CN113591884B - Character recognition model determination method, device, equipment and storage medium - Google Patents

Abstract

This application discloses a method, device, equipment and storage medium for determining a character recognition model, which belongs to the technical field of character recognition. The method includes: the electronic device obtains device information of a device selected by the user for running the character recognition model, and obtains a sample uploaded by the user with annotation information, where the annotation information can be used to describe the character in the corresponding sample. Attributes. The electronic device can select the network model to be trained corresponding to the device information and the label information from the network model collection based on the device information and the label information. In this way, the character recognition model obtained by training the network model to be trained based on the sample uploaded by the user It is more suitable to run on the device selected by the user, and the character recognition model is more suitable for identifying characters in objects to be recognized that are similar to the samples uploaded by the user, that is, it can accurately identify characters in objects to be recognized that are similar to the samples uploaded by the user. The rate is higher.

Description

Character recognition model determination method, device, equipment and storage medium

Technical field

The present application relates to the field of character recognition technology, and in particular to a method, device, equipment and storage medium for determining a character recognition model.

Background technique

OCR (Optical Character Recognition, optical character recognition) refers to a technology that can recognize characters to be recognized in images to be recognized. At present, OCR is widely used in logistics, medical care, finance, insurance and other fields.

Generally, users can obtain a general OCR model and use the general OCR model to recognize characters to be recognized in different images to be recognized. Wherein, the general OCR model refers to a model trained through a general model training method. However, due to the large differences between different images to be recognized, the general OCR model has a low recognition accuracy of characters to be recognized in different images to be recognized, that is, the general OCR model has poor generalization ability.

Contents of the invention

The present application provides a method, device, equipment and storage medium for determining a character recognition model, which can solve the problem of determining a character recognition model in related technologies. The technical solutions are as follows:

On the one hand, a method for determining a character recognition model is provided, and the method includes:

Obtain device information, and obtain samples uploaded by the user with annotation information. The device information is information about the device selected by the user for running the character recognition model. The annotation information is used to describe the attributes of the characters in the corresponding sample. ;

Based on the device information and the annotation information, determine a network model to be trained from a network model set, where the network model set includes at least two network models;

The network model to be trained is trained based on the samples uploaded by the user to obtain the character recognition model.

In a possible implementation of the present application, the annotation information includes field categories and/or field rules of the characters in the corresponding samples, and the annotation information is determined from a network model collection based on the device information and the annotation information. Network models to be trained include:

According to the device information, select a network model subset from the network model set, wherein the size and structure of the network models included in the network model subset match the device;

According to the field categories and/or field rules of the samples uploaded by the user, the corresponding network model is determined from the network model subset to obtain the network model to be trained.

In a possible implementation of this application, the network model to be trained is trained based on the samples uploaded by the user to obtain the character recognition model, which includes:

Divide the samples uploaded by the user into a training sample set and a test sample set, where the samples in the training sample set and the samples in the test sample set have the same data distribution rule;

Train the network model to be trained multiple times based on the training sample set;

After each training obtains the first network model, the current first network model is tested based on the test sample set to obtain test indicators;

If the current test index of the first network model is higher than the test index of the first network model obtained by the previous training, continue training and testing; until the current test index of the first network model is lower than the first test index obtained by the previous training. Test indicators of the network model, end training and testing, and determine the first network model obtained from the previous training as the character recognition model.

In a possible implementation of this application, the current first network model includes a detection network and an identification network, and the current first network model is tested based on the test sample set to obtain test indicators, including:

Test the detection network based on the test sample set to obtain test indicators of the detection network, and the test indicators of the detection network are used to indicate the detection accuracy of the detection network;

Test the recognition network based on the test sample set to obtain test indicators of the recognition network, and the test indicators of the recognition network are used to indicate the character recognition accuracy of the recognition network;

The test indicators of the detection network and the test indicators of the identification network are determined as test indicators of the current first network model.

In a possible implementation of this application, the method further includes:

Generate a test report, the test report including one or more of the test indicators of the detection network, the test indicators of the identification network, and model feedback information. The model feedback information is used to instruct the user to update the uploaded sample. and/or update annotation information in uploaded samples.

In a possible implementation of the present application, before determining the first network model obtained by the previous training as the character recognition model, the method further includes:

Perform fixed-point quantization processing on the first network model obtained from the previous training to obtain the first network model after fixed-point quantization;

Determining the first network model obtained from the previous training as the character recognition model includes:

The fixed-point quantized first network model is determined as the character recognition model.

On the one hand, a device for determining a character recognition model is provided, and the device includes:

The acquisition module is used to obtain device information and obtain samples with annotation information uploaded by the user. The device information is the information of the device selected by the user for running the character recognition model. The annotation information is used to describe the corresponding sample. Properties of the characters in ;

A determination module, configured to determine a network model to be trained from a network model set based on the device information and the annotation information, where the network model set includes at least two network models;

A training module, configured to train the network model to be trained based on the samples uploaded by the user to obtain the character recognition model.

