Method, device and equipment for training a tabloid recognition model and storage medium

Abstract

The application provides a method, device and equipment for training a hand-written report text recognition model and a storage medium, wherein the method comprises the following steps: step 1, extracting two report image to form an original training group, obtaining a first signal corresponding to each of the two report images, and a second signal corresponding to each of the two report images; step 2, aligning energy points corresponding to the two first signals to obtain a first alignment path, and aligning energy points corresponding to the two second signals to obtain a second alignment path; step 3, calculating a comprehensive alignment distance and a comprehensive deformation distance according to the coordinates of the energy points on the first alignment path and the second alignment path, and constructing a new training group; step 4, repeating steps 1 to 3, training a linear support vector machine using the plurality of new training groups and determining a decision slope, and constructing a Gaussian kernel support vector machine through the decision slope; and step 5, training the Gaussian kernel support vector machine using the original training group to obtain a hand-written report text recognition model.

Description

Technical Field

[0001] This invention relates to the field of image recognition and classification technology, and in particular to a training method, apparatus, device, and storage medium for a handwritten message recognition model. Background Technology

[0002] Handwritten message recognition is an important research direction in the field of computer vision. It has strong versatility and can be widely used in application scenarios such as financial statement recognition, mail sorting, and grade statistics.

[0003] In related technologies, traditional template matching methods require a large number of templates and cannot cope with changes in handwriting style, which has limitations. Convolutional neural network-based models can basically adapt to changes in handwriting style, but the generalization ability and robustness of the models are poor, and they cannot handle font deformation due to changes in perspective. Therefore, expanding the training set through data augmentation and other techniques increases the amount of two-dimensional information in the images in the training set, leading to overfitting problems during the training process.

[0004] Based on the above analysis of the development status of this technical field, the existing technical solutions lack a training method that extracts one-dimensional information from the image to represent image features and considers the differences in alignment distance and deformation distance respectively. Summary of the Invention

[0005] The purpose of this invention is to provide a training method, apparatus, device, and storage medium for a handwritten message recognition model, aiming to solve the above-mentioned problems in the prior art.

[0006] According to a first aspect of the present invention, a method for training a handwritten message recognition model is provided, comprising:

[0007] Step 1: Extract two message images to form the original training group, and obtain the first signal corresponding to each of the two message images and the second signal corresponding to each of the two message images.

[0008] Step 2: Use dynamic time programming to align the energy points corresponding to the two first signals to obtain the first alignment path, and align the energy points corresponding to the two second signals to obtain the second alignment path.

[0009] Step 3: Based on the coordinates of the energy points on the first alignment path and the second alignment path, calculate the alignment distance and deformation distance corresponding to each path. Add the alignment distances of the two paths to obtain the comprehensive alignment distance, and add the deformation distances of the two paths to obtain the comprehensive deformation distance. Construct a new training group using the comprehensive alignment distance and the comprehensive deformation distance.

[0010] Step 4: Repeat steps 1 to 3 to construct multiple new training groups, use multiple new training groups to train the linear support vector machine, determine the decision slope of the linear support vector machine, and construct a Gaussian kernel support vector machine based on the decision slope.

[0011] Step 5: Train the Gaussian kernel support vector machine using the original training set. After training, the trained handwritten message recognition model is obtained.

[0012] According to a second aspect of the present invention, a training apparatus for a handwritten message recognition model is provided, comprising:

[0013] The subsampling module is used to extract two message images to form the original training group, and to obtain the first signal corresponding to each of the two message images, as well as the second signal corresponding to each of the two message images.

[0014] The dynamic alignment module is used to align the energy points corresponding to two first signals using a dynamic time planning method to obtain a first alignment path, and to align the energy points corresponding to two second signals to obtain a second alignment path.

[0015] The new training group construction module is used to calculate the alignment distance and deformation distance of each path based on the coordinates of the energy points on the first alignment path and the second alignment path. The alignment distance of the two paths is added together to obtain the comprehensive alignment horizontal distance, and the deformation distance of the two paths is added together to obtain the comprehensive deformation distance. The new training group is constructed using the comprehensive alignment distance and the comprehensive deformation distance.

