WO2021212658A1

WO2021212658A1 - Ocr image sample generation method and apparatus, print font verification method and apparatus, and device and medium

Info

Publication number: WO2021212658A1
Application number: PCT/CN2020/099064
Authority: WO
Inventors: 陈伟杰
Original assignee: 平安国际智慧城市科技股份有限公司
Priority date: 2020-04-24
Filing date: 2020-06-30
Publication date: 2021-10-28
Also published as: CN111626124A

Abstract

Disclosed are an OCR image sample generation method and apparatus, a print font verification method and apparatus, a device and a medium, which relate to artificial intelligence. The method comprises: receiving an image generation instruction, and acquiring an image sample; inputting the image sample into a preset font typesetting generation model; acquiring first annotation information by means of performing text detection and character recognition on the image sample, and obtaining a simulation result generated by means of the reconstruction of the font typesetting generation model; inputting the image sample and a simulated image into a preset style compositing model, such that the style compositing model extracts style features and content features, and generates a composite result; and acquiring an OCR image sample label and also recording a composite image as an OCR image sample corresponding to the image sample, and associating the OCR image sample with the OCR image sample label. By means of the method, an OCR image sample with the same texture style as an image sample is automatically generated, and automatic annotation with a sample label is realized.

Description

OCR image sample generation, print verification methods, devices, equipment and media

This application requires the priority of a Chinese patent application filed with the Chinese Patent Office with the application number CN202010333257.4 and titled "OCR image sample generation, printed matter verification method, device, equipment and medium" on April 24, 2020, which The entire content is incorporated into this application by reference.

Technical field

This application relates to the field of artificial intelligence data modeling, and in particular to an OCR image sample generation and printed verification method, device, computer equipment and storage medium.

Background technique

At present, with the development of science and technology in society and the growing digital age, the application of text recognition based on OCR technology has been widely used; OCR (Optical Character Recognition), Chinese called optical character recognition, is used to make use of optical technology and computer technology. The text written on or on paper is read and converted into a format that the computer can accept and understand. In the prior art, more and more application scenarios (such as: application scenarios involving finance, insurance, and smart security) need to identify text information in printed documents for verification. A very large number of document samples (and manual labeling of sample labels) are needed to train the neural network, and usually only a very small number of document samples can be obtained. It is difficult to obtain such a large document sample, and due to its unique interference (For example, in the printing process, the font is likely to become broken or ink sticking), making OCR recognition still difficult, resulting in low accuracy and precision of the trained document recognition model, resulting in a high verification error rate, requiring manual re-verification. Great waste of cost and low efficiency.

Summary of the invention

This application provides an OCR image sample generation, print verification method, device, computer equipment and storage medium, which realizes the automatic generation of OCR image samples with the same texture style as the image samples, and automatically annotates the OCR image sample tags, reducing labor Cost and time, and can improve the accuracy and reliability of OCR recognition results and print verification.

An OCR image sample generation method, including:

Receiving an image generation instruction to obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

Input the image sample into a preset font typesetting generation model, by performing text detection and character recognition on the image sample, the font typesetting generation model recognizes the first annotation information of the image sample, and obtains the The font typesetting generation model reconstructs and generates a simulation result according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image, a second font label, and a second typesetting label And the second label information;

The image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model performs simulation on the simulation based on the style features and the content features. Image style transfer and synthesis are performed to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

Mark the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time record the composite image as corresponding to the image sample And associate the OCR image sample with the OCR image sample tag.

A print verification method, including:

Receive certificate verification instructions, obtain the printed version and verification information of the pending certificate;

Input the printed document to be trained into the document recognition model that has been trained; the document recognition model is trained through the OCR image sample generated by the above-mentioned OCR image sample generation method;

Perform OCR recognition on the printed document to be documented through the document recognition model, and obtain an OCR recognition result output by the document recognition model, where the OCR recognition result includes the document-related text information in the printed document to be documented;

Comparing the OCR recognition result with the verification information to determine whether the printed document to be certified meets the verification information;

If the printed version of the certificate to be issued meets the verification information, it is determined that the verification is passed;

If the printed body of the document to be issued does not conform to the verification information, confirm that the verification is not passed, and prompt on the display interface.

An OCR image sample generating device, including:

A receiving module, configured to receive an image generation instruction and obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

The input module is used to input the image sample into a preset font typesetting generation model, and by performing text detection and character recognition on the image sample, to obtain the first annotation information of the image sample recognized by the font typesetting generation model , And obtain a simulation result generated by reconstruction of the font typesetting generation model according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image and a second font label , The second typesetting label and the second annotation information;

The synthesis module is configured to input the image sample and the simulated image into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model is based on the style features and the content The feature performs style transfer and synthesis on the simulated image to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

A generating module for marking the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time recording the composite image as The OCR image sample corresponding to the image sample, and the OCR image sample is associated with the OCR image sample tag.

A printed body verification device includes:

The acquisition module is used to receive the certificate verification instruction, and obtain the printed version and verification information of the certificate;

The training module is used to input the printed document to be into the trained document recognition model; the document recognition model is trained through the OCR image sample generated by the above-mentioned OCR image sample generation method;

The recognition module is configured to perform OCR recognition on the printed document to be documented through the document recognition model, and obtain the OCR recognition result output by the document recognition model. The OCR recognition result includes the document-related printed document to be documented Text information;

The comparison module is configured to compare the OCR recognition result with the verification information, and determine whether the printed document to be document meets the verification information;

The first determining module is configured to determine that the verification is passed if the printed document to be document meets the verification information;

The second determination module is configured to confirm that the verification is not passed if the printed document to be document does not meet the verification information, and prompt on the display interface.

A computer device includes a memory, a processor, and a computer program that is stored in the memory and can run on the processor, and the processor implements the following steps when the processor executes the computer program:

Perform OCR recognition on the printed document to be documented by the document recognition model, and obtain an OCR recognition result output by the document recognition model, where the OCR recognition result includes the document-related text information in the printed document to be documented;

A computer-readable storage medium having a computer program stored on the computer-readable storage medium, wherein, when the computer program is executed by a processor, the following steps are implemented:

The OCR image sample generation method, device, computer equipment, and storage medium provided in this application obtain image samples containing the first font label, the first typesetting label, and the first texture style label; and input it into the font typesetting generation model. Perform text detection and text recognition on the image sample to obtain the recognized first annotation information; reconstruct according to the first font label, the first typesetting label, and the first annotation information to generate a simulation result; The image sample and the simulated image are input into a style synthesis model, and the style synthesis model performs style transfer and synthesis on the simulated image according to the extracted style feature and the content feature to generate a synthesis result; the synthesis The result includes the synthesized image and the first texture style label; the second font label, the second typesetting label, the second annotation information and the first texture style label are marked as OCR image sample labels, and all The composite image is recorded as an OCR image sample corresponding to the image sample, and the OCR image sample is associated with the OCR image sample tag. Therefore, the present application realizes the automatic generation of an OCR image with the same texture style as the image sample Samples, and accurately mark the OCR image sample labels on the OCR image samples, reducing the labor cost and time of collecting image samples, and can quickly obtain OCR image samples in the desired scene, which is more targeted and is a follow-up model The training improves the accuracy and reliability, reduces the labor cost of labeling the OCR image sample labels, avoids errors in manual labeling, and improves labeling accuracy.