In a possible implementation of this application, the annotation information includes field categories and/or field rules of characters in the corresponding sample, and the determination module is used to:

According to the device information, select a network model subset from the network model set, wherein the size and structure of the network models included in the network model subset match the device;

According to the field categories and/or field rules of the samples uploaded by the user, the corresponding network model is determined from the network model subset to obtain the network model to be trained.

In a possible implementation of this application, the training module is used for:

Divide the samples uploaded by the user into a training sample set and a test sample set, where the samples in the training sample set and the samples in the test sample set have the same data distribution rule;

Train the network model to be trained multiple times based on the training sample set;

After each training obtains the first network model, the current first network model is tested based on the test sample set to obtain test indicators;

If the current test index of the first network model is higher than the test index of the first network model obtained by the previous training, continue training and testing; until the current test index of the first network model is lower than the first test index obtained by the previous training. Test indicators of the network model, end training and testing, and determine the first network model obtained from the previous training as the character recognition model.

In a possible implementation of this application, the current first network model includes a detection network and a recognition network, and the training module is used to:

Test the detection network based on the test sample set to obtain test indicators of the detection network, and the test indicators of the detection network are used to indicate the detection accuracy of the detection network;

Test the recognition network based on the test sample set to obtain test indicators of the recognition network, and the test indicators of the recognition network are used to indicate the character recognition accuracy of the recognition network;

The test indicators of the detection network and the test indicators of the identification network are determined as test indicators of the current first network model.

In a possible implementation of this application, the device further includes:

Generating module, used to generate a test report, the test report includes one or more of the test indicators of the detection network, the test indicators of the identification network, and model feedback information, the model feedback information is used to instruct the user Update the uploaded sample and/or update the annotation information in the uploaded sample.

In a possible implementation of this application, the training module is also used to:

Perform fixed-point quantization processing on the first network model obtained from the previous training to obtain the first network model after fixed-point quantization;

Determining the first network model obtained from the previous training as the character recognition model includes:

The fixed-point quantized first network model is determined as the character recognition model.

On the one hand, an electronic device is provided, and the electronic device includes:

processor;

Memory used to store instructions executable by the processor;

Wherein, the processor is configured to implement the method for determining a character recognition model described in the above aspect.

On the one hand, a computer-readable storage medium is provided. Instructions are stored on the computer-readable storage medium. When the instructions are executed by a processor, the method for determining the character recognition model described in the above-mentioned aspect is implemented.

In one aspect, a computer program product containing instructions is provided, which when run on a computer causes the computer to execute the method for determining a character recognition model described in the above aspect.

The technical solution provided by this application can at least bring the following beneficial effects:

Obtain the device information of the device selected by the user for running the character recognition model, and obtain the sample uploaded by the user with annotation information, where the annotation information can be used to describe the attributes of the characters in the corresponding sample. The electronic device can select the network model to be trained corresponding to the device information and the label information from the network model collection based on the device information and the label information. In this way, the character recognition model obtained by training the network model to be trained based on the samples uploaded by the user It is more suitable to run on the device selected by the user, and the character recognition model is more suitable for identifying characters in objects to be recognized that are similar to samples uploaded by users, that is, it can accurately identify characters in objects to be recognized that are similar to samples uploaded by users. The rate is higher.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments will be briefly introduced below. 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 also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of a method for determining a character recognition model provided by an embodiment of the present application;

Figure 2 is a schematic diagram of a sample provided by the embodiment of the present application;

Figure 3 is a schematic structural diagram of a device for determining a character recognition model provided by an embodiment of the present application;

FIG. 4 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

Before giving a detailed explanation of the method for determining the character recognition model provided by the embodiment of the present application, the implementation environment involved in the embodiment of the present application is first introduced.

The method for determining the character recognition model provided by the embodiments of the present application can be applied to an AI (Artificial Intelligence, artificial intelligence) open platform. The AI open platform supports human-computer interaction. For example, users can upload collected samples to the AI open platform. , the AI open platform can process samples uploaded by users. For another example, users can perform operations such as labeling the sample in the AI open platform.

As an example, the AI open platform may include electronic equipment, and the electronic equipment may be used to train a network model to obtain a character recognition model. For example, the electronic device may be a notebook computer, a portable computer, a desktop computer, a server, etc., which is not limited in the embodiments of the present application.

After introducing the implementation environment involved in the embodiments of the present application, the method for determining the character recognition model provided by the embodiments of the present application will be introduced in detail with reference to the accompanying drawings.

Figure 1 is a flow chart of a method for determining a character recognition model provided by an embodiment of the present application. This method can be applied in the above implementation environment. Please refer to Figure 1. The method includes the following steps.

Step 101: Obtain device information and obtain samples with annotation information uploaded by the user. The device information is information about the device selected by the user for running the character recognition model. The annotation information is used to describe the attributes of the characters in the corresponding sample.

