JPH07239913A - Character recognition device - Google Patents

Abstract

PURPOSE:To provide a character recognition device with excellent real time performance in which a mis-reading rate is reduced through learning. CONSTITUTION:The character recognition device 1 has a recognition section (1st recognition means) 4, a discrimination section 6, a re-recognition section (2nd recognition means) 7, and a correction means 8. When the recognition section 5 decides a character pattern candidate with a highest similarity (evaluation value) in the reading stage, the judging section 6 judges the candidate. When the judging section 6 judges the candidate as a candidate liable to incur mis-reading, the re-recognition section 7 applies recognition processing to an input character pattern corresponding to the candidate. A correction means 8 corrects a judging criterion and a processing criterion so that the judging section 6 and the re-recognition section 7 make surer judgement and recognition based on the judging result by the judging section 6 and the recognition result by the re-recognition section 7 in the learning stage.

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,
More specifically, it relates to a character recognition device using a neural network.

[0002]

2. Description of the Related Art Recently, it has been attempted to perform character recognition using a neural network.

In the character recognition apparatus using this neural network, when an error occurs in the recognition result of a certain input character pattern, a correct character pattern candidate can be obtained in the next character recognition for the input character pattern. By correcting the arc weight value in the neural network with directionality, it is possible to stably perform good character recognition according to changes in the characteristics of the input character pattern.

[0004]

However, the conventional character recognition device using a neural network has a problem that it lacks real-time property because it performs character recognition on all characters by the neural network.

Therefore, the present invention has been made in view of the above circumstances, and an object of the present invention is to provide a character recognition device which can reduce the misreading rate by learning and which is excellent in real time. is there.

[0006]

In order to achieve the above object, a character recognition apparatus according to claim 1 collates an input character pattern with a recognition dictionary to obtain an evaluation value, and the character recognition result is the most effective as a recognition result of the input character pattern. A first recognizing unit that determines a character pattern candidate having a high evaluation value, and a determining unit that determines whether or not the character pattern candidate determined by the first recognizing unit is a candidate that is likely to cause misreading according to a changeable determination standard. When,
When the determination unit determines that the candidate is likely to cause misreading, the second character recognition unit that recognizes the character by performing feature extraction according to a process standard that can be changed for the input character pattern corresponding to the candidate, and the determination unit. Correction means capable of correcting the judgment standard and the processing standard so that the judgment means and the second recognition means can make more reliable judgment and recognition based on the judgment result of the means and the recognition result of the second recognition means, respectively. It is characterized by having.

Further, the character recognition device according to the second aspect collates an input character pattern with a recognition dictionary to obtain an evaluation value,
First recognition means for deciding a plurality of character pattern candidates having a relatively high evaluation value as a recognition result of the input character pattern;
It is determined whether or not the character pattern candidate having the highest evaluation value determined by the first recognizing means is a candidate that is likely to cause misreading according to a changeable determination criterion based on each evaluation value obtained by the first recognizing means. If the first determination unit and the first determination unit determine that the candidate is likely to cause misreading, feature extraction is performed on the input character pattern corresponding to the candidate according to a changeable processing standard, and the character is recognized. Second recognition means and determination as to whether or not the character recognized by the second recognition means is present in a predetermined upper number of character pattern candidates among the plurality of character pattern candidates determined by the first recognition means. And a second judging means for rejecting the character when there is no character, and the judging means and the second recognizing means based on the judgment result of the judging means and the recognition result of the second recognizing means. It is characterized in that it has a correction means capable of correcting the criteria and treatment standards to allow a more reliable determination and recognition, respectively.

The character recognition device according to a third aspect of the invention is characterized in that at least the judging means and the second recognizing means are constituted by a neural network.

[0009]

The character recognition device according to the first aspect will be described separately for the reading stage and the learning stage.

