KR102264988B1 - Traditional Korean character Hanja Recognition System and method using thereof - Google Patents

Abstract

The present invention relates to a Chinese character recognition system and a Chinese character recognition method using the same, and according to the present invention, an input unit for receiving a Chinese character image, which is an image of a Chinese character to be identified; a feature extraction unit including a plurality of image recognizers to which a deep learning model for extracting image features from a Chinese character image is applied; an identified Chinese character selection unit that derives reliability for each of the plurality of image features and selects an identified Chinese character corresponding to a Chinese character to be identified; and an output unit for receiving and outputting the identified Chinese characters from the identified Chinese character selection unit; Disclosed is a technology capable of increasing the accuracy of recognizing Chinese characters even though it is difficult to recognize, or the quality of various handwriting or sizes, such as cursive or cursive type, or the quality of Chinese character records is deteriorated.

Description

Traditional Korean character Hanja Recognition System and method using thereof

The present invention relates to a Chinese character recognition system capable of providing an identification result so that a user can recognize Chinese characters by accurately identifying Chinese characters written in cursive or cursive fonts mainly used in ancient documents such as traditional records, and a Chinese character recognition method using the same.

In recent years, interest in discovering cultural contents is increasing worldwide, and various efforts are being made to diversify cultural contents. Among them, in the field of bibliography, research activities are being carried out to analyze classical historical materials that are mainly used, and to explore traditional culture or historical events or backgrounds based on the analyzed contents. For this purpose, accurate identification and analysis of historical materials is conducted. This should be done.

However, most of the historical materials are written in an ancient language that is not often used in the present era, so it is difficult to analyze the historical documents. Among them, in the case of characters written in cursive or cursive fonts, which are part of Chinese characters, accurate identification and identification of characters This is not easy. Accordingly, there is a demand for an advanced recognition technology capable of accurately identifying characters written in such cursive or cursive fonts.

Republic of Korea Patent Publication No. 10-1996-0008529 Republic of Korea Patent Publication No. 10-2002-0088124

An object of the present invention is to solve the conventional problems as described above, and to provide a Chinese character recognition system that can accurately identify Chinese characters written in cursive or cursive fonts mainly used in ancient documents such as traditional records, and a Chinese character recognition method using the same. is in

A Chinese character recognition system according to an embodiment of the present invention for achieving the above object includes an input unit for receiving a Chinese character image that is an image of a Chinese character to be identified; a feature extracting unit including a plurality of image recognizers to which a deep learning model for extracting image features from the Chinese character image received from the input unit is applied; an identified Chinese character selection unit for deriving reliability for each of the plurality of image features received from the feature extraction unit, and selecting an identified Chinese character corresponding to the Chinese character to be identified; and an output unit for receiving and outputting the identified Chinese character selected by the identified Chinese character selecting unit from the identified Chinese character selecting unit; It may be characterized as one of including.

Here, the Chinese character image transmitted from the input unit to the feature extraction unit has the form of a multi-channel image matrix, and a plurality of the image recognizers that have received the Chinese character image having the form of a multi-channel image matrix are each of a row vector. Extracting an image feature having a shape may be another feature.

Here, among the plurality of image recognizers to which the deep learning model is applied, it may be further characterized that there is the image recognizer that has been trained through a sensitivity normalization method.

Here, among the plurality of image recognizers to which the deep learning model is applied, another feature may be that there is the image recognizer that has been trained through a robustness normalization method.

Here, among the plurality of image recognizers to which the deep learning model is applied, another feature may be that there is the image recognizer that has been trained through a sensitivity and robustness normalization method.

Here, another feature may be that the identified character selection unit calculates a reliability matrix composed of reliability scores by using an activation function for each of the plurality of image features received from the feature extraction unit.

Here, the identified character selection unit selects a maximum value from each column of the calculated reliability matrix to generate a reliability score-row vector, and the confidence score corresponds to the maximum reliability score with the largest reliability score in the generated reliability score-row vector Another feature may be to select a Chinese character to be identified as the identified Chinese character corresponding to the Chinese character to be identified.