The character recognition model is a network model trained by an electronic device. The character recognition model can be used to identify characters in objects to be recognized. For example, the character recognition model can be used to identify characters in images, or the character recognition model can be used to identify characters in images. Models can be used to recognize characters in videos and more.

The device information may be device ID (Identity, identification), device IP (Internet Protocol Address, Internet Protocol address), etc., and the electronic device can determine which device or devices the user selects to run the character recognition model based on the device information.

For example, the user can select a cloud device to run the character recognition model, in which case the electronic device can obtain the device information of the cloud device. The user can also select the edge device to run the character recognition model. In this case, the electronic device can obtain the device information of the edge device.

Among them, the computing power of cloud devices is usually strong, while the computing power of edge devices is usually weak. For example, the cloud device can be a server, a computer, etc., and the edge device can be a mobile phone, NVR (Network Video Recorder, Network Video Recorder), etc.

The samples uploaded by the user may be samples collected by the user based on the sample collection device, and the samples may be static images, dynamic images, videos, etc. This embodiment does not limit this. For example, the sample uploaded by the user includes characters. Correspondingly, as shown in Figure 2, the user can label the sample based on the characters in the sample to generate label information, which can be used to describe the characters in the sample. properties.

For example, the annotation information may include field positions, field categories, field rules, field content, etc., which is not limited in this embodiment. The electronic device can determine information related to the characters in the sample based on the annotation information.

Among them, the field position refers to the location of the area where the characters in the sample are located, the field category refers to the category to which the fields in the sample belong, the field rules refer to the composition rules of the characters in the fields in the sample, and the field content refers to The fields in the sample contain the contents of the characters.

For example, as shown in Figure 2, the field categories of this sample include carriage number and vehicle model. The field rule for carriage number is a fixed 7-digit number, the field content for vehicle model is "C62AK", and the field content for carriage number is "4503388" .

It should be noted that when the user collects samples based on the sample collection device, the user can adjust the angle and distance of the sample collection device so that the characters in the collected samples remain horizontal and the character height of the characters in the collected samples is greater than The 30 pixels and the character categories in the collected samples include all character categories that the user needs to recognize by the character recognition model. Furthermore, when the user completes sample collection, the user can upload the collected samples to the AI open platform.

That is, the electronic device can obtain the device information, so that it can determine which device or devices the user selects to run the character recognition model based on the device information. The electronic device can also obtain the sample with annotation information uploaded by the user, so that based on the annotation The information determines the properties of the characters in the sample.

As an example, the electronic device can also perform preprocessing on the samples uploaded by the user. For example, the electronic device can perform size normalization processing, filtering processing, smoothing processing, etc. on the samples uploaded by the user. In this embodiment, No restrictions.

Step 102: Based on the device information and annotation information, determine the network model to be trained from the network model set, which includes at least two network models.

The network models in the network model set may be pre-trained and then stored in the electronic device.

For example, the network models in the network model set may be trained based on the pre-training sample set. The samples in the pre-training sample set may include samples uploaded by users, samples collected by electronic devices, samples synthesized by electronic devices, samples downloaded from the cloud, etc. This embodiment is not limited to this.

It should be noted that different network models usually correspond to different pre-training sample sets. The different network models can be used to identify characters belonging to different field categories, or the different network models can be used to identify characters with different field rules. characters, or the different network models can be used to identify characters belonging to different field categories and with different field rules.

For example, when the characters in the samples in the pre-training sample set are carriage numbers and the network model is trained based on the pre-training sample set, the network model can be used to identify the carriage number. In the case where the characters in the samples in the pre-training sample set are 7-digit numbers and the network model is trained based on the pre-training sample set, the network model can be used to recognize 7-digit characters. When the characters in the samples in the pre-training sample set are carriage numbers and the carriage number is a 7-digit number, and the network model is trained based on the pre-training sample set, the network model can be used to identify the 7-digit carriage number. .

When the network model set includes multiple network models, the electronic device can select a model to be trained in the network model set that matches the device information and the annotation information based on the device information and the annotation information.

As an example, the annotation information includes the field categories and/or field rules of the characters in the corresponding samples. Based on the device information and annotation information, the implementation of determining the network model to be trained from the network model collection may include the following sub-steps:

1. According to the device information, select a network model subset from the network model collection, where the size and structure of the network models included in the network model subset match the device.

For example, the sizes and structures of the multiple network models included in the network model set may be different. For example, the network model set may include a network model with a larger size or a network model with a smaller size. The network model The collection can include network models with more complex structures or network models with simpler structures.

Generally, devices with stronger computing power can run network models with larger sizes and more complex structures, while devices with weaker computing power can run network models with smaller sizes and simpler structures. In this way, the electronic device can select a network model that matches the computing capability of the device from the network model collection based on the device information.

That is, if the device information indicates that the device is a device with weak computing power, the electronic device can select a network model with a smaller size and a simpler structure for the device; if the device information indicates that the device is a device with strong computing power, Then the electronic device can choose a network model with a larger size and a more complex structure for the device.

