Underwriting method, underwriting device, electronic device, and storage medium

Abstract

This application provides an underwriting method, underwriting device, electronic device, and storage medium, belonging to the field of artificial intelligence technology and applied in the financial technology field. The method includes: acquiring a policy image and policy entry data of a target policy; performing data integrity verification on the policy image and policy entry data to obtain an integrity verification result; obtaining an image identifier of the policy image and a data identifier of the policy entry data based on the integrity verification result, and generating a policy recognition request for the target policy based on the image identifier and data identifier; calling a preset policy recognition model according to the policy recognition request, and performing text recognition on the policy image using the called policy recognition model to obtain reference policy data; performing a data consistency comparison between the policy entry data and the reference policy data using the called policy recognition model to obtain a data comparison result; and performing underwriting on the target policy based on the data comparison result. This application embodiment can improve underwriting efficiency.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applied to the field of financial technology, particularly to an underwriting method, underwriting device, electronic device, and storage medium. Background Technology

[0002] To facilitate the smooth operation of insurance business, it is necessary to review various policy data. The existing policy review process still relies heavily on manual methods, verifying the consistency between the data in the policy image files and the policy data entered into the system one by one. This verification process is time-consuming, resulting in low underwriting efficiency. Summary of the Invention

[0003] To overcome the problem of low underwriting efficiency in related technologies, the main objective of this application is to propose an underwriting method, underwriting device, electronic device, and storage medium, which aims to improve underwriting efficiency.

[0004] To achieve the above objectives, a first aspect of this application provides an underwriting method, the method comprising: Obtain the policy image and policy entry data of the target policy; The policy image and the policy data entered are subjected to data integrity verification to obtain an integrity verification result; wherein, the integrity verification result is used to indicate whether the policy image and the policy data are complete or incomplete; Based on the integrity verification result, obtain the image identifier of the policy image and the data identifier of the policy entry data, and generate a policy recognition request for the target policy based on the image identifier and the data identifier; The policy recognition request is used to call a preset policy recognition model, and the policy image is used to perform text recognition through the called policy recognition model to obtain reference policy data. The policy identification model that has been invoked is used to compare the consistency of the policy data and the reference policy data to obtain a data comparison result; wherein, the data comparison result is used to indicate whether the policy data and the reference policy data are consistent or inconsistent. The target policy is underwritten based on the data comparison results.

[0005] In some embodiments, the step of performing text recognition on the policy image using the invoked policy recognition model to obtain reference policy data includes: The policy image is denoised using a multispectral denoising network to obtain a denoised image; The denoised image is then subjected to surface correction to obtain a reference image; The reference policy data is obtained by performing text recognition on the reference image using the policy recognition model.

[0006] In some embodiments, the multispectral denoising network is trained according to the following steps: Acquire noisy policy images, reference spectral features, and clean policy images; The noisy policy image is subjected to feature extraction by an encoder to obtain sample policy coding features; wherein, the sample policy coding features include shared coding features; The sample policy encoding features are decoded by the first decoder to obtain the restored policy image and candidate image features; wherein, the candidate image features are output by the intermediate network layer of the first decoder. The shared encoded features and the candidate image features are decoded by a second decoder to obtain the predicted spectral features. The target loss is obtained by calculating the loss based on the noisy policy image, the reference spectral features, the predicted spectral features, the clean policy image, and the restored policy image; The encoder, the first decoder, and the second decoder are updated based on the target loss to obtain the multispectral denoising network.

[0007] In some embodiments, the step of calculating the target loss based on the noisy policy image, the reference spectral features, the predicted spectral features, the clean policy image, and the restored policy image includes: Feature extraction is performed on the restored policy image to obtain candidate spectral features; Perform a Fourier transform on the clean policy image to obtain the first frequency domain features, and perform a Fourier transform on the restored policy image to obtain the second frequency domain features. Generate a text mask for the noisy policy image, perform mask calculation on the clean policy image based on the text mask to obtain a first text feature, and perform mask calculation on the restored policy image based on the text mask to obtain a second text feature; The target loss is obtained by calculating the loss based on the reference spectral features, the predicted spectral features, the clean policy image, the restored policy image, the candidate spectral features, the first frequency domain features, the second frequency domain features, the first text features, and the second text features.

