JPH04304586A - Device for recognizing character - Google Patents

Abstract

PURPOSE:To obtain a character recognizing device which controls the load dispersal of a character recognition processing to make all recognition time minimum when plural character recognizing means exist. CONSTITUTION:Information of the number of a character in a character kind and recognition object such as a KANJI (Chinese character), an alphabet, etc., from a format information storing means 3 and information of the number of the character and an average character size from a character segmenting means 4 are transmitted to a neural network means 8 and the recognizing time of an applying field is predicted so that a control means 7a executes the load dispersal of the field recognition processing through the use of the predicted value so as to make all recognition processing time minimum. The input parameter of the neural network means 8 at every reading of a respective slip and recognition processing time being actual read result are stored as learning data and the learning of the neural network is executed through the use of the data. The optimum load dispersal comes to be executed at every slip and a through-put is especially improved at the time of successively reading the same slip.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device, and more particularly to a device for recognizing a form using a plurality of microprocessors based on format information.

0002

2. Description of the Related Art FIG. 10 is a block diagram showing the structure of a conventional character recognition device. In the figure, reference numeral 1 denotes a recording medium such as a form or paper on which a character pattern to be read is written or printed, and hereinafter will be described as a form. 2
3 is a scanning means such as an image scanner that scans the form 1 and converts it photoelectrically; 3 is a format information storage means comprising a ROM or the like in which format information such as the position of the reading target area on the form and the type of characters to be recognized is stored in advance; 4; 5 and 6 are character cutting means for inputting the character string image of each field from the scanning means 2 and cutting out each character by using the positional information of the field to be read obtained from the format information storage means 3; A first character recognition means and a second character recognition means which perform recognition by comparing a reference pattern and an input character pattern sent from the character cutting means 4 and output a recognition result, 7 is the scanning means 2, and a format. It is a control means for controlling the information storage means 3, the character cutting means 4, and the character recognition means 5 and 6. In addition, the above-mentioned character cutting first stage 4, character recognition means 5
, 6 and the control means 7 are realized by a microprocessor or the like.

Next, the operation will be explained. In the form 1 shown in FIG. 11, 9 is field 1 and the number "1",
"2", "3", and "4" are entered, and 10 is field 2, where the kanji "east" and "kyo" are entered.
11 is field 3 and contains the numbers "7", "8", "9"
is entered, and 12 is field 4, in which the kanji characters "bun", "ji", "ken", and "shiki" are entered. Figure 11
The character pattern of the form 1 is scanned by the scanning means 2 and photoelectrically converted. The control means 7 receives the position information on the form of the field to be cut out stored in the format information storage means 3 and sends it to the character cutting means 4. Character cutting means 4
inputs the photoelectrically converted character string pattern of each field using the positional information of fields 1, 2, 3, and 4 to be read on the form sent from the control means 7;
Performs the process of cutting out character patterns for each character.

Next, the control means 7 controls the character cutting means 4.
Of the character patterns cut out in step 1, the character pattern in field 1 is sent to the first character recognition means 5, and the character pattern in field 2 is sent to the second character recognition means 6 for recognition. The first character recognition means 5 and the second character recognition means 6 compare the sequentially sent character patterns with the reference pattern to be recognized, and output the character with the smallest difference as a recognition result.

FIG. 12 shows load distribution indicating whether each field is processed by the first or second character recognition means, the time for recognizing each field, and the throughput of the recognition time. Since the recognition times for fields 1 and 2 are as shown in FIG. 12, the first character recognition means finishes processing field 1 at time 7 [s]. The control means 7 instructs the first character recognition means 5 to perform recognition processing on the field 3 for which recognition processing has not yet been completed. Similarly, the control means 7 causes the second character recognition means 6 at time 9 [s] to
Instructs field 4 recognition processing.

Similarly, a form 1 shown in FIG. 13 having a format similar to that in FIG. 11 is read, and character recognition control as shown in FIG. 14 is performed. The above process is then repeated for other forms.

[0007]

The conventional character recognition device is constructed as described above, and as shown in the above operation example, the control means 7 controls the first character recognition means after the character recognition process has been completed. 5 and the second character recognition means 6, fields for which character recognition processing has not yet been completed are simply assigned, so the recognition processing time when processing the form shown in FIG. 11 is as shown in FIG. 12. As shown, it is 32 seconds. Then, after the first character recognition means finished recognizing field 3, there was no field to process and the first character recognition means went into a waiting state for 17 seconds, which caused the problem that the two microprocessors could not operate efficiently. .