For example, if the device information indicates that the device is an edge device, the electronic device can select a network model with a smaller size and a simpler structure for the edge device. If the device information indicates that the device is a cloud device, the electronic device can select the network model for the edge device. For cloud devices, choose a network model with larger size and more complex structure.

2. According to the field categories and/or field rules of the samples uploaded by the user, determine the corresponding network model from the network model subset to obtain the network model to be trained.

For example, the field type may be ID number, mobile phone number, carriage number, car model, etc., which is not limited in this embodiment.

For example, the field rule may be that the field includes 7 digits, or the field rule may be that the field is composed of letters and numbers, or the field rule may be that the first character in the field is a capital letter, etc. This embodiment is suitable for This is not limited.

For example, there may be a corresponding relationship between field categories and network models, and there may also be a corresponding relationship between field rules and network models. For example, as shown in Table 1, the field category corresponding to network model 1 is vehicle model, the field category corresponding to network model 2 is carriage number, the field category corresponding to network model 3 is carriage number, and the field rule corresponding to network model 3 is carriage number. The number is a 6-digit number. The field categories corresponding to network model 4 are car model and compartment number. The field categories corresponding to network model 5 are car model and compartment number. The field rules corresponding to network model 5 are that the compartment number is 7 digits.

Table 1

network model Field category Field rules Network model 1 car model No rules Network model 2 Car number No rules Network model 3 Car number 6 digits Network model 4 Car model + carriage number No rules Network model 5 car model No rules Network model 5 Car number 7 digits

It should be noted that the correspondence between field categories and network models, and the correspondence between field rules and network models can be predetermined and then stored in the electronic device.

In the case where multiple network models are included in the network model subset, the electronic device may select a network model corresponding to the field category in the network model subset based on the field category, or the electronic device may select the network model corresponding to the field category in the network model subset based on the field rule. Select a network model corresponding to the field rule, or the electronic device can select a network model corresponding to the field category and field rule in a subset of network models based on the field category and field rule.

For example, as shown in Table 1, in the case where the network model subset includes network model 1, network model 2, network model 3, network model 4, and network model 5, if the annotation information indicates that the field category is vehicle model, Then the electronic device can select network model 1 as the network model to be trained. If the annotation information indicates that the field rule is a 6-digit number, the electronic device can select network model 3 as the network model to be trained. If the annotation information indicates that the field category is car model and compartment number, and the field rule is that the compartment number is a 7-digit number, then the electronic device can select network model 5 as the network model to be trained.

Step 103: Train the network model to be trained based on the samples uploaded by the user to obtain a character recognition model.

That is, after determining the samples uploaded by the user and the network model to be trained, the electronic device can train the network model to be trained based on the samples uploaded by the user, thereby obtaining a character recognition model.

As an example, the network model to be trained is trained based on the samples uploaded by the user, and the implementation of the character recognition model can include the following sub-steps:

1. Divide the samples uploaded by the user into a training sample set and a test sample set. The samples in the training sample set and the samples in the test sample set have the same data distribution rules.

Among them, the samples in the training sample set can be used to train the network model to be trained to obtain the trained network model. The samples in the test sample set can be used to test the trained network model to evaluate the character recognition effect of the trained network model.

The data distribution rule can be set according to the actual situation. The data distribution rule can be a rule related to character height, character category, field category, etc., which is not limited in this embodiment. For example, the same data distribution law may include that the character height of the samples in the training sample set is basically the same as the character height of the samples in the test sample set, and/or the same data distribution law may include that the character categories of the samples in the training sample set are basically the same as those in the test sample set. The character categories of the samples in the sample set are basically the same, and/or the same data distribution pattern may include the field categories of the samples in the training sample set being basically the same as the field categories of the samples in the test sample set, etc.

For example, methods for dividing samples uploaded by users may include hold-out, cross-validation, bootstrapping, etc., which are not limited in this embodiment.

That is, the electronic device can count the data distribution of the samples uploaded by the user, and then divide the samples uploaded by the user into a training sample set and a test sample set based on the data distribution, so that the data distribution pattern of the samples in the training sample set is consistent with the data distribution pattern of the samples in the test sample set. The data distribution rules are the same.

For example, if the data distribution rules are the same, including that the character height of the samples in the training sample set is basically the same as the character height of the samples in the test sample set, the electronic device can count the character heights of the samples uploaded by the user, so as to calculate the character height based on the character height. The samples uploaded by the user are divided into a training sample set and a test sample set, so that the character height of the samples in the training sample set is between 30-100 pixels, and the character height of the samples in the test sample set is also between 30-100 pixels. between.

For example, if the data distribution rules are the same, including that the character categories of the samples in the training sample set are basically the same as the character categories of the samples in the test sample set, the electronic device can count the character categories of the samples uploaded by the user, so as to classify the characters based on the character categories. The samples uploaded by the user are divided into a training sample set and a test sample set, so that the character categories of the samples in the training sample set are numbers, and the character categories of the samples in the test sample set are also numbers.

