then, the new set of parameter values is output to the activity calculation section 82.

The activity calculating section 82 receives the input from the density histogram calculating section 3, calculates the value of dm (i) of (Equation 15) in the section input from the normal extraction position presenting section 10, and corrects the parameters. It is output to the unit 83. The section input from the normal cutout position presentation unit 10 is a cutout error section specified from the outside, and this will be described later in the description of the normal cutout position presentation unit 10.

The parameter correction unit 83 has the activity calculation unit 8
When the value at the coordinate position input from the normal-cut-out position presentation unit 10 is larger than the value at the other coordinate position by referring to the value at each coordinate position input from 2, the density histogram distribution in the input section Is output to the density histogram memory 72 in the parameter control unit 6, and the set of parameter values at this time is output to the corresponding parameter control unit 6
It is output to the parameter value memory 73 therein. If the value at the coordinate position input from the normal cut-out position presenting unit 10 is smaller than the values at other coordinate positions, a signal is sent to the parameter changing unit 81, and the parameter changing unit 81 outputs a new set of parameter values. Then, the process is repeated again to search for a larger set of parameter values. The coordinate position input from the normal cutout position presentation unit 10 is a correct cutout position instructed from the outside, which will be described later in the description of the normal cutout position presentation unit 10.

FIG. 9 is a block diagram of the normal-cut-out position presenting section 10, in which 91 is an image display section, 92 is a normal-cut-out position instruction section,
Reference numeral 93 is a normal cutting position output unit.

The image display unit 91 reads the input character string image from the memory 2 and presents it to the user as an image, and also receives the input from the cutout position determining unit 8 and presents the cutout result on the image in the same manner. ..

Next, the user who sees the result presented on the image display section 91, if the result is wrong,
The section position of the erroneous portion and the correct cut-out position are designated by the normal cut-out position instruction unit 92, and output to the normal cut-out position output unit 93. The normal cut-out position output unit 93 receives the input from the normal cut-out position instruction unit 92, the section position of the error portion to the activity calculation unit 82 in the learning unit 9, and the correct cut-out position coordinate position in the learning unit 9. It is output to the parameter value correction unit 83. Although it is the user who individually gives the normal cut-out position, character recognition processing and further language processing are performed, the processing parameters are changed for the grammatically incorrect part, and the cut-out processing is performed again, and the grammatically correct part is given. The cutout position may be the normal cutout position.

Next, the operation of the initial value determination unit 7 will be described with reference to the drawings. The initial value determination unit 7 includes the density histogram calculation unit 3
According to the input value from, the cutout position, the coordinate that is unlikely to be the cutout position, and other coordinates are surely determined, and the cutout position calculation unit 22 at each coordinate is determined.
The contents of the output value memory 34 therein are initialized in each case.

FIG. 10 is a concrete configuration diagram of the initial value determination unit 7, 101 is a character string edge detection unit, 102 is a high density coordinate detection unit, 103 is an initial value calculation unit, and 61 is a maximum value detection unit. is there.

The character string end detection unit 101 has a value greater than or equal to a certain value among the respective inputs corresponding to the coordinates in the character string direction from the density histogram calculation unit 3, the smallest coordinate is the character string start point, and the largest coordinate is the character. The information is sent to the initial value calculation unit 103 as the end point of the column.

The maximum value detecting section 61 receives the input from the density histogram calculating section 3, detects the maximum value of the input value for each constant section, and outputs it to the high density coordinate detecting section 102.

The high density coordinate detecting section 101 outputs, to the initial value calculating section 103, coordinates having a value larger than the value determined by the input value from the maximum value detecting section 61 among the inputs from the density histogram calculating section 3.

The initial value calculation unit 103 includes a character string end detection unit 1
Cutout position calculation unit 22 corresponding to the input value from 01
The content of the output value memory 34 is L1, the high-density coordinate detection unit 1
Cutout position calculation unit 22 corresponding to the input value from 02
The contents of the output value memory 34 therein are initially set to L2, and the contents of the other output value memory 34 are initially set to L3. However, L1, L2, and L3 satisfy (Equation 24).

[0117]

[Equation 24]

In this way, the smallest value (value close to 0) is given to the coordinate position which is obviously not the cutout position, and the large value (value close to 1 is originally given to the coordinate position where the possibility of the cutout position is originally high. ) As an initial value,
The correct clipping position can be estimated at high speed and accurately as compared with the case where the same value is uniformly used as the initial value.

Finally, the character pattern output unit 9 reads out the image between the adjacent cutout positions from the input character string image held in the memory 2 and sequentially outputs it as a character pattern.

As described above, according to this embodiment, the density smoothing network unit 4 smoothes the projection histogram distribution, and the cutout position estimation network unit 5 refers to the result and parallelizes the character cutout positions. The optimum cut-out position for the entire character string can be obtained by estimating the. As a result, character segmentation processing that is resistant to pitch variation, noise, and blurring of characters is achieved, and high-speed processing can be performed by parallel processing.