In the reading step, when the first recognition means determines the character pattern candidate having the highest evaluation value, the determination means makes a determination for that candidate. Then, when it is determined that the candidate is likely to cause misreading, the second recognizing means performs the recognition process on the input character pattern corresponding to the candidate. In the learning stage, if the determination result of the determination means is incorrect, the correction means corrects the determination criteria so that the determination means can make a more reliable determination. If the recognition result of the second recognizing means is incorrect, the correcting means corrects the processing standard so that the second recognizing means can perform more reliable recognition. As the learning of the determining means progresses, it becomes more reliable to determine whether or not the candidate is likely to cause misreading, and as the learning of the second recognizing means progresses, the recognition by the second recognizing means becomes more reliable and the misreading rate decreases. To do.

Further, according to the character recognition device of the second aspect, since the correctness of the re-recognition result by the second recognition means is judged by the second judgment means, the misreading rate is further reduced.

According to the character recognition device of the third aspect, the correction standard is used to modify the arc load value in the neural network that constitutes the determination means and the second recognition means, thereby making the determination standard and the processing standard. Can be fixed quickly.

[0013]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a schematic block diagram showing an embodiment of the character recognition apparatus of the present invention.

The apparatus 1 of this embodiment includes a scanner unit 2 for detecting an image of a document, a character cutout unit 3 for cutting out a character pattern for each character from the image of the document detected by the scanner unit 2, and this character cutout. A recognition unit 5 as a first recognition unit that determines a plurality of character pattern candidates by collating the input character pattern from the unit 3 with the character pattern candidates stored in the recognition dictionary unit 4, and the recognition unit 5 determines the character pattern candidates. A determination unit 6 that determines whether the generated character pattern candidate is a candidate that is likely to cause misreading according to a changeable determination criterion;
When the determination unit 6 determines that the candidate is likely to cause misreading, feature extraction is performed according to a changeable processing standard for an input character pattern corresponding to the candidate, and the character extraction is performed again as a second recognition unit that recognizes the character. Based on the recognition result of the recognition unit 7, the judgment unit 6 and the recognition result of the re-recognition unit 7, the judgment standard and the processing standard are modified so that the judgment unit 6 and the re-recognition unit 7 can perform more reliable judgment and recognition, respectively. And a correction means 8 which can be used.

Next, the details of each of the above parts will be described.

The scanner unit 2 comprises a light source for irradiating a document with light, and a photoelectric conversion element for receiving reflected light from the document and converting it into an electric signal. The image of is detected.

The recognition dictionary section 4 stores a large number of character pattern candidates to be matched with the input character pattern from the character cutting section 3.

The recognition unit 5 collates the input character pattern with the character pattern candidates stored in the recognition dictionary unit 4, obtains the degree of similarity as an evaluation value by, for example, a pattern matching method, and compares it as the recognition result of the input character pattern. Of multiple character pattern candidates with a high degree of physical similarity is determined by the character code,
It is output to the determination unit 6 together with the similarity. The evaluation value is not limited to the similarity and may be another evaluation value. Also,
The evaluation method is not limited to the pattern matching method and may be another method.

The correction means 8 corrects the display section 9 such as a CRT display, the input section 10 equipped with a keyboard, a mouse, etc., the judgment standard of the judgment section 6 and the processing standard of the re-recognition section 7, and The control unit 11 controls each unit of the apparatus 1.

The determination unit 6 is composed of, for example, a neural network as shown in FIG. That is, the determination unit 6 includes an input layer 60 including a plurality of input nodes 60a, an intermediate layer 61 including a plurality of intermediate nodes 61a connected to each input node 60a via an arc 63 through a network, and each intermediate node. 61a and an output layer 62 including a pair of output nodes 62a connected to each other via a network via an arc 64.

The similarity (first rank to nth rank) for each character pattern candidate calculated by the pattern matching method by the recognition unit 5 is input to each input node 60a. The determination result is output from the output node 62a to the control unit 11.