The printed matter verification method, device, computer equipment and storage medium provided in this application obtain the printed matter to be document and verification information by receiving the document verification instruction; input the printed matter to be documented into the OCR generated by the above-mentioned OCR image sample generation method A certificate recognition model trained and completed by image samples; OCR recognition is performed on the printed document to be issued through the certificate recognition model, and the OCR recognition result output by the certificate recognition model is obtained; and the OCR recognition result is compared with the verification Information is compared to determine whether the printed document to be documented meets the verification information; if the printed document to be documented meets the verification information, it is determined that the verification is passed; if the printed document to be documented does not comply with the verification information, It is confirmed that the verification is not passed, and a prompt is displayed on the display interface. Therefore, this application uses the OCR image sample generated by the OCR image sample generation method to train a certificate recognition model for a specific scene, and performs automatic recognition and automatic verification through the certificate recognition model. The automatic verification of printed documents for specific scenarios is realized, which is highly targeted, improves recognition accuracy, improves recognition efficiency and reliability, improves user experience, and saves labor costs.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present application more clearly, the following will briefly introduce the drawings that need to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative labor.

FIG. 1 is a schematic diagram of an application environment of an OCR image sample generation method or a printed body verification method in an embodiment of the present application;

Fig. 2 is a flowchart of a method for generating OCR image samples in an embodiment of the present application;

3 is a flowchart of step S20 of the OCR image sample generation method in an embodiment of the present application;

4 is a flowchart of step S20 of the OCR image sample generation method in another embodiment of the present application;

5 is a flowchart of step S30 of the OCR image sample generation method in an embodiment of the present application;

FIG. 6 is a flowchart of a printed matter verification method in an embodiment of the present application;

FIG. 7 is a flowchart of step S200 of the printed body verification method in an embodiment of the present application;

Fig. 8 is a functional block diagram of an OCR image sample generating device in an embodiment of the present application;

FIG. 9 is a schematic block diagram of a printed matter verification device in an embodiment of the present application;

Fig. 10 is a schematic diagram of a computer device in an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

The OCR image sample generation method provided by this application can be applied in the application environment as shown in Fig. 1, in which the client (computer equipment) communicates with the server through the network. Among them, the client (computer equipment) includes, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, cameras, and portable wearable devices. The server can be implemented as an independent server or a server cluster composed of multiple servers.

In an embodiment, as shown in FIG. 2, a method for training a recognition model is provided, and the technical solution mainly includes the following steps S10-S40:

S10: Receive an image generation instruction, and obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label.

Understandably, after receiving the image generation instruction, the image generation instruction is a request to be triggered after selecting and confirming the image sample that needs to be generated. The trigger mode can be set according to requirements, for example, in the application The platform interface provides a trigger button that can be triggered by clicking, sliding, etc., to acquire the image sample, and the acquisition method can be set as required. For example, the acquisition method can be the The image sample acquires the image sample, acquires the image sample according to the storage path of the image sample included in the image generation instruction, and so on.

Wherein, the image sample is an image that contains printed matter and is related to a certificate, that is, an image file obtained after a certificate is copied. The image sample can be a photo file or a scanned image file. The image sample label is associated, the image sample label is a label assigned to annotate the content in the image sample, and the image sample label includes a first font label, a first typesetting label, and a first texture style label, The first font label is a label of information such as font style and font size corresponding to the text content in the image sample, such as Song Ti No. 5, Li Shu Xiao 4, etc., and the first typesetting label is in the image sample The text content corresponding to the label of the typesetting information, such as single-column arrangement, double-column equal-width arrangement, double-column unequal-width arrangement, etc., the first texture style label is transferred during scanning, copying, etc. The texture information label formed by the image sample, such as wrinkle, deformation, etc.

S20. Input the image sample into a preset font typesetting generation model, by performing text detection and character recognition on the image sample, obtain the first annotation information of the image sample that the font typesetting generation model recognizes, and obtain The font typesetting generation model reconstructs and generates a simulation result according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image, a second font label, and a second Typesetting labels and second annotation information.

Understandably, the font typesetting generation model includes performing text detection and text recognition processing on the input image sample to obtain the first annotation information related to the text in the image sample, and the text detection is The image sample is scanned to detect the text area containing the text and the area coordinates of the text area. The scanning detection method can be set according to requirements, such as extracting text features to determine the text area of the text, and identifying multiple text areas through edge detection. The text boundary determines the text area of the text, etc. The text area is a quadrilateral area containing text content, the area coordinates are the coordinates of the four points of the text area in the image sample, and the area coordinates include four A coordinate value (the coordinate value includes the abscissa and the ordinate), the text recognition is the recognition of each text area in the image sample to obtain the text content in each text area, and the The text area, the area coordinates associated with the text area, and the text content of the text area are recorded as the first information of the image sample, and all the first information is marked as the first annotation information .

Wherein, the font typesetting generation model further includes a reconstruction model, and the reconstruction model performs reconstruction processing according to the first font label, the first typesetting label, and the first annotation information to generate a simulated image, The simulation result of the second font label, the second typesetting label and the second annotation information. The reconstruction model may be a neural network model set according to requirements. Preferably, the reconstruction model is based on GAN (Generative Adversarial Network , Generate a neural network model trained by a confrontation network) model, and the simulated image is an image file containing text content corresponding to the second font label, the second typesetting label, and the second annotation information, so The second font label may be consistent with the first font label, or may be the font label with the highest approximate value to the first font label, and the font label is the font style and font size of all font styles, For example, Song Ti No. 5, Kai Ti No. 6 and so on, the highest approximate value is that in the font label, in addition to the first font label, the approximate value of the font style of the first font label plus the approximate value of the first font label The highest value obtained by the approximate font size. The approximate font style value is a measurement index value that differs from one of the font styles in all font styles. The more similar the font style, the larger the measurement index value. The range of the metric value can be set from 0 to 100. For example, the font style of the first font label is "Song Ti", and the font style of "Imitated Song" is the most similar to "Song Ti" (you can use the reconstruction model based on GAN model Encode the font style of the first font label, and then decode the encoded code, which can also be defined by a preset rule for the difference between the living body printing and the calligraphy font), and the measurement index value is 100; The approximate value of the font size is the size measurement index value of the difference between each font size and one of the font sizes. The closer the font size is, the larger the size measurement index value. For example, the font size of the first font label is "number five". The font size of “11 points” is the most similar to “No. 5” (the font style of the first font label can be encoded through the reconstruction model based on the GAN model, and then the encoded code can be decoded, or through The difference between “No. 5” and “No. 5” is the smallest. When the difference is negative, take its absolute value plus one to obtain the difference), and its measurement index value is 100; the second typesetting label may be A typesetting label is consistent, or it may be the typesetting label with the highest approximation value to the first typesetting label. The approximate value is the typesetting value corresponding to a measure of the difference between each typesetting label and one of the typesetting labels. The closer the typesetting label is, The larger the typesetting value, for example: the first typesetting label is double-column equal-width arrangement, the typesetting value between three-column equal-width arrangement and double-column equal-width arrangement is 1, and double-column unequal-width arrangement and double The typesetting value between the columns with equal width is 0 (the most similar); A label information is encoded, random noise is introduced to the encoded first label information and decoded to generate information close to the first label information, for example: the first label information is "I like juice" and its coordinates , The second label information is "I like fruits" and the corresponding coordinates.