[0008] Furthermore, when the units assigned to the first character recognition means 5 and the second character recognition means 6 are assigned not to field units but to character units, the first character recognition means 5 and the second character recognition means The character recognition means 6 is used almost equally, but each time the recognition process for each character is completed, an interrupt is generated from the character recognition means 5 and 6 to the control means 7, and each time the interrupt occurs, it is determined which field will be next. Since it is necessary to control which characters are assigned, the overhead of the control means 7 increases, resulting in a problem that the overall throughput decreases.

[0009] Furthermore, when reading forms of the same format repeatedly as in this conventional example, the processing time for each field is almost the same between sheets, so even when reading the second and subsequent sheets, Similar to the first version, this method had the problem of inefficient load distribution and reduced throughput.

[0010] This invention was made in order to solve the above-mentioned problems, and is a character recognition system that can control load distribution when there are multiple character recognition means so as to maximize the overall throughput. The purpose is to obtain a recognition device.

[0011]

[Means for Solving the Problems] A character recognition device according to the present invention includes a prediction means for predicting the processing time of each recognition area;
Based on the predicted recognition time obtained from the above prediction means, the system determines the assignment of each field to a plurality of character recognition means so as to maximize throughput, and is equipped with a control means for controlling each character recognition means. be. Further, a neural network is used as a prediction means, inputting information from a format information storage means and information from a character cutting means, and outputting a predicted recognition time. Furthermore, using a neural network as a prediction means, the recognition time of the character recognition means is input, and the predicted recognition time is output.

0012

[Operation] The prediction means in this invention calculates the predicted recognition time of each recognition area, and the control means determines the allocation so as to optimize the load distribution for each character recognition means based on the predicted recognition time. Take control.

[0013] Furthermore, the neural network serving as the prediction means uses the information in the format information storage means and the information in the character cutting means as learning data to predict the recognition time. Furthermore, the neural network as a prediction means uses the recognition time of the character recognition means as learning data to predict the recognition time of subsequent forms.

[0014]

[Example] Example 1. An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a character recognition device employing an embodiment, and the same or corresponding parts as in the conventional example shown in FIG. In the figure, 8 inputs information on the number of characters in the field cut out by the character cutting means 4 and information on the field stored in the format storage means 3 via the control means 7a each time a form is read, and It is a neural network means that outputs the predicted recognition time and also accumulates the format information of each field and the actual recognition time of that field each time a form is read.

Further, the control means 7a of the present invention has the following functions added to the functions of the conventional control means 7. That is,
A function of inputting the number of characters and the format information of each field stored in the format information storage means 3 to the neural network means 8 when the number of characters of each field in the form is determined, the neural network means 8
The first character recognition means 5 recognizes each field in the form to be read based on the predicted recognition time of each field output from the first character recognition means 5;
The function is to create assignment information indicating which of the second character recognition means 6 to assign each field to, and to assign each field to a plurality of character recognition means from that information. A function of storing the obtained number of characters, average character size information in the field, and actual recognition time of each field in the neural network means 8, and a function of storing the information in the neural network means 8 at the time when a certain number of forms are read. This is a function that causes the neural network means 8 to learn using data.

Next, the operation of this embodiment will be explained. It should be noted that explanations that overlap with those described in the [Prior Art] section will be omitted, and only newly added or changed means will be explained here. A scanning means 2 reads a form 1 as shown in FIG. 11, and a character cutting means 4 cuts out characters in each field.

The control means 7a obtains information on the number of characters in each field and average character size as shown in FIG. 2 from the character extraction means 4, and format information as shown in FIG. 3 from the format information storage means 3. Next, the control means 7
a sends the information shown in FIGS. 2 and 3 to the neural network means 8 for each field. When the neural network means 8 is initialized, it is assumed that the predicted recognition time Tf of the field is calculated using the following equation and the network is trained. Tf=1.5×(number of characters in field)
[Seconds]・・・(1)

[0018] Therefore,
Neural network means 8 when reading the form 1 in FIG. 11 with the neural network means 8 at the time of initialization
The output of the predicted recognition time is shown in FIG. Learning during initialization is not performed until complete convergence, and learning is stopped when the error falls below an appropriate threshold, so the value calculated using equation (1) will not be completely achieved. , the results include some errors.

Next, based on the predicted recognition time of the neural network means 8, the control means 7a allocates character recognition means as shown in FIG. 5, and performs recognition control accordingly. In this case, the prediction of recognition time is simply calculated based on the number of characters information, so the recognition time will be a different value from the actual recognition time. Load distribution based on this value is the same as in the conventional example, so Figure 12
As shown in , the recognition time is 32 seconds.