[0016] The Gaussian kernel construction module is used to repeatedly call the subsampling module, dynamic alignment module, and new training group construction module to construct multiple new training groups, train linear support vector machines using multiple new training groups, determine the decision slope of the linear support vector machine, and construct Gaussian kernel support vector machines based on the decision slope.

[0017] The Gaussian kernel support vector machine training module is used to train a Gaussian kernel support vector machine using the original training set. After training, a trained handwritten message recognition model is obtained.

[0018] According to a third aspect of the present invention, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the training method for the handwritten message recognition model provided in the first aspect of the present disclosure.

[0019] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which an information transmission implementation program is stored, wherein when executed by a processor, the program implements the steps of the training method for the handwritten message recognition model provided in the first aspect of the present disclosure.

[0020] The technical solution provided by the embodiments of the present invention includes the following beneficial effects: using Radon transform and subsampling to transform a two-dimensional image into two corresponding one-dimensional signals reduces the amount of information contained in the image and achieves rotation invariance of the image; using dynamic time programming to align the energy points corresponding to the one-dimensional signals and construct a new training set describing the alignment distance and deformation distance can cope with font deformations from different viewpoints, enabling the new training set to accurately represent image features and improve the training effect of the handwritten newspaper recognition model.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the training method for the handwritten message recognition model according to an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of energy map subsampling according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of dynamic time planning alignment according to an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the training device for the handwritten message recognition model according to an embodiment of the present invention;

[0027] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0029] Method Implementation Examples

[0030] According to an embodiment of the present invention, a training method for a handwritten message recognition model is provided. Figure 1 This is a flowchart of the training method for the handwritten message recognition model according to an embodiment of the present invention, as shown below. Figure 1 As shown, the training method for the handwritten message recognition model according to an embodiment of the present invention specifically includes:

[0031] In step S110, two message images are extracted to form the original training set, and the first signal corresponding to each of the two message images and the second signal corresponding to each of the two message images are obtained. Specifically, this includes:

[0032] Radon transform is applied to two message images in the training group to generate an energy map with the rotation angle as the y-axis and energy as the x-axis. Radon transform is a mathematical integral transform that can obtain the projection of the image at different angles. The image after Radon transform contains the same information as the original image. The energy of Radon transform at different angles is defined as shown in Equation 1:

[0033]

[0034] Where E(p) represents energy, R(p, τ) represents Radon transform, p represents the slope of the line, and τ represents the intercept. During the Radon transform process, an energy map is generated with the rotation angle as the y-axis and energy as the x-axis. The energy map transforms slowly in the y-axis direction and has high information redundancy, so subsampling is performed in the x-axis direction.

[0035] Subsampling is performed on the energy maps corresponding to each message image. The signal corresponding to the point with the maximum energy value is selected as the first signal, and the signal corresponding to the first signal with a rotation angle coordinate that differs from the first signal by 90 degrees is selected as the second signal. This yields the first signal and the second signal for each of the two message images. Figure 2 This is a schematic diagram of energy map subsampling according to an embodiment of the present invention, as shown below. Figure 2 As shown, taking a message image from the training group as an example, Figure 2 In Figure (b), the Radon transformation is performed on the handwritten message 9. Figure (c) is the energy map corresponding to the Radon transformation. Only two points can be extracted within a 180-degree range. That is, the angle point with the largest energy value is taken as the first signal, which is 26.4 degrees in this embodiment. The second signal is the point corresponding to -63.6 degrees.

[0036] In step S120, the energy points corresponding to the two first signals are aligned using dynamic time programming to obtain a first alignment path, and the energy points corresponding to the two second signals are aligned to obtain a second alignment path. Specifically, this includes:

[0037] Dynamic time programming is a method for measuring the similarity between two time series. It minimizes the overall distance difference between the two series through dynamic programming.

[0038] Figure 3 This is a schematic diagram of dynamic time planning alignment according to an embodiment of the present invention, as shown below. Figure 3 As shown, dynamic time programming is used to align the energy curve corresponding to handwritten message 8 with the energy curve corresponding to handwritten message 9. Figure 3 This demonstrates the effect of aligning one of the paths. Figure 3 (c) shows the diagonal curve as the first alignment path in the embodiment. Figure 3 (d) represents the longitudinal difference after time alignment with the first path.