Chinese character recognition method according to an embodiment of the present invention for achieving the above object, the Chinese character recognition system according to the present invention, the input step of receiving a Chinese character image that is an image of a Chinese character to be identified; an extraction step of extracting a plurality of image features from the Chinese character image received in the input step; a reliability derivation step of deriving a reliability score from the plurality of image features extracted in the extraction step; and an identified Chinese character selection step of selecting an identified Chinese character corresponding to the Chinese character to be identified from a plurality of the reliability scores derived in the reliability deriving step.

Here, as a step performed after the identified Chinese character selecting step, an output step of outputting the identified Chinese character selected in the identified Chinese character selecting step so that it can be recognized from the outside may be further included.

Here, the selection of the identified Chinese character in the identified Chinese character selection step may be performed by selecting a Chinese character corresponding to the maximum reliability score, which is the largest value among a plurality of the reliability scores, as the identified Chinese character.

The Chinese character recognition system and the Chinese character recognition method using the same according to the present invention are difficult to recognize due to the passage of a long time, or various handwriting or sizes such as cursive or cursive fonts, high similarity between characters, and the quality of Chinese character records It increases the accuracy of Chinese character recognition. Therefore, it has the effect of contributing to recognizing the Chinese characters of traditional records, which were difficult to recognize in the past. In addition, it has the effect of contributing to the improvement of the efficiency of archival records that digitize and preserve traditional records.

1 is a block diagram schematically showing a Chinese character recognition system according to an embodiment of the present invention.
2 is a flowchart schematically illustrating a Chinese character recognition method using a Chinese character recognition system according to an embodiment of the present invention.

Hereinafter, a preferred embodiment will be described with reference to the accompanying drawings so that the present invention can be understood in more detail.

After a description of the Chinese character recognition system according to an embodiment of the present invention with reference to FIG. 1, a Chinese character recognition method according to an embodiment of the present invention will be described with further reference to FIG. 2 .

For reference, for convenience of explanation and understanding of the embodiment of the present invention, 'Chinese character to be identified' refers to a Chinese character to be identified by the user, and 'Chinese character to be identified' is a Chinese character according to an embodiment of the present invention. It refers to the identification result obtained by identifying 'the Chinese character to be identified' through a recognition system or a Chinese character recognition method.

1 is a block diagram schematically showing a Chinese character recognition system according to an embodiment of the present invention.

Referring to FIG. 1 , the Chinese character recognition system 100 according to an embodiment of the present invention may include an input unit 110 , a feature extraction unit 120 , an identified Chinese character selection unit 130 , and an output unit 140 . have.

The input unit 110 receives a Chinese character image that is an image of a Chinese character to be identified.

An example of such an input unit 110 may be a device such as a camera.

The Chinese character image transmitted from the input unit 110 to the feature extraction unit 120 has the form of a multi-channel image matrix. The input unit 110 makes the Chinese character image in the form of an image matrix of 3 RGB channels with a size of (W*H*3). Here, W is the width of the image, and H is the height of the image.

As described above, the Chinese character image in the form of a matrix is transmitted from the input unit 110 to the image recognizers 121 , 122 , 123 , 124 , and 125 of the feature extraction unit 120 .

The feature extraction unit 120 includes a plurality of image recognizers 121 , 122 , 123 , 124 , and 125 for extracting image features from the Chinese character image received from the input unit 110 . That is, the feature extraction unit 120 can be said to be an aggregate including a plurality of image recognizers (121, 122, 123, 124, 125).

The image recognizers 121 , 122 , 123 , 124 , and 125 perform a specific number of convolution and pooling operations to extract image features from a Chinese character image.

Multiple image recognizers (121, 122, 123, 124, 125) that received a Chinese character image in the form of a multi-channel image matrix like an RGB 3-channel image matrix of (W*H*3) Extract image features in the form of row vectors.