Next, the control means 7a stores the input parameters sent to the neural network means 8 and the recognition time of each field actually read as learning data in the neural network means 8. The learning data stored in the neural network means 8 is shown in FIG. The control means 7a causes the neural network means 8 to perform learning using the accumulated learning data. As with the initialization of the network, learning ends when the error falls below a certain threshold.

Next, the form 1 shown in FIG. 13 is read in the same manner, and format information and information on the cutting result are sent to the neural network means 8. FIG. 7 shows the input parameters and output results of the neural network means 8. Note that the format information is the same as the form shown in FIG. Since there is no change in the input data other than a slight change in the input font size, the predicted recognition time output of the neural network means 8 in FIG. 7 is almost the same as the actual processing time when reading the form in FIG. 11. are the same. The control means 7a performs load distribution as shown in FIG. 8 based on this result. FIG. 9 shows the results of recognition processing of the fields of the form shown in FIG. 13 using the load distribution results. As can be seen from FIG. 9, the recognition processing time when processing the form shown in FIG. It is only 1 second.

Example 2. In the first embodiment, the neural network means 8 is used to predict the recognition time, but in this embodiment, the same operation can be expected even if the calculation formula (1) is used for prediction.

Example 3. In the above example, the number of characters, average character size, character type, and number of target characters were used as input to the neural network means for predicting recognition time, but if all of these information is not used, or other information such as character frame size is used. However, you can expect similar behavior.

Example 4. In the above embodiment, the case where there are two character recognition means has been described, but the same operation can be expected when there are three or more character recognition means.

Example 5. In the above embodiment, the processing unit is divided into field units, but the same effect can be expected even if the processing unit is divided into character units or larger units.

Example 6. In the above embodiment, the learning was performed after reading the first sheet, but the same effect can be expected even if the learning is performed after reading a plurality of sheets.

[0027]

As described above, according to the present invention, a prediction means is provided for inputting information obtained before performing recognition processing when reading a form, and calculating a predicted recognition time for each recognition area to be read. We found a load distribution method that maximizes throughput from the predicted recognition time of the recognition area, and controlled multiple character recognition methods based on the load distribution information, resulting in a significant improvement in recognition processing throughput. be.

[0028] Furthermore, a neural network is used as a prediction means, and learning parameters for training the neural network are stored each time a form is read, and the neural network is trained at appropriate intervals, so that the same format can be used. When repeatedly reading and processing multiple forms, the recognition time can be predicted more accurately and the throughput can be optimized.

[Brief explanation of drawings]

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

FIG. 2 is a diagram showing information obtained from character cutting means 4. FIG.

FIG. 3 is a diagram showing information obtained from format information storage means 3. FIG.

FIG. 4 is a diagram showing the output of the neural network means 8 at the time of initialization.

FIG. 5 is a diagram showing load distribution determined by the control means 7a.

6 is a diagram showing learning data stored in the neural network means 8. FIG.

FIG. 7 is a diagram showing the input and output of the neural network means 8 after learning.

FIG. 8 is a diagram showing load distribution determined by the control means 7a when reading the form in FIG. 13;

FIG. 9 is a diagram showing throughput when controlled using the load distribution information of FIG. 8;

FIG. 10 is a configuration diagram showing a conventional character recognition device.

FIG. 11 is a diagram showing a form 1 to be read first.

FIG. 12 is a diagram showing load distribution determined by the control means 7 when reading the form in FIG. 11;

FIG. 13 is a diagram showing the second form 1 to be read.

14 is a diagram showing load distribution determined by the control means 7 when reading the form in FIG. 13. FIG.

[Explanation of symbols]

1. Report 2 Scanning means 3 Format information storage means 4 Character cutting means 5 First character recognition means 6 Second character recognition means 7a Control means 8 Neural network means

Claims (3)

[Claims] 1. Scanning means for scanning a recording medium; format information storage means for storing format information of the position of a field to be recognized in the recording medium and the type of recognized character;
Character cutting means for cutting out characters from the character string obtained from the scanning means using the information in the format information storage means, a plurality of character recognition means for recognizing the characters cut out from the character cutting means, and processing time for each recognition area. , a control means for controlling the scanning means, the format information storage means, the character cutting means, the character recognition means, and the prediction means, based on the processing time predicted by the prediction means. A character recognition device characterized in that the load of the plurality of character recognition means is distributed. 2. Character recognition according to claim 1, wherein a neural network is used as the prediction means, and information from the format information storage means and information from the character extraction means are inputted to predict the processing time. Device. 3. The character recognition device according to claim 1, wherein a neural network is used as the prediction means, and the processing time is predicted by inputting the recognition time of the character recognition means.