For example, if the data distribution rules are the same, including that the field categories of the samples in the training sample set are basically consistent with the field categories of the samples in the test sample set, the electronic device can count the field categories of the samples uploaded by the user, so as to classify the fields based on the field categories. The samples uploaded by the user are divided into a training sample set and a test sample set, so that the field category of the samples in the training sample set is the carriage number, and the field category of the samples in the test sample set is also the carriage number.

2. Train the network model to be trained multiple times based on the training sample set. After each training obtains the first network model, the current first network model is tested based on the test sample set to obtain test indicators. If the current test index of the first network model is higher than the test index of the first network model obtained in the previous training, continue training and testing. Until the test index of the current first network model is lower than the test index of the first network model obtained from the previous training, the training and testing are ended, and the first network model obtained from the previous training is determined as the character recognition model.

Usually, as the network model to be trained is trained, the character recognition effect of the first network model will gradually improve. However, when the training reaches a certain level, the character recognition effect of the first network model will gradually decrease. In this case, the electronic device can select the first network model with the best character recognition effect as the character recognition model based on the test index of the first network model.

That is, the electronic device can train the network model to be trained multiple times based on the training sample set, and test the first network model obtained after each training based on the test sample set to obtain the test index. If the current test index of the first network model is higher than the test index of the first network model obtained by the previous training, it means that the test index of the first network model obtained by the previous training is not optimal, that is, the test index obtained by the previous training The first network model is not the best network model for character recognition. In this case, the electronic device can continue to train and test the current first network model until the test index of the current first network model is lower than the test index of the first network model obtained by the previous training. The electronic device It can be determined that the test index of the first network model obtained by the previous training is optimal, that is, the first network model obtained by the previous training is the network model with the best character recognition effect. In this case, the electronic device can The first network model obtained through one training is determined as the character recognition model.

Of course, in a possible implementation, when the test index of the current first network model is lower than the test index of the first network model obtained by the previous training, the first network model obtained by the previous training can be The model is determined as a candidate character recognition model. Since the test index of the first network model may decrease briefly and then continue to increase, the electronic device may continue to train and test multiple times to further determine whether the test index of the candidate character recognition model is optimal. If the electronic device continues to train and test multiple times and the multiple test indicators obtained are lower than the test indicators of the candidate character recognition model, it can be determined that the test indicators of the candidate character recognition model are optimal, that is, the candidate character recognition model is The network model with the best character recognition effect. In this case, the electronic device can determine the candidate character recognition model as the character recognition model.

As an example, the current first network model includes a detection network and a recognition network. The current first network model is tested based on the test sample set. The implementation method of obtaining the test indicators can be: test the detection network based on the test sample set. , the test indicators of the detection network are obtained. The test indicators of the detection network are used to indicate the detection accuracy of the detection network. The recognition network is tested based on the test sample set to obtain the test indicators of the recognition network. The test indicators of the recognition network are used to indicate the character recognition accuracy of the recognition network. The test indicators of the detection network and the test indicators of the identification network are determined as the test indicators of the current first network model.

The detection network can be used to detect field positions and field categories in the sample. For example, the detection network can use a quadrilateral text detection algorithm, a general target detection algorithm, etc. This embodiment is not limited to this.

Among them, the recognition network can be used to identify the field content in the sample. For example, the recognition network can be a network based on CTC (Connectionist Temporal Classification, Connectionist Temporal Classification), a network combined with Attention Mechanism (attention mechanism), etc. This embodiment does not limit this.

For example, the electronic device can input the samples in the test sample set to the detection network and output the field positions and field categories, so that the test indicators of the detection network can be determined based on the field positions and field categories output by the detection network. The electronic device can input the field area corresponding to the field position to the recognition network and output the field content in the field area, so that the test indicators of the recognition network can be determined based on the field content output by the recognition network.

Among them, the test indicators of the detection network can be recall rate, accuracy rate, harmonic average of recall rate and accuracy rate, etc.

Among them, the accuracy rate refers to the proportion of the marked field positions detected by the detection network among all detected field positions, and the recall rate refers to the proportion of the marked field positions detected by the detection network among all the marked field positions. The proportion of field positions.

For example, if the test sample set includes 50 labeled field positions, and the detection network detects 40 field positions, of which 30 field positions are labeled field positions, the accuracy rate is 30/40=0.75, and the recall rate is 30/50=0.6.

It should be noted that when the test indicators of the detection network include multiple of the recall rate, accuracy rate, and the harmonic average of the recall rate and accuracy rate, the weight can be set for each indicator separately, so that based on the multiple indicators And the weight corresponding to each indicator is used to determine the test indicators of the detection network.