If there are x number of Chinese characters that are likely to be selected as the Chinese characters to be identified as corresponding to the Chinese characters to be identified, the extracted image feature has the form of a row vector of (1*x).

As described above, each of the plurality of image recognizers 121 , 122 , 123 , 124 , and 125 extracts image features from the Chinese character image received from the input unit 110 .

A deep learning model for extracting image features is applied to each of the plurality of image recognizers 121 , 122 , 123 , 124 , and 125 .

Here, as the deep learning model, various deep learning models such as a convolution neural network (CNN), a densely connected convolutional network (DenseNet), or a residual network (ResNet) are possible, and a Visual Geometry Group Network (VGGNet) is also preferable as a deep learning model.

It is preferable that the image recognizers 121, 122, 123, 124, and 125 to which such a deep learning model is applied are trained through the following normalization.

Sensitivity regularization makes it better to distinguish between-class images, which look similar but are different Chinese character images.

And robustness regularization makes it better to distinguish between-class images, which look different but are actually the same Chinese character image.

Through such normalization, the same kind of Chinese character images (within-class) can be distinguished as the same, and different kinds of Chinese character images (between-class) can be distinguished as more different.

Therefore, among the plurality of image recognizers 121 , 122 , 123 , 124 , and 125 to which the deep learning model is applied, it is also preferable that the image recognizer 121 is trained through the sensitivity normalization method.

In addition, among the plurality of image recognizers 121, 122, 123, 124, and 125 to which the deep learning model is applied, it is also preferable that there is an image recognizer 122 that has been trained through a robustness normalization method.

In addition, among the plurality of image recognizers 121, 122, 123, 124, and 125 to which the deep learning model is applied, it is also preferable that there are image recognizers 123, 124, 125 that have been trained through the sensitivity and robustness normalization method.

Among the plurality of image recognizers 121, 122, 123, 124, and 125 to which the deep learning model is applied, some image recognizers 121 are trained through the sensitivity normalization method, and other image recognizers 122 are robustness normalized. Learning is made through the method, and it is also preferable that the remaining image recognizers 123, 124, and 125 are trained through the sensitivity and robustness normalization method.

As described above, it is preferable that each of the plurality of recognizers 121 , 122 , 123 , 124 , and 125 is in a state that has been learned through the normalization method.

은 가중치 갱신을 위한 학습률(learning rate)이며,

은 각

층에 해당하는 가중치의 미분값이다. 본 발명에서는 가중치의 미분값을 구할 수 있는 손실함수에 민감도 혹은 강건성 정규화를 적용한다. The weight update method follows the general method below. At this time

is the learning rate for weight update,

silver angle

It is the derivative of the weight corresponding to the layer. In the present invention, sensitivity or robustness normalization is applied to a loss function that can obtain a differential value of a weight.

The loss function to which the normalization method is applied can be defined as follows.

That is, the average of the differential values of the last layers of the image recognizers 123, 124, and 125 is calculated for normalization of sensitivity and robustness, and R, which is a regularizer, is calculated.

을 나누어 준다. In the case of the loss value of the sensitivity normalization, it is the value obtained by adding the regularizer R and the margin value m, which is the loss value of the CNN (convolution neural network).

share the

가 0으로 나누어져 무한대가 되는 경우를 방지하기 위하여 필요하다. Here the margin is

This is necessary to prevent the case where is divided by 0 to become infinity.

Loss of sensitivity normalization:

을 바탕으로 이미지인식기(121)를 구성하는 가중치를 갱신하기 위한 값은 다음의 식과 같이 구해진다. This newly defined loss value

A value for updating the weight constituting the image recognizer 121 based on

는 l번째 층의 입력 특징맵이며,

은 l번째 층으로 전파되는 에러값이다.here

is the input feature map of the l-th layer,

is the error value propagated to the l-th layer.

에 레귤라이저 R을 더해줌으로써 정의된다. In the case of the loss value of robustness normalization, the loss value of CNN is

It is defined by adding a regulator R to

Loss of robustness normalization:

The value for updating the weight of the image recognizer based on robustness normalization is obtained as follows.