[0008] In some embodiments, the step of calculating the target loss based on the reference spectral features, the predicted spectral features, the clean policy image, the restored policy image, the candidate spectral features, the first frequency domain features, the second frequency domain features, the first text features, and the second text features includes: The loss is calculated based on the reference spectral features and the predicted spectral features to obtain the first loss; The loss is calculated based on the clean policy image and the restored policy image to obtain the second loss; The third loss is obtained by calculating the loss based on the candidate spectral features and the reference spectral features; The fourth loss is obtained by calculating the loss based on the first frequency domain feature and the second frequency domain feature; Based on the first and second text features, a loss is calculated to obtain the fifth loss; The target loss is obtained by weighted summation of the first loss, the second loss, the third loss, the fourth loss, and the fifth loss.

[0009] In some embodiments, performing surface correction on the denoised image to obtain a reference image includes: The denoised image is then subjected to perspective correction to obtain an initial image; The initial image is subjected to skeleton extraction to obtain the text line skeleton; Depth estimation is performed based on the text line skeleton to obtain the depth map of the denoised image; The reference image is obtained by performing coordinate mapping on the depth map.

[0010] In some embodiments, the policy entry data includes a policy entry field and the field entry values ​​of the policy entry field; the reference policy data includes a reference policy field and the reference field values ​​of the reference policy field; and the step of comparing the policy entry data and the reference policy data for data consistency using the invoked policy recognition model to obtain a data comparison result includes: The policy recognition model is used to compare the consistency of the policy entry fields and the reference policy fields to obtain the field comparison results. The policy recognition model is used to compare the value entered in the field with the value of the reference field to obtain the value comparison result. The field comparison results and the value comparison results are integrated to obtain the data comparison results.

[0011] To achieve the above objectives, a second aspect of this application provides an underwriting device, the device comprising: The acquisition module is used to acquire the policy image and policy entry data of the target policy; The verification module is used to perform data integrity verification on the policy image and the policy entry data, and obtain an integrity verification result; wherein, the integrity verification result is used to indicate whether the policy image and the policy entry data are complete or incomplete; The request module is used to obtain the image identifier of the policy image and the data identifier of the policy entry data according to the integrity verification result, and generate a policy recognition request for the target policy according to the image identifier and the data identifier. The recognition module is used to call a preset policy recognition model according to the policy recognition request, and to perform text recognition on the policy image through the called policy recognition model to obtain reference policy data; The comparison module is used to perform a data consistency comparison between the policy entry data and the reference policy data using the invoked policy recognition model, and obtain a data comparison result; wherein, the data comparison result is used to indicate whether the policy entry data and the reference policy data are consistent or inconsistent; The underwriting module is used to underwrite the target policy based on the data comparison results.

[0012] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0013] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0014] The underwriting method, underwriting device, electronic device, and computer-readable storage medium proposed in this application acquire the policy image and policy entry data of the target policy to determine whether the content of the policy image is consistent with the policy entry data. Data integrity verification is performed on the policy image and policy entry data to obtain the integrity verification result. By introducing a data integrity triggering mechanism, invalid model calls can be prevented, improving the reliability of model calls. Based on the integrity verification result, the image identifier of the policy image and the data identifier of the policy entry data are obtained, and a policy recognition request for the target policy is generated based on the image identifier and data identifier, so as to perform model calls based on the policy recognition request. Since both the image identifier and the data identifier are unique identifiers, generating the policy recognition request based on the unique identifier can achieve state isolation and result merging in scenarios where the same policy is uploaded multiple times, ensuring underwriting stability. The system automatically verifies the consistency between policy images and entered policy data by calling a pre-defined policy recognition model based on the policy recognition request. This process obtains reference policy data, and the system then compares the entered policy data with the reference policy data using the same model to determine the comparison results. This automated verification solves the problem of time-consuming manual verification. Underwriting is then performed on the target policy based on the comparison results, improving underwriting efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart of the underwriting method provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart of step S140 in the middle; Figure 3 yes Figure 2 The flowchart of the training process of the multispectral denoising network in step S210; Figure 4 yes Figure 3 The flowchart of step S350 in the text; Figure 5 yes Figure 4 The flowchart of step S440 in the middle; Figure 6 yes Figure 2 The flowchart of step S220 in the text; Figure 7 yes Figure 1 The flowchart of step S150 in the middle; Figure 8 This is a schematic diagram of the underwriting device provided in the embodiments of this application; Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0019] To facilitate the smooth operation of insurance business, it is necessary to review various policy data. The existing policy review process still relies heavily on manual methods, verifying the consistency between the data in the policy image files and the policy data entered into the system one by one. This verification process is time-consuming, resulting in low underwriting efficiency.