A load value (real number) indicating the strength of the connection between the nodes 60a, 61a, 62a is set for each arc 63, 64. By setting this load value,
This means that the judgment standard for the judgment unit 6 to make the judgment is set.

When each similarity calculated by the recognition unit 5 is input to each input node 60a, the character pattern candidate having the highest similarity recognized by the recognition unit 5, that is, the first character pattern candidate, which is recognized by the recognition unit 5, is set according to the set criterion. It is determined whether or not the first-ranked character pattern candidate is a candidate that is likely to cause misreading, and the determination result is output from the output node 62a to the control unit 11. For example, when it is determined that the candidate is likely to cause misreading, “1” is output, and in other cases, “0” is output.

Each load value is controlled by the control unit 11 based on the operation of the input unit 10 such as by inputting the correct character by moving the cursor to the misread character in the recognition result displayed on the display unit 8, for example. , Error backpropagation (backpropagation) method. By correcting these load values, the judgment criteria will be learned and corrected. That is, when the case of one node is described as an example, the output y of that node is represented by y = f (Σωi · xi) (1) where xi is the input data and ωi is the load value. On the other hand, as case data, {y '(s), x
i ′ (s) | s = 1, ..., S} exists, and the output value obtained by inputting x ′ (s) into the node of the load value ω is y (ω,
x ′ (s)), the error function R (ω) is It is represented by. Here, in equation (1), xi = xi ′ (determination unit 6
When the input is the similarity when the erroneous determination is made, the output y (= f (Σωi ·
xi ′)) is obtained, ωi is gradually changed to obtain ωi that minimizes the error function R (ω).
This is performed for each node, and the processing standard of the determination unit 6 is corrected.

As the criterion, for example, the similarity of the first rank of the character pattern candidate is lower than a predetermined value, or the score difference between the first rank and the second rank of the character pattern candidate is predetermined. If the value is less than or equal to the value, it is determined that the candidate is likely to cause misreading,
Other than that is determined to be correct reading. For example, if the first, second, third, and fourth similarities of the character pattern candidates are 100 points, 90 points, 80 points, and 70 points, respectively,
Since the score difference in the degree of similarity between the second rank and the second rank is small, the first character pattern candidate is determined to be a candidate that easily causes misreading. Further, when the first rank of the character pattern candidates is 100 points and all of the second and subsequent ranks are 0 points, the score difference between the first rank and the second rank is large, and thus it is determined to be correct reading.

The re-recognition unit 7 is composed of, for example, a neural network as shown in FIG. That is,
The re-recognition unit 7 includes an input layer 70 including a plurality of input nodes 70a, an intermediate layer 71 including a plurality of intermediate nodes 71a connected to each input node 70a via an arc 73, and each intermediate node 71a. And a plurality of output nodes 72 connected in a network via
and an output layer 72 made of a.

At each input node 70a, a character cutout unit 3 is provided.
Of the character patterns cut out in step 2, the character patterns of the image that are determined as candidates that are likely to cause misreading by the determination unit 6 are input as patterns 0 and 1. The recognition result is output from the output node 72a to the control unit 11.

A load value (real number) indicating the strength of the connection between the nodes 70a, 71a, 72a is set for each arc 73, 74. By setting this load value,
This means that the processing standard for the re-recognition unit 7 to recognize is set.

When a character pattern determined to be a candidate for misreading is input to the input node 70a, feature extraction is performed according to the set processing standard, the character is recognized as a character code, and the recognition result is The output node 72a is adapted to output to the control unit 11.

Each load value is controlled by the control unit 11 based on the operation of the input unit 10 such as by inputting a correct character by moving the cursor to a misread character in the recognition result displayed on the display unit 8, for example. , Error backpropagation (backpropagation) method. By modifying these load values, the processing criteria will be learned and modified. That is, similar to the determination unit 6, when the input of xi = xi '(character pattern misread by the re-recognition unit 7) is given to the equation (1), the output close to the teacher data (correct character code) y' y (= f (Σωi · xi ′))
So that ωi is gradually changed to obtain ωi that minimizes the error function R (ω), this is performed for each node, and the processing standard of the re-recognition unit 7 is corrected.