[0039] When the two curves are exactly the same, that is, when there is no difference in horizontal time and vertical points, Figure 3 (c) The alignment path represented by the diagonal is a straight line. Therefore, the difference in distance after alignment is mainly divided into the alignment distance in the horizontal direction and the deformation distance in the vertical direction.

[0040] In step S130, based on the coordinates of the energy points on the first and second alignment paths, the alignment distance and deformation distance corresponding to each path are calculated. The alignment distances of the two paths are added together to obtain the comprehensive alignment distance, and the deformation distances of the two paths are added together to obtain the comprehensive deformation distance. A new training set is constructed using the comprehensive alignment distance and the comprehensive deformation distance. Specifically, this includes:

[0041] Calculate the alignment distance corresponding to the first alignment path using Formula 2. And the alignment distance corresponding to the second alignment path

[0042]

[0043] In the formula, Indicates the first signal. Indicates the second signal, PT ix [j] represents the x-axis coordinate of energy point j in the alignment path, PT iy [j] represents the y-coordinate of energy point j in the alignment path, and i represents the path number of the alignment path. and The sum of these values ​​gives the overall alignment distance E. 1 ;

[0044] The deformation distance corresponding to the first alignment path is calculated using Formula 3. and the deformation distance corresponding to the second alignment path

[0045]

[0046] In the formula, DPT ix [j] represents the x-axis coordinate of energy point j after taking the first difference of the aligned path, DPT iy [j] represents the y-coordinate of the energy point j after taking the first-order difference of the aligned path. and The sum of these values ​​yields the combined deformation distance E. 2 .

[0047] The overall alignment distance E 1 and the combined deformation distance E 2 The new training group is added to the original training group, and a new label is defined for the original training group after the feature is added according to the label definition rules. The label definition rules are as follows: if the original labels of the two message images are the same, the new label is defined as 1; if the original labels of the two message images are different, the new label is defined as 0.

[0048] In step S140, steps S110 to S130 are repeated to construct multiple new training sets. Linear support vector machines are trained using these new training sets. The decision slope of the linear support vector machine is determined, and a Gaussian kernel support vector machine is constructed based on the decision slope. Specifically, this includes:

[0049] A linear support vector machine (SVM) is used to perform binary classification on each new training group, yielding the decision slope of the corresponding decision surface. The negative of the decision slope is used as the weight coefficient of the Gaussian kernel function in the Gaussian kernel SVM. The form of the Gaussian kernel function is defined by Equation 4.

[0050]

[0051] In the formula, K(·) represents the Gaussian kernel function, exp() represents the exponential function with base e, and E 1 +aE 2 σ represents the distance metric of the Gaussian kernel support vector machine during training, a represents the weight coefficient, and σ represents the control decision boundary parameter.

[0052] In step S150, a Gaussian kernel support vector machine is trained using multiple sets of original training groups, and the trained handwritten message recognition model is obtained after training.

[0053] The method further includes:

[0054] In step S160, the trained handwritten message recognition model is used to obtain the classification result of the message image to be tested. The comparison of the performance of the handwritten message recognition model of the present invention with other classic classification methods is shown in Table 1:

[0055] Table 1. Comparison of Classification Results

[0056]

[0057] As shown in Table 1, under the same training conditions, the accuracy of the handwritten message recognition model obtained by this invention reaches 99.15%.

[0058] In summary, to address the existing problems, this invention proposes a training method for a handwritten message recognition model. This method uses Radon transform to generate an energy map, performing subsampling in the horizontal direction of the energy map to avoid redundant information caused by slow vertical sampling. This transforms the two-dimensional image into two corresponding one-dimensional signals, reducing the amount of information contained in the image and achieving rotation invariance. A dynamic time programming method is used to align the energy points corresponding to the one-dimensional signals, and the sequence of one of the energy maps is stretched or compressed to minimize the difference in energy curves between the two images in the alignment path. A new training set describing the alignment distance and deformation distance is constructed, capable of handling font deformation from various viewpoints, ensuring the new training set accurately represents image features. A linear support vector machine (SVM) is trained using this new training set. A Gaussian kernel SVM is constructed using the decision slope of the linear SVM. The Gaussian kernel function has weight coefficients determined by the decision slope when measuring distance, considering the different importance of different distances in the total distance measurement, thus improving the accuracy of decision-making and the training effect of the handwritten message recognition model.