[0020] Based on this, embodiments of this application provide an underwriting method, underwriting device, electronic device, and computer-readable storage medium, aiming to improve underwriting efficiency.

[0021] The underwriting method, underwriting device, electronic device, and computer-readable storage medium provided in this application are specifically described through the following embodiments. First, the underwriting method in this application embodiment is described.

[0022] The underwriting method provided in this application relates to the field of artificial intelligence technology. The underwriting method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the underwriting method, but is not limited to the above forms.

[0023] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0024] Figure 1 This is an optional flowchart of the underwriting method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S160.

[0025] Step S110: Obtain the policy image and policy entry data of the target policy; Step S120: Perform data integrity verification on the policy image and the policy entry data to obtain the integrity verification result; wherein, the integrity verification result is used to indicate whether the policy image and the policy entry data are complete or incomplete. Step S130: Obtain the image identifier of the policy image and the data identifier of the policy entry data based on the integrity verification result, and generate a policy recognition request for the target policy based on the image identifier and the data identifier. Step S140: Call the preset policy recognition model according to the policy recognition request, and perform text recognition on the policy image through the called policy recognition model to obtain reference policy data; Step S150: The policy entry data and the reference policy data are compared using the invoked policy recognition model to obtain the data comparison result; wherein, the data comparison result is used to indicate whether the policy entry data and the reference policy data are consistent or inconsistent. Step S160: Underwrite the target policy based on the data comparison results.

[0026] Steps S110 to S160 as shown in the embodiments of this application automatically verify the consistency between the policy image and the policy data by calling the policy recognition model, which solves the problem of long time consumption of manual verification and improves underwriting efficiency.

[0027] In step S110 of some embodiments, the target policy is the policy to be underwritten. Image data of the target policy is acquired through a scanner, mobile terminal camera, or upload interface to obtain a policy image. The entry information of the target policy in the system is then obtained to obtain policy entry data. The system can be an insurance business system, a policy quality inspection platform, etc.

[0028] Policy images can exist in PDF or JPG format. The policy image is transferred to the database, generating an image identifier, denoted as `file_key_one`. The policy entry data is structured data, containing data items such as the entrusted amount, management fee percentage, and handling service fee. This policy entry data is converted to a JSON file and stored in the database, generating a data identifier, denoted as `file_key_two`. The image identifier and data identifier are written to the request record table for subsequent use. The image identifier is a unique identifier for the policy image, and the data identifier is a unique identifier for the policy entry data.

[0029] In step S120 of some embodiments, a data integrity triggering mechanism is introduced to ensure data completeness before AI invocation, thereby guaranteeing the reliability of AI invocation. Data integrity is verified on the policy image and policy entry data. If both the policy image and the policy entry data have image identifiers, the integrity verification result is determined to be that the policy image and policy entry data are complete; if neither the policy image nor the policy entry data has an image identifier, the integrity verification result is determined to be that the policy image and policy entry data are incomplete. It should be noted that if the policy image lacks an image identifier, the policy image is incomplete; similarly, if the policy entry data lacks a data identifier, the policy entry data is incomplete.

[0030] In step S130 of some embodiments, if the integrity verification result indicates that the policy image and the policy entry data are complete, then the image identifier of the policy image and the data identifier of the policy entry data are obtained, and a policy recognition request for the target policy is generated based on the image identifier, the data identifier, and the request ID. The policy recognition request is used to request the invocation of the AI ​​model.

[0031] If the integrity check result indicates that the policy image is incomplete or the policy data is incomplete, the data completion process will be triggered.

[0032] The process involves sending a policy recognition request to a target queue, retrieving the request from the queue, and then invoking a policy recognition model based on the retrieved request. The model is then used to perform text recognition on the policy image to obtain reference policy data. When invoking the policy recognition model, it checks if a similar policy recognition request is already being processed. If so, it prioritizes processing the existing request to avoid concurrency conflicts. This achieves state isolation and result merging in scenarios where the same policy is uploaded multiple times, ensuring underwriting stability.

[0033] Please see Figure 2 In some embodiments, step S140 may include, but is not limited to, steps S210 to S230: Step S210: Perform multispectral denoising on the policy image using a multispectral denoising network to obtain a denoised image; Step S220: Perform surface correction on the denoised image to obtain a reference image; Step S230: Use the policy recognition model to perform text recognition on the reference image to obtain reference policy data.