에 앞서 정의한 민감도 및 강건성 손실값을

,

계수로 가중치 합을 구해서 더해준다.In order to integrate these loss values, the CNN loss

The sensitivity and robustness loss values defined previously in

,

Calculate the weighted sum of the coefficients and add them.

계수를 적용하여

와 정규화 항 사이의 계수를 정의하여 얼마나 각 항을 반영할지 정하여, 최종적으로 총 에러값인

를 구한다. in addition to this

by applying a coefficient

Determine how much each term is to be reflected by defining the coefficient between and the regularization term, and finally the total error value

save

When the weight of the image recognizer is updated with the newly defined error value, it has the following gradient value.

을 각 계수인

,

에 대해 편미분을 하여 경사도 값을 구한다. 그리고, 경사도 값을 바탕으로

,

값을 갱신한다. At this time, the total error value so that the ratio between the sensitivity and robustness error values can also be learned during the learning process.

is each coefficient

,

The slope value is obtained by performing partial differentiation with respect to . And, based on the slope value

,

update the value

를 적용하여 학습시킬 수 있다.At this time, the learning rate of the coefficients

can be learned by applying

After implementing the adaptive sensitivity and robustness normalization in this way, it is preferable to train each of the image recognizers 121 , 122 , 123 , 124 , and 125 according to the normalization.

The results extracted by a plurality of different image recognizers 121, 122, 123, 124, and 125 are synthesized using the ensemble method.

A plurality of image recognizers 121, 122, 123, 124, 125 of the image extractor 120 includes an image recognizer 121 to which sensitivity normalization is applied, an image recognizer 122 to which robustness normalization is applied, and an image to which both sensitivity and robustness normalization are applied. There is a recognizer 123 .

For example, one image recognizer 121 to which sensitivity normalization is applied, one image recognizer 122 to which robustness normalization is applied, and three image recognizers 123, 124, 125 to which both sensitivity and robustness are applied are image extractors 120 ) may have a form provided within. The image recognizers 121 , 122 , 123 , 124 , and 125 are preferably trained in advance according to the definition of the loss value equation and the gradient equation previously defined.

The image recognizer 121 to which the sensitivity normalization is applied is preferably learned using the aforementioned sensitivity loss value.

It is preferable that the image recognizer 122 to which the robustness normalization is applied is learned using only the aforementioned robustness loss value.

을 바탕으로 하여 학습될 수 있다. In the case of the image recognizer 123 to which both sensitivity and robustness are applied, the total loss value is

can be learned based on

,

계수,

계수, 가중치 갱신을 위한 학습률(

), 계수들의 학습률(

)을 서로 다른 값으로 설정된 것이 바람직하다. In order to be able to extract different image features for each image recognizer (121, 122, 123, 124, 125)

,

Coefficient,

Learning rate for updating coefficients and weights (

), the learning rate of the coefficients (

) is preferably set to different values.

Accordingly, each of the image recognizers 121, 122, 123, 124, and 125 is learned with different image recognizers 121, 122, 123, 124, and 125 during the learning process.

As mentioned above, the Chinese character image received from the input unit 110 is transmitted to each image recognizer 121 , 122 , 123 , 124 , and 125 in the form of a three-dimensional matrix. And each learned image recognizer (121, 122, 123, 124, 125) extracts different image features.

When there are x possible types of Chinese characters, the image recognizers 121, 122, 123, 124, and 125 output an image feature vector of (1*x).

Then, the output image feature vector is transmitted to the identified character selection unit 130 .

In this way, by taking into account the minute fluctuations and high-frequency components in the inputted Chinese character image through sensitivity normalization, they look similar but actually distinguish different Chinese characters.

And robustness normalization can extract image features even from images with ink stains, low resolution, or complex backgrounds, so it can distinguish between seemingly different but actually identical Chinese characters.