In one embodiment, as shown in FIG. 3, in the step S20, the image sample is input into a preset font typesetting generation model, and the image sample is subjected to text detection and text recognition to obtain all the images. The font typesetting generation model that recognizes the first annotation information of the image sample includes:

S201: Perform text detection on the image sample through the font typesetting generation model, and at the same time extract the text features of the image sample, and obtain the image recognized by the font typesetting generation model according to the extracted text features The regional result of the sample; the regional result includes a number of text regions containing text and region coordinates associated with each of the text regions.

Understandably, the text feature extraction of the image sample is performed on the image sample through the font typesetting generation model, and the text features are features such as Chinese words and sentences, letter words, etc., and the font typesetting generation model is based on the The text features are recognized, and a number of text regions containing text content in the image sample and the region coordinates associated with each of the text regions are identified. The text region is a quadrilateral region containing text content, and the region coordinates are Is the coordinates of the four points of the text area in the image sample, the area coordinates include four coordinate values (the coordinate values include the abscissa and the ordinate), which associate all the text areas and the text areas The area coordinates are determined as the area result.

S202. Extract the character feature of each text area through the font typesetting generation model, and obtain each of the character features identified by the font typesetting generation model according to the extracted character feature of each text area. The text content of the text area.

Understandably, the character feature extraction is performed on each of the text regions through the font typesetting generation model, and the character features are Chinese characters, letters, etc., according to the extracted features of each text region Character features, the font typesetting generation model outputs the text content of each text area, and obtains all the text content.

S203. Record the text area, the area coordinates associated with the text area, and the text content of the text area as the first information of the image sample, and mark all the first information as all Describe the first label information.

Understandably, the text area has a one-to-one correspondence with the area coordinates and a one-to-one correspondence with the text content, and the text area, the corresponding area coordinates and the corresponding text content are combined together Marked as the first information of the image sample, the image sample may contain several pieces of the first information, and then all the first information is marked as the first annotation information corresponding to the image sample .

In this way, by performing text detection and text recognition on the image sample through the font typesetting generation model, the first annotation information in the image sample can be accurately obtained.

In one embodiment, as shown in FIG. 4, in the step S20, the acquiring the font typesetting generation model is performed according to the first font label, the first typesetting label, and the first annotation information. The simulation results generated by reconstruction include:

S204. Input the first font label, the first typesetting label, and the first annotation information into a reconstruction model in the font typesetting generation model; the reconstruction model is completed based on a GAN model.

Understandably, the GAN model (GAN, Generative Adversarial Networks) is a deep neural network model of a generative confrontation network, and the reconstruction model is a deep neural network model trained and trained based on the GAN model. The first font label, the first typesetting label and the first annotation information are input into the reconstruction model.

S205: Perform combined reconstruction through a generator in the reconstruction model, and obtain the simulated image, the second font label, the second typesetting label, and the second annotation information output by the reconstruction model.

Understandably, the first font label, the first typesetting label, and the first annotation information are combined into a transition image with the same size as the image sample, and the transition image is preferably a blank background, Its content includes the first font label, the first typesetting label, and the first annotation information; the reconstruction model includes the generator, and the main task of the generator is to learn the transition image through The transition image is encoded, and the encoded code is decoded to generate an image file that is very close to the transition image. The reconstruction model also includes a discriminator whose main task is to The image file generated by the generator is differentiated from the transition image, and the authenticity is judged. In the iterative training process of the reconstruction model, the generator continuously strives to make the generated image file closer and closer. The transition image, and the discriminator is constantly trying to identify the authenticity of the image file. Through the game between the generator and the discriminator, with repeated iterations, the generator and the discriminator are finally The generator has reached a balance. The generator generates the analog image close to the transition image, and the discriminator has difficulty identifying the difference between the analog image and the image sample. Thus, the The reconstruction model reconstructs the simulated image.

The simulated image is an image file containing text content corresponding to the second font label, the second typesetting label, and the second annotation information. Preferably, the size of the simulated image may be the same as the size of the image sample. The size of the second font label is the same as that of the first font label, for example: the first font label is Times New Roman No. 5, the second font label is imitation Song 11 points; the second typesetting label is similar to the first font label. One typesetting label is close, for example: the first typesetting label is two-column equal-width arrangement, and the second typesetting label is two-column unequal-width arrangement; The label information is encoded, random noise is introduced to the encoded first label information and decoded to generate information close to the first label information, for example: the first label information is "I like juice" and its coordinates, The second label information is "I like fruits" and the corresponding coordinates.

S206. Record the simulated image, the second font label, the second typesetting label, and the second annotation information as a simulation result output by the font typesetting generation model.

Understandably, the simulation result includes the simulated image, the second font label, the second typesetting label, and the second annotation information, and the simulated image, the second font label, the The second typesetting label and the second annotation information are associated with each other.

In this way, through the reconstruction model trained on the basis of the GAN model in the font typesetting generation model, the simulation results including the simulated image, the second font label, the second typesetting label and the second annotation information are reconstructed and generated, thereby realizing automatic Generate simulation images associated with the second font label, second typesetting label and second annotation information, that is, automatically annotate the simulated image, thereby automatically generating a simulated image with the first font label and the first typesetting label related to the image sample .

S30. Input the image sample and the simulated image into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model compares the style features and the content features with each other. The simulated image undergoes style transfer and synthesis to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label.

Understandably, the image sample and the simulated image are input into the style synthesis model, and the style synthesis model extracts style features and content features, and the style features are features such as wrinkles, background gray levels, and stripes. , The content feature is a feature related to the first font label, the first typesetting label, the second font label, and the second typesetting label. The simulated image performs style transfer, that is, the simulated image is used as the initial image, all the pixel values of the initial image are obtained, and the total loss value is obtained according to the style feature and the content feature, and iteratively updated continuously through gradient descent For all the pixel values, until the total loss value reaches a preset condition, the preset condition can be set according to demand. For example, the preset condition can be set to no longer decrease, and the updated The initial image is determined to be a composite image, and the composite image is an image of the simulated image that is optimized by transferring the texture information in the image sample, and the texture information is the process of transferring the texture information into an image file during operations such as scanning and copying. Therefore, the texture information provided in the composite image is consistent with the first texture style tag, and the composite image is associated with the first texture style tag to associate the composite image with the first texture style tag. A texture style label is determined as the synthesis result.