For example, when the test indicators of the detection network include recall and precision, you can set the weight of the recall to 0.6 and the weight of the precision to 0.4. If the recall is 0.6 and the precision is 0.75, then the detection network The test index is 0.6*0.6+0.75*0.4=0.66.

Among them, the test indicator of the recognition network can be the field recognition rate. The field recognition rate refers to the proportion of all fields that are correctly recognized by the recognition network. The correct recognition by the recognition network means that each character in the field is correctly recognized. Correctly recognized by the identified network.

That is, the electronic device can call the test sample set to test the detection network, that is, based on the character positions and field categories output by the detection network, and the character positions and field categories marked by the user, the test indicators of the detection network are obtained, thereby determining the detection network field detection accuracy. The electronic device can call the test sample set to test the recognition network, that is, based on the field content output by the recognition network and the field content marked by the user, the test indicators of the recognition network are obtained, thereby determining the character recognition accuracy of the recognition network.

It should be noted that if there are multiple field categories in the sample uploaded by the user, the current first network model can include multiple recognition networks, and each recognition network can be used to identify fields of the field category corresponding to the recognition network. content.

In a possible implementation, the electronic device can also use the test sample set to conduct an overall test on the current first network model, that is, input the test sample set into the current first network model, and output the field content, thereby according to the user's annotation The field content of , and the field content output by the current first network model determine the character recognition effect of the current first network model.

As an example, the electronic device can also perform the following operations: generate a test report. The test report includes one or more of test indicators for detecting the network, test indicators for identifying the network, and model feedback information. The model feedback information is used to instruct the user Update the uploaded sample and/or update the annotation information in the uploaded sample.

For example, when the user annotates field rules, if it is detected that the field content annotated by the user in a certain sample does not comply with the field rule, it can be determined that the user annotated the sample incorrectly. In this case, the electronic device can generate model feedback information to prompt the user that the sample is incorrectly labeled, so that the user can re-label the sample based on the model feedback information. Alternatively, if the character recognition model has a low field recognition rate for the field content of a certain field category, the electronic device can generate model feedback information to prompt the user to collect more samples including the field content of the field category.

That is, the electronic device can send one of the test indicators for detecting the network, the test indicator for identifying the network, and the model feedback information to the user as a test report, or the electronic device can send the test indicators for detecting the network, the test indicators for identifying the network, and the like. And multiple items in the model feedback information are sent to the user as a test report. In this way, the user can determine the character recognition effect of the character recognition model based on the test report, and update the sample based on the model feedback information.

As an example, before determining the first network model obtained by the previous training as the character recognition model, the electronic device may also perform the following operations: perform fixed-point quantization processing on the first network model obtained by the previous training, and obtain the fixed-point quantized The first network model. Determining the first network model obtained by the previous training as the character recognition model may be implemented by: determining the first network model after fixed-point quantization as the character recognition model.

Among them, fixed-point quantization processing refers to the process of converting the first network model obtained from the previous training from a floating-point model to a fixed-point model.

That is, the first network model obtained by the previous training is a floating point model. Since the device selected by the user to run the character recognition model may not support running the floating point model, the electronic device can use the first network model obtained by the previous training. A network model is subjected to fixed-point quantification processing to obtain a first network model after fixed-point quantification, so that the electronic device can determine the first network model after fixed-point quantification as a character recognition model.

Exemplarily, the electronic device can also send the character recognition model to the device selected by the user. In this way, the character recognition model can be run in the device selected by the user, so that the character recognition model can be used for uploading to the device by the user. Recognize the characters in the object to be recognized.

In the embodiment of the present application, the electronic device obtains the device information of the device selected by the user for running the character recognition model, and obtains the sample uploaded by the user with annotation information, where the annotation information can be used to describe the corresponding sample. Character properties. The electronic device can select the network model to be trained corresponding to the device information and the label information from the network model collection based on the device information and the label information. In this way, the character recognition model obtained by training the network model to be trained based on the samples uploaded by the user It is more suitable to run on the device selected by the user, and the character recognition model is more suitable for identifying characters in objects to be recognized that are similar to samples uploaded by users, that is, it can accurately identify characters in objects to be recognized that are similar to samples uploaded by users. The rate is higher.

Figure 3 is a schematic structural diagram of a device for determining a character recognition model according to an exemplary embodiment. The device for determining a character recognition model can be implemented by software, hardware, or a combination of both. The determining device of the character recognition model may include:

The acquisition module 310 is used to obtain device information and samples uploaded by the user with annotation information. The device information is information about the device selected by the user for running the character recognition model. The annotation information is used to describe the corresponding Properties of the characters in the sample;

Determining module 320, configured to determine a network model to be trained from a network model set based on the device information and the annotation information, where the network model set includes at least two network models;

The training module 330 is used to train the network model to be trained based on the samples uploaded by the user to obtain the character recognition model.