The learning method of the neural network forming the judgment unit 6 and the re-recognition unit 7 is the back propagation method here, but is not limited to this. Also,
The number of layers is not limited to the above three layers, and may be four or more layers. Further, if the judgment unit 6 and the re-recognition unit 7 can learn,
The configuration is not limited to the neural network, and other configurations may be used.

Next, the operation of this embodiment will be described by dividing it into a reading stage and a learning stage according to the flowchart of FIG.

First, in the reading stage, the scanner unit 2
However, under the control of the control unit 11, when the image of the document is detected by optically scanning the entire document, the character cutting unit 3
The image of the document is cut out character by character and output to the recognition unit 5. Here, one of the cut-out character patterns is "closed" as shown in FIG.

The recognition unit 5 performs a recognition process on the character pattern using the character pattern candidates stored in the recognition dictionary unit 4, and a determination unit 6 determines a predetermined number of character pattern candidates and similarities in descending order of similarity. Output (S1). An example of the output is shown in FIG. As shown in the figure, the character pattern candidates and the similarity are the first place with “Seki” (240 points),
"Closed" (222 points) in second place, "Open" (203) in third place
It is assumed that “gate” (158 points) is output to the sixth place.

The determination unit 6 determines whether or not the first-ranked character pattern candidate (“Seki” in the example of FIG. 5) is a candidate that tends to cause misreading, and outputs the determination result to the control unit 11 ( S
2). For example, if it is determined that the candidate is likely to cause misreading,
"1" indicating that is output, and when it is determined that the candidate is the correct reading, "0" indicating that is output.

The control unit 11 analyzes the judgment result of the judgment unit 6 (S3), and the judgment result is "1" (a candidate that is likely to cause misreading).
If it is determined that the character pattern (“closed” in the example of FIG. 4) corresponding thereto is input to the re-recognition unit 7. Under the control of the control unit 11, the re-recognition unit 7 performs feature extraction on the character pattern (“closed”) according to the processing standard,
The character is recognized (S4).

The recognition result is displayed on the display section 8 under the control of the control section 11 (S5). in this case,
Seki is displayed and it is misread.

Next, in the learning stage, the operator compares the recognition result displayed on the display unit 8 with the document to be read, and when he / she finds the misread character "Seki", operates the input unit 10. , The cursor is aligned with the misread character "Seki" and the correct character "closed" is input (S6).

Subsequently, the judgment result of the judgment unit 6 in the step S3 is analyzed (S7), and the judgment result is "0".
If it is, it can be considered that the result is erroneously read by the recognition unit 5 and erroneously determined by the determination unit 6.
The similarity with respect to each character pattern candidate as shown in is input to the determination unit 6, and the learning is performed by correcting the determination criterion in the determination unit 6 so that the output is “1” (a candidate that is likely to cause misreading) (S8). ). This enables more reliable determination.

In step S7, the determination result is "1".
If it is, it can be considered that the result is erroneously read by the recognition unit 5, correctly determined by the determination unit 6, and further erroneously read by the re-recognition unit 7. Therefore, the control unit 11 re-recognizes the character pattern “closed”. 7 and re-recognition unit 7 so that the output is "closed".
Learning is performed by correcting the processing standard in (S9). This enables more reliable recognition.

According to the present embodiment as described above, as the learning of the judging unit 6 progresses, it becomes more reliable to judge whether or not there is a candidate that is likely to cause misreading, and as the learning of the re-recognizing unit 7 progresses, the re-recognizing unit 7 increases. The recognition by 7 becomes more reliable, and the misreading rate can be reduced. Further, by correcting the arc load value in the neural network forming the judgment unit 6 and the re-recognition unit 7 by the correction means 8, the judgment standard and the processing standard can be rapidly modified, and the real-time property is excellent. .