In one embodiment, as shown in FIG. 5, in the step S30, the image sample and the simulated image are input into a preset style synthesis model, and the style synthesis model extracts style features and content features, so The style synthesis model performs style transfer and synthesis on the simulated image according to the style feature and the content feature to generate a synthesis result, including:

S301, using the simulated image as an initial image, and acquiring all pixel values of the initial image;

Understandably, the simulated image is used as the initial image, that is, when the image sample and the simulated image are input into a preset style synthesis model, the initial image is consistent with the simulated image, and the initial image It contains a number of pixels, each of the pixels corresponds to one of the pixel values, and the pixel value is the value assigned by the pixel by measuring the color, and the range of the pixel value can be set according to requirements .

S302: Extract the style feature of the image sample and the style feature of the initial image through the style synthesis model, and calculate a style loss value according to the style feature of the image sample and the style feature of the initial image.

Understandably, the style synthesis model is a deep neural network model obtained by transfer learning, and the network structure of the style synthesis model is obtained by transfer learning, such as the network structure of the transfer learning VGG19 model, and the style feature is Features such as wrinkles, background gray levels, spots, etc., where the style loss value is calculated by calculating the style feature of the image sample and the style feature of the initial image through a style loss function to obtain the image sample and the initial image sample The style difference.

S303: Extract the content feature of the image sample and the content feature of the initial image through the style synthesis model, and calculate a content loss value based on the content feature of the image sample and the content feature of the initial image.

Understandably, the content feature is a feature related to the first font label, the first typesetting label, the second font label, and the second typesetting label, and the content loss value is a loss through content The function calculates the content feature of the image sample and the content feature of the initial image to obtain the content difference between the image sample and the initial image sample.

S304: Perform weighting processing on the style loss value and the content loss value to obtain a total loss value.

Understandably, the weighting process is to input the style loss value and the content loss value into a loss weighting function, and the total loss value is calculated by the loss weighting function, and the loss weighting function is:

L=w ₁ ×L ₁ +w ₂ ×L ₂

in:

L ₁ is the style loss value;

L ₂ is the content loss value;

w1 is the weight of the loss function of the style loss value;

w ₂ is the weight of the loss function of the content loss value;

L is the total loss value.

S305: Perform gradient descent using the L-BFGS algorithm, and when the total loss value does not reach a preset condition, iteratively update all the pixel values in the initial image until the total loss value reaches the preset condition At this time, the updated initial image is determined to be a composite image.

Understandably, gradient descent is performed through the L-BFGS algorithm, that is, the total loss value is continuously reduced through the L-BFGS algorithm. The L-BFGS algorithm is a method for solving unconstrained nonlinear problems, and the preset conditions It can be set according to requirements. For example, the preset condition can be set to no longer decrease, and when the total loss value does not reach the preset condition (no more decrease), all the items in the initial image are updated iteratively. For the pixel value, until the total loss value reaches the preset condition (no more drop), the updated initial image is determined to be a composite image.

S306: Associate the first texture style label with the synthesized image, and determine the synthesized image and the first texture style label as the synthesis result.

Understandably, the synthesized image is associated with the first texture style tag, so that the synthesized image and the first texture style tag are determined as the synthesis result.

In this way, the style transfer of the simulated image from the image sample is carried out through the style synthesis model, and the synthesized image with the first texture style label associated with the image sample is automatically generated, and the synthesized image is automatically generated from the image sample, which is the training of the subsequent model. Provide effective samples, shorten the time for collecting samples, reduce the cost of collecting samples, and improve efficiency.

S40. Mark the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time record the composite image as the same as the image An OCR image sample corresponding to the sample, and associate the OCR image sample with the OCR image sample label.

Understandably, the OCR image sample is annotated with OCR image sample tags, that is, the second font tags, the second typesetting tags, the second annotation information, and the first texture style tags are marked as OCR images The OCR image sample label corresponding to the sample thus reduces the labor cost of labeling the OCR image sample label, avoids errors in manual labeling, and improves labeling accuracy.

This application obtains an image sample by receiving an image generation instruction; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label; and the image sample is input The preset font typesetting generation model uses text detection and text recognition on the image sample to obtain the font typesetting generation model to recognize the first annotation information of the image sample, and obtain the font typesetting generation model according to the A simulation result generated by reconstructing the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image, a second font label, a second typesetting label, and second annotation information; The image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model performs simulation on the simulation based on the style features and the content features. The image undergoes style transfer and synthesis to generate a synthesis result; the synthesis result includes a synthesized image and a first texture style label; the second font label, the second typesetting label, the second annotation information and the first A texture style label is marked as an OCR image sample label, and the composite image is recorded as an OCR image sample corresponding to the image sample, and the OCR image sample is associated with the OCR image sample label. This application can be applied in the field of smart security, so as to promote the construction of smart cities.

Therefore, this application realizes that by obtaining an image sample containing the first font label, the first typesetting label, and the first texture style label; inputting it into the font typesetting generation model, and performing text detection and text recognition on the image sample to obtain The identified first annotation information; reconstruct according to the first font label, the first typesetting label, and the first annotation information to generate a simulation result; synthesize the image sample and the simulated image input style Model, the style synthesis model performs style transfer and synthesis on the simulated image according to the extracted style feature and the content feature to generate a synthesis result; the synthesis result includes the synthesis image and the first texture style label; The second font label, the second typesetting label, the second annotation information, and the first texture style label are marked as OCR image sample labels, and the composite image is recorded as corresponding to the image sample OCR image samples, and associate the OCR image samples with the OCR image sample tags. Therefore, the present application realizes the automatic generation of OCR image samples with the same texture style as the image samples, and accurately performs operations on the OCR image samples. Annotating OCR image sample labels reduces the labor cost and time of collecting image samples, and can quickly obtain OCR image samples in the required scene, which is more targeted, and improves the accuracy and reliability of subsequent model training, and The labor cost for labeling the OCR image sample label is reduced, errors caused by manual labeling are avoided, and labeling accuracy is improved.

The image recognition method provided in this application can be applied in the application environment as shown in Fig. 1, in which the client (computer equipment) communicates with the server through the network. Among them, the client (computer equipment) includes, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, cameras, and portable wearable devices. The server can be implemented as an independent server or a server cluster composed of multiple servers.

In an embodiment, as shown in FIG. 6, a printed matter verification method is provided, and the technical solution mainly includes the following steps S100-S600:

S100: Receive a certificate verification instruction, and obtain a printed version of a certificate to be issued and verification information.

Understandably, the document verification instruction is a request triggered after selecting and confirming the printed document to be verified and the verification information that need to be verified, and the printed document to be verified is the document that needs to be verified after being scanned or copied. An image file, the verification information is a verification carrier that provides the printed document to be verified, the verification information can be obtained from the information related to the printed document to be entered by the user at the client, or from a database It is obtained by querying information related to the printed document to be documented, such as the entered bank card number or bank name.