[0059] Device Examples

[0060] According to an embodiment of the present invention, a training device for a handwritten message recognition model is provided. Figure 4 This is a schematic diagram of the training device for the handwritten message recognition model according to an embodiment of the present invention, as shown below. Figure 4 As shown, the training device for the handwritten message recognition model according to an embodiment of the present invention specifically includes:

[0061] The subsampling module 40 is used to extract two message images to form an original training set, and to obtain the first signal and the second signal corresponding to each of the two message images. Specifically, it is used for:

[0062] Radon transform is performed on the two message images in the training group to generate an energy map with the rotation angle as the y-axis and the energy as the x-axis during the Radon transform process.

[0063] Subsampling is performed on the energy maps corresponding to each message image. The signal corresponding to the point with the maximum energy value is selected as the first signal, and the signal corresponding to the first signal with a rotation angle coordinate that differs from the first signal by 90 degrees is selected as the second signal.

[0064] The dynamic alignment module 42 is used to align the energy points corresponding to the two first signals using a dynamic time planning method to obtain a first alignment path, and to align the energy points corresponding to the two second signals to obtain a second alignment path.

[0065] The new training set construction module 44 is used to calculate the alignment distance and deformation distance of each path based on the coordinates of the energy points on the first and second alignment paths. It then adds the alignment distances of the two paths to obtain the comprehensive alignment distance and adds the deformation distances of the two paths to obtain the comprehensive deformation distance. Finally, it constructs a new training set using the comprehensive alignment distance and comprehensive deformation distance. Specifically, it is used for:

[0066] Calculate the alignment distance corresponding to the first alignment path using Formula 1. And the alignment distance corresponding to the second alignment path

[0067]

[0068] In the formula, Indicates the first signal. Indicates the second signal, PT ix [j] represents the x-axis coordinate of energy point j in the alignment path, PT iy [j] represents the y-coordinate of energy point j in the alignment path, and i represents the path number of the alignment path. and The sum of these values ​​gives the overall alignment distance E. 1 ;

[0069] The deformation distance corresponding to the first alignment path is calculated using Formula 2. and the deformation distance corresponding to the second alignment path

[0070]

[0071] In the formula, DPT ix [j] represents the x-axis coordinate of energy point j after taking the first difference of the aligned path, DPT iy [j] represents the y-coordinate of the energy point j after taking the first-order difference of the aligned path. and The sum of these values ​​yields the combined deformation distance E. 2 .

[0072] The overall alignment distance E 1 and the combined deformation distance E 2 The new training group is added to the original training group, and a new label is defined for the original training group after the feature is added according to the label definition rules. The label definition rules are as follows: if the original labels of the two message images are the same, the new label is defined as 1; if the original labels of the two message images are different, the new label is defined as 0.

[0073] The Gaussian kernel construction module 46 is used to repeatedly call the subsampling module 40, the dynamic alignment module 42, and the new training group construction module 44 to construct multiple new training groups. These new training groups are then used to train the linear support vector machine (SVM), determine the decision slope of the SVM, and construct the Gaussian kernel SVM based on the decision slope. Specifically, it is used for:

[0074] A linear support vector machine (SVM) is used to perform binary classification on each new training group, obtaining the decision slope of the corresponding decision surface of the linear SVM. The inverse of the decision slope is used as the weight coefficient of the Gaussian kernel function in the Gaussian kernel SVM. The form of the Gaussian kernel function is defined by Equation 3:

[0075]

[0076] In the formula, K(·) represents the Gaussian kernel function, exp() represents the exponential function with base e, and E 1 +aE 2 σ represents the distance metric of the Gaussian kernel support vector machine during training, a represents the weight coefficient, and σ represents the control decision boundary parameter.

[0077] The Gaussian kernel support vector machine training module 48 is used to train the Gaussian kernel support vector machine using the original training group. After training, a trained handwritten message recognition model is obtained.

[0078] The device further includes:

[0079] The model uses module 410 to obtain the classification results of the text image to be tested using the trained handwritten message recognition model.