[0034] In step S210 of some embodiments, image quality directly determines text recognition accuracy. To improve image quality, the policy image is input into a multispectral denoising network for multispectral denoising to obtain a denoised image. The multispectral denoising network uses a cross-spectral attention mechanism to take spectral features as priors, guiding the RGB channels to reconstruct a clean signal. When a paper policy is scanned or photographed, it is simultaneously illuminated by visible light and near-infrared (NIR), resulting in RGB+NIR four-channel data. The multispectral denoising network utilizes these four images of the same scene at different wavelengths to denoise on a signal-to-spatial scale rather than simply on a spatial scale, achieving high-quality denoising results.

[0035] In step S220 of some embodiments, it is almost impossible for the policy to be perfectly flat when it is photographed or scanned. Once curling, creases, perspective or trapezoidal deformation occurs, the text lines will exhibit non-linear geometric distortion. In order to ensure geometric consistency, surface correction is performed on the denoised image to obtain a reference image.

[0036] In step S230 of some embodiments, a reference image is input into a policy recognition model for text recognition to convert image data into text data, thereby obtaining reference policy data. The policy recognition model may employ a gated convolutional neural network, a visual transformer model, or the like.

[0037] Through the above steps S210 to S230, accurate text recognition can be performed on the image.

[0038] Please see Figure 3 In some embodiments, the training process of the multispectral denoising network in step S210 may include, but is not limited to, steps S310 to S360: Step S310: Obtain the noisy policy image, reference spectral features, and clean policy image; Step S320: Extract features from the noisy policy image using an encoder to obtain sample policy coding features; wherein, the sample policy coding features include shared coding features; Step S330: The sample policy encoding features are decoded by the first decoder to obtain the restored policy image and candidate image features; wherein, the candidate image features are output by the intermediate network layer of the first decoder. Step S340: The shared coded features and candidate image features are decoded by the second decoder to obtain the predicted spectral features; Step S350: Calculate the loss based on the noisy policy image, reference spectral features, predicted spectral features, clean policy image, and restored policy image to obtain the target loss; Step S360: Update the encoder, the first decoder, and the second decoder according to the target loss to obtain the multispectral denoising network.

[0039] In step S310 of some embodiments, a reference spectral feature matching the noisy policy image and a clean policy image are obtained. The noisy policy image and the clean policy image are images for the same policy; the noisy policy image contains noise, while the clean policy image does not contain any noise. The reference spectral feature is the near-infrared (NIR) channel component of the clean policy image.

[0040] In step S320 of some embodiments, the noisy policy image is input into a preset model, which includes an encoder, a first decoder, and a second decoder. The encoder extracts features from the noisy policy image to obtain sample policy encoded features, which include shared encoded features and skip connection features. The skip connection features are a four-level multi-scale feature list, used by the first decoder for skip connections.

[0041] In step S330 of some embodiments, the first decoder is used to provide the denoised RGB image. The first decoder performs feature decoding on the shared coding features and skip connection features to obtain the restored policy image and candidate image features. The candidate image features are output from the intermediate network layer of the first decoder, obtained by adding or concatenating the original output of the intermediate network layer with skip connection features of the same scale.

[0042] In step S340 of some embodiments, the second decoder is used for NIR reconstruction. The second decoder performs feature decoding on the shared coded features and candidate image features to reconstruct the NIR vector and obtain the predicted spectral features. The first decoder and the second decoder share the same shared coded features.

[0043] In step S350 of some embodiments, in order to measure the difference between the reference spectral features and the predicted spectral features, and the difference between the clean policy image and the restored policy image, a loss calculation is performed based on the noisy policy image, the reference spectral features, the predicted spectral features, the clean policy image, and the restored policy image to obtain a target loss, so as to guide the model training process and optimize the model parameters based on the target loss.

[0044] In step S360 of some embodiments, the target loss is minimized, and the model parameters of the encoder, the first decoder, and the second decoder are updated synchronously to obtain a multispectral denoising network.

[0045] It should be noted that when using a multispectral denoising network to denoise the policy image, the encoder can extract features from the policy image, and the first decoder can decode the extracted features to obtain the denoised image.

[0046] Steps S310 to S360 above optimize the model parameters, thereby improving the model's denoising effect on the image.