S200: Input the printed document to be into a trained document recognition model; the document recognition model is trained through the OCR image sample generated by the above-mentioned OCR image sample generation method.

Understandably, the printed document to be document is input to the document recognition model, and the document recognition model is an OCR image sample generated by the above-mentioned OCR image sample generation method and a neural network model trained on the image sample.

S300. Perform OCR recognition on the printed document to be documented through the document recognition model, and obtain an OCR recognition result output by the document recognition model, where the OCR recognition result includes document-related text information in the printed document to be documented .

Understandably, the OCR recognition is to read the text of the printed document to be documented through OCR (Optical Character Recognition) technology, and the OCR recognition result includes that the printed document to be documented is related to the document Text information.

S400: Compare the OCR recognition result with the verification information, and determine whether the printed document to be certified meets the verification information.

Understandably, the OCR recognition result is checked with the verification information to determine whether the printed document to be checked passes.

S500: If the printed document to be document meets the verification information, it is determined that the verification is passed.

Understandably, if the OCR recognition result corresponding to the printed document to be certified is consistent with the verification information, it is determined that the printed document to be certified is verified.

S600: If the printed document to be document does not meet the verification information, confirm that the verification fails, and prompt on the display interface.

Understandably, if the OCR recognition result corresponding to the printed document to be documented is inconsistent with the verification information, it is determined that the printed document to be documented has failed the verification, and a prompt is displayed on the display interface that the display interface It is the display interface of the terminal device corresponding to the customer. The content of the prompt can be set according to the needs. For example, the content of the prompt is "The verification information is wrong, please re-enter the verification information!".

In one embodiment, as shown in FIG. 7, before step S200, that is, before inputting the printed document to be into the trained document recognition model, the method includes:

S2001: Obtain a certificate sample set; the certificate sample set includes a number of certificate samples, and one certificate sample is associated with a sample label; when the certificate sample is an image sample, the sample label is an image sample label; in the certificate sample When it is an OCR image sample, the sample label is an OCR image sample label; the OCR image sample is generated by the above-mentioned OCR image sample generation method; the number of the image samples in the certificate sample set is less than the number of the OCR image samples.

Preferably, the OCR image sample is an image generated by the above-mentioned OCR image sample generation method through the image samples in the certificate sample set, and the OCR image sample has been annotated by the OCR image sample generation method. OCR image sample label, one image sample can generate multiple OCR image samples by the OCR image sample generation method, and the number of the image samples in the certificate sample set is less than the number of the OCR image samples, thus saving The time for collecting the document sample set and the manual time for labeling the document sample, and the OCR image sample label can be accurately labelled on the OCR image sample.

S2002: Input the document sample set into a deep learning OCR model containing initial parameters.

Understandably, the initial parameters can be set according to requirements, for example, the initial parameters are randomly assigned parameter values, or the initial parameters are preset parameter values, and so on.

S2003: Perform OCR recognition on the credential sample through the deep learning OCR model, and obtain a training recognition result of the credential sample output by the deep learning OCR model.

Understandably, the OCR recognition is to read the printed text of the to-be-documented document through OCR (Optical Character Recognition) technology, and the training recognition result includes the document-related text in the document sample information.

S2004: Match the training recognition result with the sample label to obtain the loss value of the certificate sample.

Understandably, the training recognition result and the sample label are input into the loss function in the deep learning OCR model, and the loss value of the certificate sample is calculated by the loss function, and the loss value The difference between the training recognition result and the sample label is indicated, and the loss value is getting smaller and smaller, indicating that the training recognition result is getting closer and closer to the sample label.

S2005: When the loss value reaches a preset convergence condition, record the deep learning OCR model after convergence as a certificate recognition model completed by training.

Understandably, when the loss value reaches the preset convergence condition, it indicates that the loss value has reached the optimal result, that is, the training recognition result is already very close to the sample label, and the deep learning The OCR model has converged, and the deep learning OCR model after convergence is recorded as a certificate recognition model completed by training.

In this way, according to the sample labels of the certificate samples in the certificate sample set, the trained certificate recognition model obtained through continuous training can improve the accuracy and reliability of the OCR recognition result.

In one embodiment, after step S2004, that is, after matching the training recognition result with the sample label, and obtaining the loss value of the certificate sample, the method further includes:

S2006: When the loss value does not reach a preset convergence condition, iteratively update the initial parameters of the deep learning OCR model, until the loss value reaches the preset convergence condition, converge the The OCR model of deep learning is recorded as a certificate recognition model completed by training.

Wherein, the convergence condition may be a condition that the value of the loss value is small and will not drop after 8000 calculations, that is, the value of the loss value is small and will not drop anymore after 8000 calculations Stop training, and record the deep learning OCR model after convergence as the certificate recognition model after training; the convergence condition can also be the condition that the loss value is less than the set threshold, that is, when the loss value When it is less than the set threshold, the training is stopped, and the OCR model of the deep learning after convergence is recorded as the certificate recognition model after the training.

In this way, when the loss value does not reach the preset convergence condition, the initial parameters of the iterative neural network model are continuously updated to continuously move closer to the accurate recognition result, so that the accuracy of the recognition result becomes higher and higher.

The present application inputs the to-be-detected image to the trained remake recognition model, and outputs the recognition result of the to-be-detected image. In this way, the present application realizes the rapid and accurate recognition of the re-taken image, which improves the recognition. The accuracy and hit rate of, improve the efficiency and reliability of recognition, and save costs.

In one embodiment, an OCR image sample generating device is provided, and the OCR image sample generating device corresponds to the OCR image sample generating method in the above-mentioned embodiment in a one-to-one correspondence. As shown in FIG. 8, the OCR image sample generating device includes a receiving module 11, an input module 12, a synthesis module 13 and a generating module 14. The detailed description of each functional module is as follows:

The receiving module 11 is configured to receive an image generation instruction and obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

The input module 12 is configured to input the image sample into a preset font typesetting generation model, and by performing text detection and character recognition on the image sample, the font typesetting generation model recognizes the first annotation of the image sample Information, and obtain a simulation result generated by reconstruction of the font typesetting generation model according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image and a second font Label, second typesetting label and second annotation information;

The synthesis module 13 is configured to input the image sample and the simulated image into a preset style synthesis model, the style synthesis model extracts style features and content characteristics, and the style synthesis model is based on the style features and the The content feature performs style transfer and synthesis on the simulated image to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

The generating module 14 is configured to mark the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time record the composite image as An OCR image sample corresponding to the image sample, and associate the OCR image sample with the OCR image sample tag.

In an embodiment, the input module 12 includes:

The first extraction unit is configured to perform text detection on the image sample through the font typesetting generation model, and at the same time extract the text features of the image sample, and obtain the font typesetting generation model to recognize the extracted text features The area result of the image sample obtained; the area result includes a number of text areas containing text and the area coordinates associated with each of the text areas;

The second extraction unit is configured to extract the character features of each text area through the font typesetting generation model, and obtain the font typesetting generation model to recognize the extracted text features of each text area The text content of each of the text areas;

The marking unit is configured to record the text area, the area coordinates associated with the text area, and the text content of the text area as the first information of the image sample, and to record all the first information Mark as the first marking information.