[0080] In summary, to address the existing problems, this invention provides a training device for a handwritten message recognition model. It uses Radon transform to generate an energy map, performs subsampling in the horizontal direction of the energy map to avoid redundant information caused by slow vertical sampling, and transforms the two-dimensional image into two corresponding one-dimensional signals, reducing the amount of information contained in the image and achieving rotation invariance. A dynamic time programming method is used to align the energy points corresponding to the one-dimensional signals, and stretching or compressing the sequence of one of the energy maps ensures that the energy curves of the two images in the alignment path have small differences. A new training set describing the alignment distance and deformation distance is constructed, capable of handling font deformation from various viewpoints, enabling the new training set to accurately represent image features. A linear support vector machine is trained using this new training set, and a Gaussian kernel support vector machine is constructed using the decision slope of the linear support vector machine. The Gaussian kernel function has weight coefficients determined by the decision slope when measuring distance, considering the different importance of different distances in the total distance measurement, thus improving the accuracy of decision-making and the training effect of the handwritten message recognition model.

[0081] Electronic device examples

[0082] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device 500 may include at least one processor 510 and a memory 520. The processor 510 can execute instructions stored in the memory 520. The processor 510 is communicatively connected to the memory 520 via a data bus. In addition to the memory 520, the processor 510 can also be communicatively connected to an input device 530, an output device 540, and a communication device 550 via the data bus.

[0083] Processor 510 can be any conventional processor, such as a commercially available CPU. The processor may also include, for example, a Graphics Processing Unit (GPU), a Field Programmable Gate Array (FPGA), a System on Chip (SOC), an Application Specific Integrated Circuit (ASIC), or a combination thereof.

[0084] The memory 520 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0085] In this embodiment of the present disclosure, the memory 520 stores executable instructions, and the processor 510 can read the executable instructions from the memory 520 and execute the instructions to implement all or part of the steps of the training method for any of the handwritten message recognition models in the above exemplary embodiments.

[0086] Computer-readable storage medium embodiments

[0087] In addition to the methods and apparatus described above, exemplary embodiments of this disclosure may also be a computer program product or a computer-readable storage medium storing the computer program product, wherein the computer program product includes computer program instructions that can be executed by a processor to implement all or part of the steps described in the training method for any of the handwritten message recognition models in the exemplary embodiments described above.

[0088] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. Programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages, and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0089] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media include: static random access memory (SRAM) having one or more electrically connected wires, electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk, or any suitable combination thereof.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for a handwritten message recognition model, characterized in that, include: Step 1: Extract two message images to form the original training set, and obtain the first signal and the second signal corresponding to each of the two message images. Specifically, this includes: Radon transform is performed on the two message images in the original training group to generate an energy map with the rotation angle as the y-axis and the energy as the x-axis during the Radon transform process. Subsampling is performed in the energy map corresponding to each message image, and the signal corresponding to the point with the maximum value of the energy point in the energy map is selected as the first signal, and the corresponding signal with a rotation angle coordinate that differs from the first signal by 90 degrees is selected as the second signal. Step 2: Use dynamic time programming to align the energy points corresponding to the two first signals to obtain the first alignment path, and align the energy points corresponding to the two second signals to obtain the second alignment path. Step 3: Based on the coordinates of the energy points on the first alignment path and the second alignment path, calculate the alignment distance and deformation distance corresponding to each path. Add the alignment distances of the two paths to obtain the comprehensive alignment distance and add the deformation distances of the two paths to obtain the comprehensive deformation distance. Construct a new training group using the comprehensive alignment distance and the comprehensive deformation distance. Step 4: Repeat steps 1 to 3 to construct multiple new training groups, train a linear support vector machine using these new training groups, determine the decision slope of the linear support vector machine, and construct a Gaussian kernel support vector machine using the decision slope. Specifically, this includes: A linear support vector machine (SVM) is used to perform binary classification on each new training group, obtaining the decision slope of the corresponding decision surface of the linear SVM. The negative of the decision slope is used as the weight coefficient of the Gaussian kernel function in the Gaussian kernel SVM. The form of the Gaussian kernel function is defined by Equation 3: Official 3; In the formula, represents the Gaussian kernel function, and exp() represents the exponential function with base e. This represents the distance metric used during the training process of the Gaussian kernel support vector machine. Indicates the overall alignment distance. Indicates the total deformation distance. This represents the weighting coefficient. Indicates the control decision boundary parameters; Step 5: Train the Gaussian kernel support vector machine using the original training set. After training, the trained handwritten message recognition model is obtained.