[0047] Please see Figure 4 In some embodiments, step S350 may include, but is not limited to, steps S410 to S440: Step S410: Extract features from the restored policy image to obtain candidate spectral features; Step S420: Perform Fourier transform on the clean policy image to obtain the first frequency domain features, and perform Fourier transform on the restored policy image to obtain the second frequency domain features. Step S430: Generate a text mask for the noisy policy image; perform mask calculation on the clean policy image based on the text mask to obtain the first text feature; perform mask calculation on the restored policy image based on the text mask to obtain the second text feature. Step S440: Calculate the loss based on the reference spectral features, predicted spectral features, clean policy image, restored policy image, candidate spectral features, first frequency domain features, second frequency domain features, first text features, and second text features to obtain the target loss.

[0048] In step S410 of some embodiments, the restored policy image is feature-encoded by an encoder to obtain restored coded features, and the restored coded features are feature-decoded by a first decoder to obtain restored decoded features. The shared features and restored decoded features in the restored coded features are then feature-decoded by a second decoder to obtain candidate spectral features.

[0049] In step S420 of some embodiments, a Fourier transform is performed on the clean policy image to obtain a first frequency domain feature, and a Fourier transform is performed on the restored policy image to obtain a second frequency domain feature.

[0050] In step S430 of some embodiments, the noisy policy image is semantically segmented using the layout-SSM model to obtain probability maps of the background, printed text, and handwritten text. Opening and dilation operations are performed on the probability maps of the printed text and the handwritten text to generate a text mask for the noisy policy image. The text mask is multiplied by the clean policy image to obtain the first text feature, and the text mask is multiplied by the restored policy image to obtain the second text feature.

[0051] In step S440 of some embodiments, in order to guide the model training process, loss is calculated based on reference spectral features, predicted spectral features, clean policy image, restored policy image, candidate spectral features, first frequency domain features, second frequency domain features, first text features and second text features to obtain the target loss.

[0052] Through the above steps S410 to S440, the loss value of model training can be obtained.

[0053] Please see Figure 5 In some embodiments, step S440 may include, but is not limited to, steps S510 to S560: Step S510: Calculate the loss based on the reference spectral features and the predicted spectral features to obtain the first loss; Step S520: Calculate the loss based on the clean policy image and the restored policy image to obtain the second loss; Step S530: Calculate the loss based on the candidate spectral features and the reference spectral features to obtain the third loss; Step S540: Calculate the loss based on the first frequency domain features and the second frequency domain features to obtain the fourth loss; Step S550: Calculate the loss based on the first and second character features to obtain the fifth loss; Step S560: The first loss, second loss, third loss, fourth loss and fifth loss are weighted and summed to obtain the target loss.

[0054] In step S510 of some embodiments, the feature data difference between the reference spectral features and the predicted spectral features is calculated to obtain a first loss. The formula for calculating the first loss is expressed as: , in, This indicates the first loss; E represents the second decoder, which is used to output predicted spectral features; E represents the encoder. This represents a noisy insurance policy image; 1 represents the reference spectral characteristics; 1 represents the L1 loss function.

[0055] In step S520 of some embodiments, for each pixel of the clean policy image, the pixel data difference between the pixel value of the clean policy image at that pixel and the pixel value of the corresponding pixel in the restored policy image is calculated to obtain a second loss. The formula for calculating the second loss is expressed as: , in, Indicates the second loss; This refers to the first decoder, which is used to output the restored policy image; This represents a clean policy image.

[0056] In step S530 of some embodiments, the feature data difference between the candidate spectral features and the reference spectral features is calculated to obtain a third loss. The formula for calculating the third loss is expressed as:

[0057] in, Indicates the third loss; This indicates the second decoder, which is used to output candidate spectral features.

[0058] In step S540 of some embodiments, to suppress over-smoothing and preserve the high-frequency energy of text edges, the feature difference between the first frequency domain feature and the second frequency domain feature is calculated to obtain a fourth loss. The formula for calculating the fourth loss is expressed as follows: , in, represents the fourth loss; FFT represents the Fourier transform.

[0059] In step S550 of some embodiments, the multi-scale structural similarity (MS-SSIM) of the first and second text features is calculated. The difference between 1 and the MS-SSIM is then subtracted to obtain the fifth loss. The formula for calculating the fifth loss is as follows: , in, The fifth loss is represented by MS-SSIM; multi-scale structural similarity is represented by char_mask; This indicates element-wise multiplication.