In addition, the ratio of the sensitivity normalization and the robustness normalization may be determined by oneself through learning.

The identified character selection unit 130 derives reliability for each of a plurality of image features received from each of the image recognizers 121 , 122 , 123 , 124 , and 125 of the feature extraction unit 120 . And based on the derived reliability, the identified Chinese characters corresponding to the Chinese characters to be identified are selected.

Since the identified Chinese character selection unit 130 applies the ensemble method to a number of image features obtained from several image recognizers 121, 122, 123, 124, and 125, it is possible to recognize the Chinese character, which is the target of identification, with high accuracy. do.

The identified character selection unit 130 calculates a reliability matrix composed of reliability scores by using an activation function for each of the plurality of image features received from the feature extraction unit 120 . Here, as the activation function, it is also preferable that a softmax function is used.

A reliability score is obtained by applying the Softmax function to image features, and the Chinese character corresponding to the highest score is selected as the Chinese character predicted by the Chinese character recognition system.

The softmax function converts the extracted image feature values into probability values so that the sum of each column becomes 1. At this time, the size of the image feature vector is maintained.

And the identified character selector 130 selects a maximum value from each column of the calculated reliability matrix to generate a reliability score-row vector, and the confidence score corresponds to the maximum reliability score with the largest reliability score in the generated reliability score-row vector. Chinese characters are selected as identifiable Chinese characters corresponding to the Chinese characters to be identified.

Based on the image feature vectors output from the image recognizers 121, 122, 123, 124, and 125, a softmax function, which is an activation function, is applied to each image feature vector to obtain a probability value of the type of Chinese character.

After applying the softmax function, each image feature vector maintains the size of (1*x), but has a reliability score in which the sum of the values in each column becomes 1.

When there are y image recognizers, since there are y reliability vectors of (1*x), a reliability matrix of size (y*x) can be constructed.

The confidence score of the j-th column in the i-th row becomes the confidence score of the j-th Chinese character predicted by the i-th image recognizer.

The maximum value of the reliability score is selected to determine the final result of recognizing the Chinese character input to the input unit from the reliability matrix composed of the reliability scores for each Chinese character for each image recognizer (121, 122, 123, 124, 125) do.

Select the maximum value in each column of the confidence matrix. The confidence score selected in the j-th column takes the maximum value of the confidence scores predicted for the j-th Chinese character by each image recognizer.

If there are x types of Chinese characters, a (1*x) confidence vector is created after selecting the maximum value for each column.

In order to determine which Chinese character the final identification target is from the generated confidence vector, the maximum value is selected again from among the reliability scores.

Then, a Chinese character corresponding to the selected maximum value is selected as an identifiable Chinese character, and this is transmitted to the output unit 140 .

If necessary, a database in which Chinese characters corresponding to each row and column of the reliability vector are stored may be separately provided, and this is also preferable.

The output unit 140 receives and outputs the identified Chinese character selected by the identified character selection unit 130 from the identified character selection unit 130 .

As the output unit 140 , a display device such as a monitor capable of displaying the identified character to be visually recognized by the user is also preferable.

In this way, by providing the user with the identified Chinese character, which is the result of recognizing the Chinese character as the identification target through the output unit 140 , the user can know that the Chinese character as the identification target is the identified character.

A Chinese character recognition method using the Chinese character recognition system 100 as described above will be further described with reference to FIG. 2 . 2 is a flowchart schematically illustrating a Chinese character recognition method using the Chinese character recognition system 100 according to an embodiment of the present invention.

2, the Chinese character recognition method using the Chinese character recognition system according to an embodiment of the present invention includes an input step (S110), an extraction step (S120), a reliability derivation step (S130) and an identified Chinese character selection step (S140) This is done, and more preferably, the output step (S300) may be further included.

<< S110 >>

The input step S110 is a step in which the Chinese character recognition system 100 receives a Chinese character image that is an image of a Chinese character to be identified.