2. The method according to claim 1, characterized in that, The method further includes: The trained handwritten message recognition model is used to obtain the classification result of the message image to be tested.

3. The method according to claim 1, characterized in that, The calculation of the alignment distance and deformation distance corresponding to each path, the sum of the alignment distances of two paths to obtain the comprehensive alignment distance, and the sum of the deformation distances of two paths to obtain the comprehensive deformation distance specifically include: Calculate the alignment distance corresponding to the first alignment path using Formula 1. and the alignment distance corresponding to the second alignment path. : Official 1; In the formula, Indicates the first signal. Indicates the second signal. Indicates the energy points in the alignment path of x Axis coordinates Indicates the energy points in the alignment path of y Axis coordinates The path number represents the alignment path. and Add them together to get the overall alignment distance ; The deformation distance corresponding to the first alignment path is calculated using Formula 2. and the deformation distance corresponding to the second alignment path. : Official 2; In the formula, This represents the energy point after taking the first-order difference of the aligned path. of x Axis coordinates This represents the energy point after taking the first-order difference of the aligned path. of y Axis coordinates, and Add them together to get the total deformation distance .

4. The method according to claim 1, characterized in that, The construction of a new training group using the integrated alignment distance and the integrated deformation distance specifically includes: The integrated alignment distance and the integrated deformation distance are added to the original training group, and a new label is defined for the original training group after adding features according to the label definition rules to obtain a new training group. The label definition rules are as follows: if the original labels of two message images are consistent, the new label is defined as 1; if the original labels of two message images are inconsistent, the new label is defined as 0.

5. A training device for a handwritten message recognition model, characterized in that, include: The subsampling module is used to extract two message images to form the original training set, obtain the first signal corresponding to each of the two message images, and the second signal corresponding to each of the two message images. Specifically, it is used for: Radon transform is performed on the two message images in the original training group to generate an energy map with the rotation angle as the y-axis and the energy as the x-axis during the Radon transform process. Subsampling is performed in the energy map corresponding to each message image, and the signal corresponding to the point with the maximum value of the energy point in the energy map is selected as the first signal, and the corresponding signal with a rotation angle coordinate that differs from the first signal by 90 degrees is selected as the second signal. The dynamic alignment module is used to align the energy points corresponding to the two first signals using a dynamic time planning method to obtain a first alignment path, and to align the energy points corresponding to the two second signals to obtain a second alignment path. The new training group construction module is used to calculate the alignment distance and deformation distance of each path according to the coordinates of the energy points on the first alignment path and the second alignment path, add the alignment distances of the two paths to obtain the comprehensive alignment distance, add the deformation distances of the two paths to obtain the comprehensive deformation distance, and construct a new training group through the comprehensive alignment distance and the comprehensive deformation distance. The Gaussian kernel construction module is used to repeatedly call the subsampling module, the dynamic alignment module, and the new training group construction module to construct multiple new training groups, train a linear support vector machine using these multiple new training groups, determine the decision slope of the linear support vector machine, and construct a Gaussian kernel support vector machine using the decision slope. Specifically, it is used for: A linear support vector machine (SVM) is used to perform binary classification on each new training group, obtaining the decision slope of the corresponding decision surface of the linear SVM. The negative of the decision slope is used as the weight coefficient of the Gaussian kernel function in the Gaussian kernel SVM. The form of the Gaussian kernel function is defined by Equation 3: Official 3; In the formula, represents the Gaussian kernel function, and exp() represents the exponential function with base e. This represents the distance metric used during the training process of the Gaussian kernel support vector machine. Indicates the overall alignment distance. Indicates the total deformation distance. This represents the weighting coefficient. Indicates the control decision boundary parameters; The Gaussian kernel support vector machine training module is used to train the Gaussian kernel support vector machine using the original training set, and the trained handwritten message recognition model is obtained after training.

6. The apparatus according to claim 5, characterized in that, The device further includes: The model uses a module to obtain the classification result of the text image to be tested using the trained handwritten message recognition model.

7. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the training method for the handwritten message recognition model as described in any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the training method for the handwritten message recognition model as described in any one of claims 1 to 4.

Images