[0060] In step S560 of some embodiments, the first loss, second loss, third loss, fourth loss, and fifth loss are weighted and summed to obtain the target loss. The formula for calculating the target loss is expressed as: , Where L represents the target loss; , , , , All are weighting coefficients.

[0061] Through the above steps S510 to S560, the target loss can be obtained, and the model parameters can be optimized based on the target loss.

[0062] Please see Figure 6 In some embodiments, step S220 may include, but is not limited to, steps S610 to S640: Step S610: Perform perspective correction on the denoised image to obtain the initial image; Step S620: Extract the skeleton from the initial image to obtain the text line skeleton; Step S630: Depth estimation is performed based on the text line skeleton to obtain the depth map of the denoised image; Step S640: Perform coordinate mapping on the depth map to obtain a reference image.

[0063] In step S610 of some embodiments, edge detection is performed on the denoised image to obtain multiple edges. The longest and most closed contour is selected from these edges to obtain a reference boundary. Small bumps are continuously removed using polygon approximation until only four vertices remain of the reference boundary. A perspective transformation algorithm is then used to perform perspective coarse correction on the four vertices to obtain the initial image.

[0064] In step S620 of some embodiments, the initial image is semantically segmented using the Layout-SSM model to divide the text region and background region, resulting in an intermediate image. The intermediate image is then skeletonized to obtain multiple thin lines. These thin lines are fitted into line segments using a probabilistic Hough transform to obtain the text line skeleton. The text line skeleton is a set of two-dimensional polylines.

[0065] In step S630 of some embodiments, a quadratic curve fitting is performed on each pixel of the text line skeleton, and the height of the deviation from the straight line is taken as the depth to obtain the depth map of the denoised image. The formula for quadratic curve fitting is: c , Where x and y represent the input and output, respectively; b and c represent curve coefficients.

[0066] The formula for calculating depth is expressed as: , Where D represents depth; and straight_line represents the height of the straight line.

[0067] In step S640 of some embodiments, the depth map is mapped to coordinates to unfold the surface into a plane, resulting in a flattened image. The flattened image is then skeletonized to obtain multiple line segments. The standard deviation of the slope for all line segments is calculated. If the standard deviation of the slope is less than a preset threshold, the flattened image is used as a reference image. The preset threshold can be set according to actual conditions, such as 0.02. If the standard deviation of the slope is greater than or equal to the preset threshold, the flattened image is used as the initial image, and steps S620 to S640 are iteratively executed repeatedly until the standard deviation of the slope is less than the preset threshold.

[0068] Through the above steps S610 to S640, geometric distortions in the image can be eliminated, improving the accuracy of text recognition.

[0069] Please see Figure 7 In some embodiments, step S150 may include, but is not limited to, steps S710 to S730: Step S710: The policy entry fields and the reference policy fields are compared using the policy recognition model to obtain the field comparison results. Step S720: The value of the field input and the value of the reference field are compared by the policy recognition model to obtain the value comparison result; Step S730: Integrate the field comparison results and value comparison results to obtain the data comparison results.

[0070] In step S710 of some embodiments, the policy entry data includes a policy entry field and the field entry value of the policy entry field, and the reference policy data includes a reference policy field and the reference field value of the reference policy field. A field existence comparison is performed between the policy entry field and the reference policy field using a policy recognition model. If some characters in the policy entry field exist in the reference policy field, the field existence comparison result is 1; otherwise, it is 0. A field semantic comparison is also performed between the policy entry field and the reference policy field using the policy recognition model. If the semantics are the same, the field semantic comparison result is 1; otherwise, it is 0. The field existence comparison result and the field semantic comparison result are weighted to obtain the field comparison result.

[0071] In step S720 of some embodiments, the value of the field input and the value of the reference field are compared by the policy recognition model. If the two values ​​are consistent, the value comparison result is 1; otherwise, it is 0.

[0072] In step S730 of some embodiments, a weighted calculation is performed on the field comparison results and value comparison results to obtain a consistency score, and the data comparison result is determined based on the consistency score. The data comparison result is used to indicate whether the policy entry data and the reference policy data are consistent or inconsistent. If the consistency score is greater than or equal to a preset threshold, the data comparison result is determined to be consistent between the policy entry data and the reference policy data; conversely, if the consistency score is less than the preset threshold, the data comparison result is determined to be inconsistent between the policy entry data and the reference policy data.