Further, the normal cut-out position presentation unit 10 feeds back information on the correct cut-out position to the learning unit 9.
Then, the learning unit 9 finds the fluctuation of the processing parameter of each user, and optimizes the processing parameter accordingly. Next, the parameter control unit 6 stores and holds the processing parameter optimized for each user, and determines the optimum processing parameter for the character string currently processed based on the density histogram distribution. As a result, the processing parameters for each user and each document can be optimized, and the processing accuracy is further improved.

Further, by providing the initial value determination unit 7, the processing speed and accuracy can be further improved by reflecting the approximate position of the correct clipping position.

[0123]

According to the present invention, the cutout position is estimated by the interaction of the density smoothing network unit and the cutout position estimation network unit by parallel processing, and the parameter of the interaction is set by the parameter control unit. Also, by setting the initial value for the interaction in advance by the initial value determination unit, the character can be cut out at a high speed, which is resistant to pitch variation, noise, and blurring of the character. Further, it is possible to deal with the fluctuation of individual users, which cannot be dealt with by uniformly setting the processing parameters, by optimizing the processing parameters by learning and having them as a dictionary, which has a great practical effect.

[Brief description of drawings]

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

FIG. 2 is a configuration diagram of a density smoothing network unit 4 and a cutout position estimation network unit 5 according to the present invention.

FIG. 3 is a configuration diagram of a density value calculation unit 21 according to the present invention.

FIG. 4 is a configuration diagram of a cutout position calculation unit 22 according to the present invention.

FIG. 5 is a configuration diagram of a parameter control unit 6 in the present invention.

FIG. 6 is a configuration diagram of a parameter standard value calculation unit 51 according to the present invention.

FIG. 7 is a configuration diagram of a parameter dictionary 52 according to the present invention.

FIG. 8 is a configuration diagram of a learning unit 9 in the present invention.

FIG. 9 is a configuration diagram of a normal cutout position presentation unit 10 according to the present invention.

FIG. 10 is a configuration diagram of an initial value determination unit 7 in the present invention.

FIG. 11 is a configuration diagram of a character cutting device in a conventional example.

[Explanation of symbols]

 1 character string image input unit 2 memory 3 density histogram calculation unit 4 density smoothing network unit 5 cutout position estimation network unit 6 parameter control unit 7 initial value determination unit 8 cutout position determination unit 9 learning unit 10 normal cutout position presentation unit 11 Character pattern output unit 21 Density value calculation unit 22 Cutting position estimation unit 31 Density partial differentiation unit 32 Density multiplication unit 33 Addition unit 34 Output value memory 35 Output unit 41 Position partial differentiation unit 42 Position multiplication unit 43 Non-linear function unit 51 Parameter Standard value calculation unit 52 Parameter dictionary 53 Matching unit 54 Parameter value output unit 61 Maximum value detection unit 62 Character height estimation unit 63 Squareness estimation unit 64 Parameter calculation unit 71 Parameter standard value memory 72 Density histogram distribution memory 73 Parameter value memory 81 Parameter change part 82 Activity calculation Output section 83 Parameter correction section 91 Image display section 92 Normal cut-out position instruction section 93 Normal cut-out position output section 101 Character string end detection section 102 High density coordinate detection section 103 Initial value calculation section 111 Character pitch estimation section 112 Gaussian filter section 113 Threshold Processing unit 114 Section memory 115 Character start point detection unit 116 Character end point detection unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Taiji Yuki, 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (2)

[Claims] 1. A character string image input section for inputting a character string image, a memory for storing and holding an input character string, and a density histogram obtained by counting the number of black pixels in the input character string direction and in the vertical direction. A density histogram calculation unit for calculating, and a parameter control unit that receives an input from the density histogram calculation unit and performs matching with a parameter dictionary having a density histogram distribution as an index, and a parameter control unit that determines an optimum parameter value, It is composed of a network that receives inputs from the density histogram calculation unit and the parameter control unit and mutually connects multi-input one-output operators corresponding to pixel positions in the input character string direction, The density histogram is minimized by minimizing the function that takes the minimum value when the values in the density histogram are almost the same. The smooth, and the density smoothing network unit to reduce the effect of noise present in the input string,
An input of the density smoothing network unit and the parameter control unit, and a network in which multi-input and one-output operators corresponding to pixel positions in the input character string direction are mutually connected, and the entire character string is formed. A clipping position estimation network unit that estimates the clipping position of a character by minimizing the function that takes the minimum value when the clipping position is optimal, and a character based on the input value from the clipping position estimation network unit. A cutout position determination unit that determines the position, a learning unit that rewrites the contents of the parameter dictionary in the parameter control unit so that the cutout position determination unit outputs a correct output, and an output result of the cutout position determination unit. Refer to the normal cut-out position presenting unit for inputting a correct output to the learning unit if it is incorrect, and the memory based on the input from the cut-out position determining unit. Character segmentation apparatus comprising the character pattern output unit for character pattern output of each Luo character. 2. The character cutout according to claim 1, further comprising an initial value determination unit that determines an output initial value of an operator in the cutout position estimation network unit based on an input from the density histogram calculation unit. apparatus.