An image of Chinese characters to be identified is input through the input unit 110 . And the input unit 110 transmits the Chinese character image in the form of an image matrix to the feature extraction unit 120 .

The Chinese character image is transmitted to a plurality of image recognizers 121 , 122 , 123 , 124 , 125 of the feature extraction unit 120 .

<< S120 >>

The extraction step (S120) is a step in which the feature extraction unit 120 extracts a plurality of image features from the Chinese character image input in the input step (S110).

A plurality of image recognizers 121 , 122 , 123 , and 124 of the feature extraction unit 120 extracts image features from the Chinese character image received from the input unit 12 .

Here, it is preferable that a deep learning model is applied to each of the plurality of image recognizers 121, 122, 123, 124, and 125 as described above.

<< S130 >>

The reliability deriving step (S130) is a step of deriving a reliability score from a plurality of image features extracted in the extraction step (S120).

The self-selection unit 130 that identifies the image features extracted from each of the plurality of image recognizers 121 , 122 , 123 , 124 and 125 is delivered. And a reliability score is derived.

The identified character selection unit 130 calculates a reliability matrix composed of reliability scores by using an activation function for each of the plurality of image features received from the feature extraction unit 120 . Here, as the activation function, it is also preferable to use a softmax function as described above.

<< S140 >>

The identified Chinese character selection step (S140) is a step of selecting an identified Chinese character corresponding to the Chinese character to be identified from a plurality of reliability scores derived in the reliability derivation step (S130).

The selection of the identified Chinese character in the identified Chinese character selection step (S140) is preferably performed by the identified Chinese character selecting unit 130 selecting the Chinese character corresponding to the maximum reliability score, which is the largest value among the plurality of reliability scores, as the identified character.

A reliability score is obtained by applying the Softmax function to image features, and the Chinese character corresponding to the highest score is selected as the identified Chinese character predicted by the Chinese character recognition system.

The identified Chinese character selection unit 130 may transmit the selected Chinese character to the output unit so that the user can know it.

<< S150 >>

As a step performed after the identified Chinese character selection step (S140), the output step (S150) is a step of outputting the identified Chinese character selected in the identified Chinese character selection step (S140) so that it can be recognized from the outside.

The identified Chinese character selected by the Chinese character selection unit 130 is output through the output unit 140 such as a monitor so that the user can recognize it. Accordingly, the user can see the Chinese characters output through the output unit 140 and know the Chinese characters to be identified.

As a step that can be made before the input step, the deep learning model as described above for the image recognizers 121, 122, 123, 124, and 125 of the Chinese character recognition system 100 to extract a number of image features from the Chinese character image It is also desirable to apply and normalize the training. Although not shown in FIG. 2, this step is called a pre-learning step, and it is also preferable to train a plurality of image recognizers 121, 122, 123, 124, and 125 in advance.

As described above, the Chinese character recognition system and the Chinese character recognition method using the same according to the present invention recognize Chinese characters through the ensemble method by applying the sensitivity normalization and robustness normalization of the image recognizer. is increased

In other words, considering the minute fluctuations and high-frequency components in the inputted Chinese character image through sensitivity normalization, they look similar but actually distinguish different Chinese characters. And robustness normalization can extract image features from images with ink stains, low resolution, or complex backgrounds, so it can distinguish between seemingly different Chinese characters, but actually the same.

In addition, the ratio of the sensitivity normalization and the robustness normalization may be determined by oneself through learning. Since the results of several image recognizers are considered together by applying the ensemble method to several image recognizers configured as described above, it is possible to recognize Chinese characters as identification objects with high accuracy.

In this way, the Chinese character recognition system and the Chinese character recognition method according to the present invention are difficult to recognize due to the passage of a long time, or the accuracy of the Chinese character recognition is improved even if the high similarity between various handwriting or sizes and characters, and the quality of the Chinese character records are deteriorated. it increases Therefore, there is an advantage of contributing to recognizing the Chinese characters of traditional records that were difficult to recognize in the past. In addition, there is an advantage that contributes to the improvement of the efficiency of archival records that digitize and preserve traditional records.