In a possible implementation of the present application, the annotation information includes field categories and/or field rules of characters in the corresponding sample, and the determination module 320 is used to:

According to the device information, select a network model subset from the network model set, wherein the size and structure of the network models included in the network model subset match the device;

According to the field categories and/or field rules of the samples uploaded by the user, the corresponding network model is determined from the network model subset to obtain the network model to be trained.

In a possible implementation of this application, the training module 330 is used to:

Divide the samples uploaded by the user into a training sample set and a test sample set, where the samples in the training sample set and the samples in the test sample set have the same data distribution rule;

Train the network model to be trained multiple times based on the training sample set;

After each training obtains the first network model, the current first network model is tested based on the test sample set to obtain test indicators;

If the current test index of the first network model is higher than the test index of the first network model obtained by the previous training, continue training and testing; until the current test index of the first network model is lower than the first test index obtained by the previous training. Test indicators of the network model, end training and testing, and determine the first network model obtained from the previous training as the character recognition model.

In a possible implementation of this application, the current first network model includes a detection network and a recognition network, and the training module 330 is used to:

Test the detection network based on the test sample set to obtain test indicators of the detection network, and the test indicators of the detection network are used to indicate the detection accuracy of the detection network;

Test the recognition network based on the test sample set to obtain test indicators of the recognition network, and the test indicators of the recognition network are used to indicate the character recognition accuracy of the recognition network;

The test indicators of the detection network and the test indicators of the identification network are determined as test indicators of the current first network model.

In a possible implementation of this application, the device further includes:

Generating module 340, configured to generate a test report. The test report includes one or more of the test indicators of the detection network, the test indicators of the identification network, and model feedback information. The model feedback information is used to indicate The user updates the uploaded sample and/or updates the annotation information in the uploaded sample.

In a possible implementation of this application, the training module 330 is also used to:

Perform fixed-point quantization processing on the first network model obtained from the previous training to obtain the first network model after fixed-point quantization;

Determining the first network model obtained from the previous training as the character recognition model includes:

The fixed-point quantized first network model is determined as the character recognition model.

In the embodiment of the present application, the electronic device obtains the device information of the device selected by the user for running the character recognition model, and obtains the sample uploaded by the user with annotation information, where the annotation information can be used to describe the corresponding sample. Character properties. The electronic device can select the network model to be trained corresponding to the device information and the label information from the network model collection based on the device information and the label information. In this way, the character recognition model obtained by training the network model to be trained based on the samples uploaded by the user It is more suitable to run on the device selected by the user, and the character recognition model is more suitable for identifying characters in objects to be recognized that are similar to samples uploaded by users, that is, it can accurately identify characters in objects to be recognized that are similar to samples uploaded by users. The rate is higher.

It should be noted that the device for determining the character recognition model provided in the above embodiment only uses the division of the above functional modules as an example to illustrate the determination of the character recognition model. In practical applications, the above functions can be allocated from different modules as needed. The functional modules are completed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device for determining the character recognition model provided in the above embodiments and the embodiment of the method for determining the character recognition model belong to the same concept. Please refer to the method embodiments for the specific implementation process, which will not be described again here.

FIG. 4 is a structural block diagram of an electronic device 400 provided by an embodiment of the present application. The electronic device 400 can be a portable mobile terminal, such as a smart phone, a tablet computer, an MP3 player (Moving Picture Experts Group Audio Layer III, moving picture experts compression standard audio layer 3), MP4 (Moving Picture Experts Group Audio Layer IV, moving picture expert) Expert compression of standard audio levels 4) players, laptops or desktop computers. The electronic device 400 may also be called a user device, a portable terminal, a laptop terminal, a desktop terminal, and other names.

Generally, the electronic device 400 includes: a processor 401 and a memory 402.

The processor 401 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 401 can adopt at least one hardware form among DSP (Digital Signal Processing, digital signal processing), FPGA (Field-Programmable Gate Array, field programmable gate array), and PLA (Programmable Logic Array, programmable logic array). accomplish. The processor 401 may also include a main processor and a co-processor. The main processor is a processor used to process data in the wake-up state, also called CPU (Central Processing Unit, central processing unit); the co-processor is A low-power processor used to process data in standby mode. In some embodiments, the processor 401 may be integrated with a GPU (Graphics Processing Unit, image processor), and the GPU is responsible for rendering and drawing content that needs to be displayed on the display screen. In some embodiments, the processor 401 may also include an AI (Artificial Intelligence, artificial intelligence) processor, which is used to process computing operations related to machine learning.

Memory 402 may include one or more computer-readable storage media, which may be non-transitory. Memory 402 may also include high-speed random access memory, and non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 402 is used to store at least one instruction, and the at least one instruction is used to be executed by the processor 401 to implement character recognition provided by the method embodiments in this application. How to determine the model.

Those skilled in the art can understand that the structure shown in FIG. 4 does not limit the electronic device 400, and may include more or fewer components than shown, or combine certain components, or adopt different component arrangements.