As another embodiment of the present embodiment, whether or not the character recognized by the re-recognition unit 7 exists in a predetermined number of character pattern candidates among the plurality of character pattern candidates determined by the recognition unit 5 or not. It is also possible to provide a second determination means for performing the above determination and rejecting the character when there is no such determination. Thereby, the misreading rate can be further reduced.

The present invention is not limited to the above embodiment,
Various modifications can be made without changing the gist of the invention.

[0045]

According to the invention described in claim 1 described in detail above, as the learning of the judging means progresses, it becomes more reliable to judge whether or not the candidate is likely to cause misreading, and the learning of the second recognizing means becomes possible. As the process proceeds, the recognition by the second recognition means becomes more reliable, so that the misreading rate can be reduced.

According to the second aspect of the invention,
Since the correctness of the re-recognition result by the recognizing means is determined by the second determining means, the misreading rate can be further reduced.

According to the third aspect of the present invention, the judgment standard and the processing standard can be promptly modified by modifying the arc load value in the neural network forming the judging means and the second recognizing means. Therefore, the real-time property is excellent.

[Brief description of drawings]

FIG. 1 is a schematic block diagram showing an embodiment of an optical character reading device of the present invention.

FIG. 2 is a configuration diagram of a determination unit according to the present exemplary embodiment.

FIG. 3 is a configuration diagram of a re-recognition unit according to the present embodiment.

FIG. 4 is a diagram showing an example of an input character pattern of this embodiment.

FIG. 5 is a diagram showing an output example of a recognition unit of the present embodiment.

FIG. 6 is a flowchart showing the operation of this embodiment.

[Explanation of symbols]

 1 Character Recognition Device 5 Recognition Unit (First Recognition Unit) 6 Judgment Unit 7 Re-recognition Unit (Second Recognition Unit) 8 Correction Unit

Claims (3)

[Claims] 1. A first recognizing means for determining an evaluation value by collating an input character pattern with a recognition dictionary, and determining a character pattern candidate having the highest evaluation value as a recognition result of the input character pattern; Judgment means for judging whether or not the character pattern candidate decided by the recognition means is a candidate for causing misreading according to a changeable judgment criterion, and if this judgment means is determined as a candidate for causing misreading, A second recognition unit that performs feature extraction for a corresponding input character pattern according to a changeable processing standard and recognizes the character, and the determination unit based on the determination result of the determination unit and the recognition result of the second recognition unit. And a second recognizing means, and a modifying means capable of modifying the judgment standard and the processing standard so as to make more reliable judgment and recognition, respectively. Character recognition device. 2. A first recognition means for determining an evaluation value by collating an input character pattern with a recognition dictionary and determining a plurality of character pattern candidates having a relatively high evaluation value as a recognition result of the input character pattern, First, it is determined whether or not the character pattern candidate having the highest evaluation value determined by the first recognition means is a candidate that is likely to cause misreading, based on each evaluation value obtained by the first recognition means, according to a changeable determination criterion. And the first deciding means, when it is determined that the candidate is liable to be misread, the feature extraction is performed according to a changeable processing standard for the input character pattern corresponding to the candidate, and the character is recognized. And the character recognized by the second recognizing means is included in a predetermined upper number of character pattern candidates among the plurality of character pattern candidates determined by the first recognizing means. Whether the character is present or not, and if it is not present, second determining means for rejecting the character, and the determining means and the second recognition based on the determination result of the determining means and the recognition result of the second recognizing means. A character recognizing device comprising: a correction unit capable of correcting the determination standard and the processing standard so that each unit can perform more reliable determination and recognition. 3. The character recognition device according to claim 1, wherein at least the judging means and the second recognizing means are constructed by a neural network.