In an embodiment, the input module 12 further includes:

An input unit for inputting the first font label, the first typesetting label, and the first annotation information into a reconstruction model in the font typesetting generation model; the reconstruction model is trained based on a GAN model Finish;

A reconstruction unit, configured to perform combined reconstruction through a generator in the reconstruction model to obtain the simulated image, the second font label, the second typesetting label, and the second annotation information output by the reconstruction model;

The output unit is configured to record the simulated image, the second font label, the second typesetting label, and the second annotation information as a simulation result output by the font typesetting generation model.

In an embodiment, the synthesis module 13 includes:

An acquiring unit, configured to use the simulated image as an initial image and acquire all pixel values of the initial image;

The first calculation unit is configured to extract the style feature of the image sample and the style feature of the initial image through the style synthesis model, and calculate it according to the style feature of the image sample and the style feature of the initial image Out the style loss value;

The second calculation unit is configured to extract the content feature of the image sample and the content feature of the initial image through the style synthesis model, and calculate it according to the content feature of the image sample and the content feature of the initial image The content loss value;

A loss unit for weighting the style loss value and the content loss value to obtain a total loss value;

The training unit is configured to perform gradient descent using the L-BFGS algorithm, and when the total loss value does not reach a preset condition, iteratively update all the pixel values in the initial image until the total loss value reaches the When the conditions are preset, the updated initial image is determined to be a composite image;

The associating unit is configured to associate the first texture style label with the synthesized image, and determine the synthesized image and the first texture style label as the synthesis result.

For the specific limitation of the OCR image sample generating device, please refer to the above limitation on the OCR image sample generating method, which will not be repeated here. Each module in the above-mentioned OCR image sample generating device can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

In one embodiment, a printed matter verification device is provided, and the printed matter verification device corresponds to the printed matter verification method in the above-mentioned embodiment one-to-one. As shown in FIG. 9, the printed body verification device includes an acquisition module 101, a training module 102, an identification module 103, a comparison module 104, a first determination module 105 and a second determination module 106. The detailed description of each functional module is as follows:

The obtaining module 101 is configured to receive a certificate verification instruction, and obtain the printed form and verification information of the to-be-certified certificate;

The training module 102 is configured to input the printed document to be trained into a document recognition model that has been trained; the document recognition model is trained by an OCR image sample generated by the OCR image sample generation method according to any one of claims 1 to 4 Finish;

The recognition module 103 is configured to perform OCR recognition on the printed document to be documented through the document recognition model, and obtain the OCR recognition result output by the document recognition model. The OCR recognition result includes the printed document to be documented and the document Relevant text information;

The comparison module 104 is configured to compare the OCR recognition result with the verification information, and determine whether the printed document to be issued meets the verification information;

The first determining module 105 is configured to determine that the verification is passed if the printed document to be document meets the verification information;

The second determining module 106 is configured to confirm that the verification is not passed if the printed document to be document does not meet the verification information, and prompt on the display interface.

For the specific limitation of the printed body verification device, please refer to the above-mentioned limitation on the printed body verification method, which will not be repeated here. Each module in the above-mentioned printed matter verification device can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. 10. The computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to realize an OCR image sample generation method or a printed body verification method.

In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The processor executes the computer program to implement the OCR image sample generation method in the above embodiment. , Or the processor implements the printed body verification method in the foregoing embodiment when the computer program is executed by the processor.

In one embodiment, a computer-readable storage medium is provided. The computer-readable storage medium may be non-volatile or volatile, and a computer program is stored thereon. When the computer program is executed by a processor, The OCR image sample generation method in the foregoing embodiment is implemented, or the printed body verification method in the foregoing embodiment is implemented when the computer program is executed by a processor.

Further, the computer-readable storage medium may mainly include a storage program area and a storage data area, where the storage program area may store an operating system, an application program required by at least one function, etc.; the storage data area may store Data created by the use of nodes, etc.

A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage. In the medium, when the computer program is executed, it may include the processes of the above-mentioned method embodiments. Wherein, any reference to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and/or volatile memory. Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not a limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The blockchain referred to in this application is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm. Blockchain, essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and the generation of the next block. The blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Those skilled in the art can clearly understand that for the convenience and conciseness of description, only the division of the above functional units and modules is used as an example. In practical applications, the above functions can be allocated to different functional units and modules as required. Module completion, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, a person of ordinary skill in the art should understand that it can still implement the foregoing The technical solutions recorded in the examples are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the application, and should be included in The scope of protection of this application and so on.

Claims

An OCR image sample generation method, which includes:

Receiving an image generation instruction to obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

Input the image sample into a preset font typesetting generation model, by performing text detection and character recognition on the image sample, the font typesetting generation model recognizes the first annotation information of the image sample, and obtains the The font typesetting generation model reconstructs and generates a simulation result according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image, a second font label, and a second typesetting label And the second label information;

The image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model performs simulation on the simulation based on the style features and the content features. Image style transfer and synthesis are performed to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

Mark the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time record the composite image as corresponding to the image sample And associate the OCR image sample with the OCR image sample tag.
The OCR image sample generation method according to claim 1, wherein the image sample is input into a preset font typesetting generation model, and the font typesetting generation is obtained by performing text detection and character recognition on the image sample The model recognizes the first annotation information of the image sample, including:

Perform text detection on the image sample through the font typesetting generation model, and at the same time extract the text features of the image sample, and obtain the image sample's value of the image sample identified by the font typesetting generation model according to the extracted text features Regional results; the regional results include a number of text regions containing text and region coordinates associated with each of the text regions;

The text feature of each text area is extracted by the font typesetting generation model, and each text area recognized by the font typesetting generation model according to the extracted text feature of each text area is obtained The text content;

The text area, the area coordinates associated with the text area, and the text content of the text area are recorded as the first information of the image sample, and all the first information is marked as the first information One label information.
The OCR image sample generation method according to claim 1, wherein said acquiring said font typesetting generation model is reconstructed and generated according to said first font label, said first typesetting label and said first annotation information Simulation results, including:

Inputting the first font label, the first typesetting label, and the first annotation information into a reconstruction model in the font typesetting generation model; the reconstruction model is trained based on a GAN model;

Performing combined reconstruction through a generator in the reconstruction model to obtain the simulated image, the second font label, the second typesetting label, and the second annotation information output by the reconstruction model;

The simulation image, the second font label, the second typesetting label, and the second annotation information are recorded as a simulation result output by the font typesetting generation model.
The OCR image sample generation method of claim 1, wherein the image sample and the simulated image are input into a preset style synthesis model, and the style synthesis model extracts style features and content features, and the style synthesis The model performs style transfer and synthesis on the simulated image according to the style feature and the content feature to generate a synthesis result, including:

Using the simulated image as an initial image, and acquiring all pixel values of the initial image;