In some embodiments, a computer-readable storage medium is also provided, and a computer program is stored in the storage medium. When the computer program is executed by a processor, the steps of the method for determining the character recognition model in the above embodiments are implemented. For example, the computer-readable storage medium may be a magnetic disk or an optical disk, EEPROM (Electrically Erasable Programmable Read Only Memory), EPROM (Erasable Programmable Read-Only Memory) ), SRAM (Static RandomAccess Memory, static anytime access memory), ROM (Read Only Memory, read-only memory), magnetic memory, flash memory, PROM (Programmable Read-Only Memory, programmable read-only memory), etc.

It is worth noting that the computer-readable storage media mentioned in this application may be non-volatile storage media, in other words, may be non-transitory storage media.

It should be understood that all or part of the steps to implement the above embodiments can be implemented through software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. The computer instructions may be stored in the computer-readable storage medium described above.

That is, in some embodiments, a computer program product containing instructions is also provided, which when run on a computer causes the computer to perform the steps of the method for determining the character recognition model described above.

The above are optional embodiments provided for this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application shall be included in the protection of this application. within the range.

Claims (8)

1. A method of determining a character recognition model, the method comprising:

Acquiring equipment information and a sample with labeling information uploaded by a user, wherein the equipment information is information of equipment selected by the user and used for running a character recognition model, the labeling information is used for describing attributes of characters in the corresponding sample, and the labeling information comprises field types and/or field rules of the characters in the corresponding sample;

selecting a subset of network models from a set of network models based on the device information, the set of network models including at least two network models, the subset of network models including network models having dimensions and structures that match the device;

determining a corresponding network model from the network model subset according to the field category and/or the field rule of the sample uploaded by the user to obtain a network model to be trained;

and training the network model to be trained based on the sample uploaded by the user to obtain the character recognition model.

2. The method of claim 1, wherein the training the network model to be trained based on the user uploaded samples to obtain the character recognition model comprises:

dividing the samples uploaded by the user into a training sample set and a test sample set, wherein the samples in the training sample set and the samples in the test sample set have the same data distribution rule;

Training the network model to be trained for a plurality of times based on the training sample set;

after a first network model is obtained through each training, testing the current first network model based on the test sample set to obtain a test index;

if the current test index of the first network model is higher than the test index of the first network model obtained by the previous training, continuing training and testing; and ending the training and testing until the test index of the current first network model is lower than the test index of the first network model obtained by the previous training, and determining the first network model obtained by the previous training as the character recognition model.

3. The method of claim 2, wherein the current first network model includes a detection network and an identification network, and wherein testing the current first network model based on the test sample set to obtain the test index includes:

testing the detection network based on the test sample set to obtain a test index of the detection network, wherein the test index of the detection network is used for indicating the detection accuracy of the detection network;

testing the recognition network based on the test sample set to obtain test indexes of the recognition network, wherein the test indexes of the recognition network are used for indicating character recognition accuracy of the recognition network;

And determining the test index of the detection network and the test index of the identification network as the test index of the current first network model.

4. The method of claim 3, wherein prior to determining the first network model from the previous training as the character recognition model, further comprising:

performing fixed-point quantization processing on the first network model obtained in the previous training to obtain a first network model after fixed-point quantization;

the determining the first network model obtained by the previous training as the character recognition model comprises the following steps:

and determining the first network model after the fixed-point quantization as the character recognition model.

5. A device for determining a character recognition model, the device comprising:

the device comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring device information and a sample with labeling information, wherein the sample is uploaded by a user, the device information is information of a device selected by the user and used for running a character recognition model, the labeling information is used for describing attributes of characters in the corresponding sample, and the labeling information comprises field types and/or field rules of the characters in the corresponding sample;

a determining module, configured to select a subset of network models from a set of network models based on the device information, where the set of network models includes at least two network models, and a size and a structure of the network models included in the subset of network models match the device; determining a corresponding network model from the network model subset according to the field category and/or the field rule of the sample uploaded by the user to obtain a network model to be trained;

And the training module is used for training the network model to be trained based on the sample uploaded by the user to obtain the character recognition model.

6. The apparatus of claim 5, wherein the training module is to:

dividing the samples uploaded by the user into a training sample set and a test sample set, wherein the samples in the training sample set and the samples in the test sample set have the same data distribution rule;

training the network model to be trained for a plurality of times based on the training sample set;

after a first network model is obtained through each training, testing the current first network model based on the test sample set to obtain a test index;

if the current test index of the first network model is higher than the test index of the first network model obtained by the previous training, continuing training and testing; and ending the training and testing until the test index of the current first network model is lower than the test index of the first network model obtained by the previous training, and determining the first network model obtained by the previous training as the character recognition model.

7. An electronic device, comprising:

a processor;

A memory storing instructions executable by the processor;

wherein the processor is configured to execute the instructions and to carry out the steps of any one of the methods of claims 1-4.

8. A computer readable storage medium having instructions stored thereon, which when executed by a processor, implement the steps of the method of any of claims 1-4.