As described above, the detailed description of the present invention has been made by the embodiments with reference to the accompanying drawings, but since the above-described embodiments have only been described with reference to the preferred embodiments of the present invention, the present invention It should not be construed as being limited only to the embodiments, and the scope of the present invention should be understood as the following claims and their equivalents.

100: Chinese character recognition system
110: server 120: feature extraction unit
121, 122, 123, 124, 125: image recognizer
130: identified character selection unit 140: output unit

Claims (10)

an input unit for receiving a Chinese character image that is an image of a Chinese character to be identified;
a feature extracting unit including a plurality of image recognizers to which a deep learning model for extracting image features from the Chinese character image received from the input unit is applied;
an identified Chinese character selection unit for deriving reliability for each of the plurality of image features received from the feature extraction unit, and selecting an identified Chinese character corresponding to the Chinese character to be identified; and
an output unit for receiving and outputting the identified Chinese character selected by the identified Chinese character selecting unit from the identified Chinese character selecting unit; includes,
The image recognizer,
Learning was made through at least one of the sensitivity and robustness regularization methods,
The sensitivity normalization method is to distinguish between-class images, which are different Chinese character images,
The robustness normalization method is to distinguish the same Chinese character image, without-class (within-class) image,
The Chinese character recognition system that the identified Chinese character selection unit collects the results extracted from the image recognizer by an ensemble method.
The method of claim 1,
The Chinese character image transmitted from the input unit to the feature extraction unit has the form of a multi-channel image matrix,
A Chinese character recognition system, characterized in that the plurality of image recognizers that have received the Chinese character image having the form of a multi-channel image matrix extract image features each having the form of a row vector.
delete delete delete The method of claim 1,
The identified character selection unit,
A Chinese character recognition system, characterized in that for each of the plurality of image features received from the feature extraction unit, a reliability matrix composed of reliability scores is calculated by using an activation function.
7. The method of claim 6,
The identified character selection unit,
Selecting a maximum value in each column of the calculated reliability matrix to generate a confidence score-row vector,
A Chinese character recognition system, characterized in that, in the generated confidence score-row vector, a Chinese character corresponding to a maximum reliability score having the highest reliability score is selected as the identified Chinese character corresponding to the Chinese character to be identified.
An input step of receiving a Chinese character image that is an image of a Chinese character to be identified by the Chinese character recognition system;
an extraction step of extracting a plurality of image features from the Chinese character image received in the input step;
a reliability derivation step of deriving a reliability score from the plurality of image features extracted in the extraction step; and
An identified Chinese character selection step of selecting an identified Chinese character corresponding to the Chinese character to be identified from a plurality of the reliability scores derived in the reliability deriving step;
An extraction step of extracting a plurality of image features from the Chinese character image input in the input step;
It is performed by a plurality of image recognizers to which a deep learning model for extracting image features from the Chinese character image is applied,
The image recognizer,
Learning was made through at least one of the sensitivity and robustness regularization methods,
The sensitivity normalization method is to distinguish between-class images, which are different Chinese character images,
The robustness normalization method is to distinguish the same Chinese character image, without-class (within-class) image,
Reliability derivation step of deriving a reliability score from the plurality of image features extracted in the extraction step;
A Chinese character recognition method using a Chinese character recognition system, comprising the step of collecting the results extracted from the image recognizer by an ensemble method.
9. The method of claim 8,
As a step made after the identified Chinese character selection step,
The Chinese character recognition method using the Chinese character recognition system, characterized in that it further comprises; an output step of outputting the identified Chinese character selected in the identified Chinese character selection step so that it can be recognized from the outside.
10. The method of claim 9,
The selection of the identified Chinese character made in the identified Chinese character selection step is
A Chinese character recognition method using a Chinese character recognition system, characterized in that it is made by selecting a Chinese character corresponding to the maximum reliability score, which is the largest value among the plurality of reliability scores, as the identified Chinese character.

Images