Extracting the style feature of the image sample and the style feature of the initial image through the style synthesis model, and calculating a style loss value according to the style feature of the image sample and the style feature of the initial image;

Extracting the content feature of the image sample and the content feature of the initial image through the style synthesis model, and calculating a content loss value based on the content feature of the image sample and the content feature of the initial image;

Weighting the style loss value and the content loss value to obtain a total loss value;

Perform gradient descent through the L-BFGS algorithm. When the total loss value does not reach the preset condition, iteratively update all the pixel values in the initial image until the total loss value reaches the preset condition, Determining the updated initial image as a composite image;

Associating the first texture style label with the synthesized image, and determining the synthesized image and the first texture style label as the synthesis result.
A print verification method, which includes:

Receive certificate verification instructions, obtain the printed version and verification information of the pending certificate;

Input the printed document to be trained into the document recognition model that has been trained; the document recognition model is trained by the OCR image sample generated by the OCR image sample generation method according to any one of claims 1 to 4;

Perform OCR recognition on the printed document to be documented by the document recognition model, and obtain an OCR recognition result output by the document recognition model, where the OCR recognition result includes the document-related text information in the printed document to be documented;

Comparing the OCR recognition result with the verification information to determine whether the printed document to be certified meets the verification information;

If the printed version of the certificate to be issued meets the verification information, it is determined that the verification is passed;

If the printed body of the document to be issued does not conform to the verification information, confirm that the verification is not passed, and prompt on the display interface.
5. The method for verification of printed matter according to claim 5, wherein, before inputting the printed matter of the to-be-documented document into a trained document recognition model, the method comprises:

Obtain a certificate sample set; the certificate sample set contains several certificate samples, and one certificate sample is associated with a sample label; when the certificate sample is an image sample, the sample label is an image sample label; when the certificate sample is OCR In the case of an image sample, the sample label is an OCR image sample label; the OCR image sample is generated by the OCR image sample generation method according to any one of claims 1 to 4; the number of the image samples in the certificate sample set is less than The number of OCR image samples;

Input the document sample set into a deep learning OCR model containing initial parameters;

Performing OCR recognition on the certificate sample through the deep learning OCR model, and obtaining the training recognition result of the certificate sample output by the deep learning OCR model;

Matching the training recognition result with the sample label to obtain the loss value of the certificate sample;

When the loss value reaches a preset convergence condition, the deep learning OCR model after convergence is recorded as a certificate recognition model completed by training.
An OCR image sample generating device, which includes:

A receiving module, configured to receive an image generation instruction and obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

The input module is used to input the image sample into a preset font typesetting generation model, and by performing text detection and character recognition on the image sample, to obtain the first annotation information of the image sample recognized by the font typesetting generation model , And obtain a simulation result generated by reconstruction of the font typesetting generation model according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image and a second font label , The second typesetting label and the second annotation information;

The synthesis module is configured to input the image sample and the simulated image into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model is based on the style features and the content The feature performs style transfer and synthesis on the simulated image to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

A generating module for marking the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time recording the composite image as The OCR image sample corresponding to the image sample, and the OCR image sample is associated with the OCR image sample tag.
A printed body verification device, which includes:

The acquisition module is used to receive the certificate verification instruction, and obtain the printed version and verification information of the certificate;

The training module is used to input the printed document to be trained into the document recognition model that has been trained; the document recognition model is trained through the OCR image sample generated by the OCR image sample generation method according to any one of claims 1 to 4 ；

The recognition module is configured to perform OCR recognition on the printed document to be documented through the document recognition model, and obtain the OCR recognition result output by the document recognition model. The OCR recognition result includes the document-related printed document to be documented Text information;

The comparison module is configured to compare the OCR recognition result with the verification information, and determine whether the printed document to be document meets the verification information;

The first determining module is configured to determine that the verification is passed if the printed document to be document meets the verification information;

The second determination module is configured to confirm that the verification is not passed if the printed document to be document does not meet the verification information, and prompt on the display interface.
A computer device includes a memory, a processor, and a computer program that is stored in the memory and can run on the processor, wherein the processor implements the following steps when the processor executes the computer program:

Receiving an image generation instruction to obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

Input the image sample into a preset font typesetting generation model, by performing text detection and character recognition on the image sample, the font typesetting generation model recognizes the first annotation information of the image sample, and obtains the The font typesetting generation model reconstructs and generates a simulation result according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image, a second font label, and a second typesetting label And the second label information;

The image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model performs simulation on the simulation based on the style features and the content features. Image style transfer and synthesis are performed to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

Mark the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time record the composite image as corresponding to the image sample And associate the OCR image sample with the OCR image sample tag.
The computer device according to claim 9, wherein said inputting said image sample into a preset font typesetting generation model, and obtaining said font typesetting generation model by performing text detection and character recognition on said image sample The first annotation information of the image sample includes:

Perform text detection on the image sample through the font typesetting generation model, and at the same time extract the text features of the image sample, and obtain the image sample's value of the image sample identified by the font typesetting generation model according to the extracted text features Regional results; the regional results include a number of text regions containing text and region coordinates associated with each of the text regions;

The text feature of each text area is extracted by the font typesetting generation model, and each text area recognized by the font typesetting generation model according to the extracted text feature of each text area is obtained The text content;

The text area, the area coordinates associated with the text area, and the text content of the text area are recorded as the first information of the image sample, and all the first information is marked as the first information One label information.
9. The computer device according to claim 9, wherein said obtaining a simulation result generated by reconstruction of said font typesetting generation model according to said first font label, said first typesetting label and said first annotation information, include:

Inputting the first font label, the first typesetting label, and the first annotation information into a reconstruction model in the font typesetting generation model; the reconstruction model is trained based on a GAN model;

Performing combined reconstruction through a generator in the reconstruction model to obtain the simulated image, the second font label, the second typesetting label, and the second annotation information output by the reconstruction model;

The simulation image, the second font label, the second typesetting label, and the second annotation information are recorded as a simulation result output by the font typesetting generation model.
The computer device according to claim 9, wherein the image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content characteristics, and the style synthesis model is based on the The style feature and the content feature perform style transfer and synthesis on the simulated image to generate a synthesis result, including:

Using the simulated image as an initial image, and acquiring all pixel values of the initial image;

Extracting the style feature of the image sample and the style feature of the initial image through the style synthesis model, and calculating a style loss value according to the style feature of the image sample and the style feature of the initial image;

Extracting the content feature of the image sample and the content feature of the initial image through the style synthesis model, and calculating a content loss value according to the content feature of the image sample and the content feature of the initial image;

Weighting the style loss value and the content loss value to obtain a total loss value;

Gradient descent is performed through the L-BFGS algorithm. When the total loss value does not reach the preset condition, all the pixel values in the initial image are iteratively updated until the total loss value reaches the preset condition, Determining the updated initial image as a composite image;

Associating the first texture style label with the synthesized image, and determining the synthesized image and the first texture style label as the synthesis result.
A computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the following steps when the processor executes the computer program:

Receive certificate verification instructions, obtain the printed version and verification information of the pending certificate;

Input the printed document to be trained into the document recognition model that has been trained; the document recognition model is trained through the OCR image sample generated by the OCR image sample generation method according to any one of claims 1 to 4;

Perform OCR recognition on the printed document to be documented by the document recognition model, and obtain an OCR recognition result output by the document recognition model, where the OCR recognition result includes the document-related text information in the printed document to be documented;

Comparing the OCR recognition result with the verification information to determine whether the printed document to be certified meets the verification information;

If the printed version of the certificate to be issued meets the verification information, it is determined that the verification is passed;

If the printed body of the document to be issued does not conform to the verification information, confirm that the verification is not passed, and prompt on the display interface.
The computer device according to claim 13, wherein, before inputting the printed document to be into the trained document recognition model, the processor further implements the following steps when executing the computer program:

Obtain a certificate sample set; the certificate sample set contains several certificate samples, and one certificate sample is associated with a sample label; when the certificate sample is an image sample, the sample label is an image sample label; when the certificate sample is OCR In the case of an image sample, the sample label is an OCR image sample label; the OCR image sample is generated by the OCR image sample generation method according to any one of claims 1 to 4; the number of the image samples in the certificate sample set is less than The number of OCR image samples;

Input the document sample set into a deep learning OCR model containing initial parameters;

Performing OCR recognition on the certificate sample through the deep learning OCR model, and obtaining the training recognition result of the certificate sample output by the deep learning OCR model;

Matching the training recognition result with the sample label to obtain the loss value of the certificate sample;

When the loss value reaches a preset convergence condition, the deep learning OCR model after convergence is recorded as a certificate recognition model completed by training.
A computer-readable storage medium having a computer program stored on the computer-readable storage medium, wherein the computer program is executed by a processor to implement the following steps:

Receiving an image generation instruction to obtain an image sample; the image sample is associated with an image sample label, and the image sample label includes a first font label, a first typesetting label, and a first texture style label;

Input the image sample into a preset font typesetting generation model, by performing text detection and character recognition on the image sample, the font typesetting generation model recognizes the first annotation information of the image sample, and obtains the The font typesetting generation model reconstructs and generates a simulation result according to the first font label, the first typesetting label, and the first annotation information; the simulation result includes a simulated image, a second font label, and a second typesetting label And the second label information;

The image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content features, and the style synthesis model performs simulation on the simulation based on the style features and the content features. Image style transfer and synthesis are performed to generate a synthesis result; the synthesis result includes the synthesized image and the first texture style label;

Mark the second font label, the second typesetting label, the second annotation information, and the first texture style label as OCR image sample labels, and at the same time record the composite image as corresponding to the image sample And associate the OCR image sample with the OCR image sample tag.
The computer-readable storage medium according to claim 15, wherein said inputting said image sample into a preset font typesetting generation model, and obtaining said font typesetting generation by performing text detection and character recognition on said image sample The model recognizes the first annotation information of the image sample, including:

Perform text detection on the image sample through the font typesetting generation model, and at the same time extract the text features of the image sample, and obtain the image sample's value of the image sample identified by the font typesetting generation model according to the extracted text features Regional results; the regional results include a number of text regions containing text and region coordinates associated with each of the text regions;

The text feature of each text area is extracted by the font typesetting generation model, and each text area recognized by the font typesetting generation model according to the extracted text feature of each text area is obtained The text content;

The text area, the area coordinates associated with the text area, and the text content of the text area are recorded as the first information of the image sample, and all the first information is marked as the first information One label information.
The computer-readable storage medium according to claim 15, wherein said acquiring said font typesetting generation model is reconstructed and generated according to said first font label, said first typesetting label and said first annotation information Simulation results, including:

Inputting the first font label, the first typesetting label, and the first annotation information into a reconstruction model in the font typesetting generation model; the reconstruction model is trained based on a GAN model;

Performing combined reconstruction through a generator in the reconstruction model to obtain the simulated image, the second font label, the second typesetting label, and the second annotation information output by the reconstruction model;

The simulation image, the second font label, the second typesetting label, and the second annotation information are recorded as a simulation result output by the font typesetting generation model.
The computer-readable storage medium of claim 15, wherein the image sample and the simulated image are input into a preset style synthesis model, the style synthesis model extracts style features and content characteristics, and the style synthesis The model performs style transfer and synthesis on the simulated image according to the style feature and the content feature to generate a synthesis result, including:

Using the simulated image as an initial image, and acquiring all pixel values of the initial image;

Extracting the style feature of the image sample and the style feature of the initial image through the style synthesis model, and calculating a style loss value according to the style feature of the image sample and the style feature of the initial image;

Extracting the content feature of the image sample and the content feature of the initial image through the style synthesis model, and calculating a content loss value according to the content feature of the image sample and the content feature of the initial image;

Weighting the style loss value and the content loss value to obtain a total loss value;

Perform gradient descent through the L-BFGS algorithm. When the total loss value does not reach the preset condition, iteratively update all the pixel values in the initial image until the total loss value reaches the preset condition, Determining the updated initial image as a composite image;

Associating the first texture style label with the synthesized image, and determining the synthesized image and the first texture style label as the synthesis result.
A computer-readable storage medium having a computer program stored on the computer-readable storage medium, wherein the computer program is executed by a processor to implement the following steps:

Receive certificate verification instructions, obtain the printed version and verification information of the pending certificate;

Input the printed document to be trained into the document recognition model that has been trained; the document recognition model is trained by the OCR image sample generated by the OCR image sample generation method according to any one of claims 1 to 4;

Perform OCR recognition on the printed document to be documented by the document recognition model, and obtain an OCR recognition result output by the document recognition model, where the OCR recognition result includes the document-related text information in the printed document to be documented;

Comparing the OCR recognition result with the verification information to determine whether the printed document to be certified meets the verification information;

If the printed version of the certificate to be issued meets the verification information, it is determined that the verification is passed;

If the printed body of the document to be issued does not conform to the verification information, confirm that the verification is not passed, and prompt on the display interface.
19. The computer-readable storage medium of claim 19, wherein, before the printed document to be entered into the trained document recognition model, when the computer program is executed by the processor, the following steps are further implemented:

Obtain a certificate sample set; the certificate sample set contains several certificate samples, and one certificate sample is associated with a sample label; when the certificate sample is an image sample, the sample label is an image sample label; when the certificate sample is OCR In the case of an image sample, the sample label is an OCR image sample label; the OCR image sample is generated by the OCR image sample generation method according to any one of claims 1 to 4; the number of the image samples in the certificate sample set is less than The number of OCR image samples;

Input the document sample set into a deep learning OCR model containing initial parameters;

Performing OCR recognition on the certificate sample through the deep learning OCR model, and obtaining the training recognition result of the certificate sample output by the deep learning OCR model;

Matching the training recognition result with the sample label to obtain the loss value of the certificate sample;

When the loss value reaches a preset convergence condition, the deep learning OCR model after convergence is recorded as a certificate recognition model completed by training.