[0073] The formula for calculating the consistency score is as follows: , Where C represents the consistency score; ES and This indicates the existence comparison results of fields and their weights; SS and This represents the semantic comparison results of fields and their weights; VS and This represents the comparison results and their weights.

[0074] Through the above steps S710 to S730, automated consistency comparison can be achieved, improving underwriting efficiency.

[0075] In step S160 of some embodiments, if the data comparison result indicates that the policy entry data and the reference policy data are consistent, then the target policy is further underwritten. If the data comparison result indicates that the policy entry data and the reference policy data are inconsistent, then the policy entry data is manually reviewed.

[0076] Please see Figure 8 This application also provides an underwriting device that can implement the above-described underwriting method. The underwriting device includes: The acquisition module 810 is used to acquire the policy image and policy entry data of the target policy; The verification module 820 is used to perform data integrity verification on the policy image and the policy entry data, and obtain the integrity verification result; wherein, the integrity verification result is used to indicate whether the policy image and the policy entry data are complete or incomplete; The request module 830 is used to obtain the image identifier of the policy image and the data identifier of the policy entry data based on the integrity verification result, and to generate a policy recognition request for the target policy based on the image identifier and the data identifier. The recognition module 840 is used to call a preset policy recognition model according to the policy recognition request, and to perform text recognition on the policy image through the called policy recognition model to obtain reference policy data. The comparison module 850 is used to compare the policy entry data and the reference policy data for data consistency using the invoked policy recognition model, and obtain the data comparison result; wherein, the data comparison result is used to indicate whether the policy entry data and the reference policy data are consistent or inconsistent; The underwriting module 860 is used to underwrite the target policy based on the data comparison results. The specific implementation of this underwriting device is basically the same as the specific embodiment of the underwriting method described above, and will not be repeated here.

[0077] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned underwriting method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0078] Please see Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 910 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 920 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 920 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 920 and is called and executed by the processor 910 using the underwriting method of the embodiments of this application. The input / output interface 930 is used to implement information input and output; The communication interface 940 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 950 transmits information between various components of the device (e.g., processor 910, memory 920, input / output interface 930, and communication interface 940); The processor 910, memory 920, input / output interface 930 and communication interface 940 are connected to each other within the device via bus 950.

[0079] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described underwriting method.

[0080] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0081] The underwriting method, underwriting device, electronic device, and computer storage medium provided in this application embodiment automatically verify the consistency between the policy image and the policy data by calling the policy recognition model, which solves the problem of long time consumption of manual verification and improves underwriting efficiency.

[0082] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0083] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0084] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0086] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0087] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0089] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0091] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An underwriting method, characterized in that, The method includes: Obtain the policy image and policy entry data of the target policy; The policy image and the policy data entered are subjected to data integrity verification to obtain an integrity verification result; wherein, the integrity verification result is used to indicate whether the policy image and the policy data are complete or incomplete; Based on the integrity verification result, obtain the image identifier of the policy image and the data identifier of the policy entry data, and generate a policy recognition request for the target policy based on the image identifier and the data identifier; The policy recognition request is used to call a preset policy recognition model, and the policy image is used to perform text recognition through the called policy recognition model to obtain reference policy data. The policy identification model that has been invoked is used to compare the consistency of the policy data and the reference policy data to obtain a data comparison result; wherein, the data comparison result is used to indicate whether the policy data and the reference policy data are consistent or inconsistent. The target policy is underwritten based on the data comparison results.

2. The method according to claim 1, characterized in that, The step of performing text recognition on the policy image using the invoked policy recognition model to obtain reference policy data includes: The policy image is denoised using a multispectral denoising network to obtain a denoised image; The denoised image is then subjected to surface correction to obtain a reference image; The reference policy data is obtained by performing text recognition on the reference image using the policy recognition model.

3. The method according to claim 2, characterized in that, The multispectral denoising network is trained according to the following steps: Acquire noisy policy images, reference spectral features, and clean policy images; The noisy policy image is subjected to feature extraction by an encoder to obtain sample policy coding features; wherein, the sample policy coding features include shared coding features; The sample policy encoding features are decoded by the first decoder to obtain the restored policy image and candidate image features; wherein, the candidate image features are output by the intermediate network layer of the first decoder. The shared encoded features and the candidate image features are decoded by a second decoder to obtain the predicted spectral features. The target loss is obtained by calculating the loss based on the noisy policy image, the reference spectral features, the predicted spectral features, the clean policy image, and the restored policy image; The encoder, the first decoder, and the second decoder are updated based on the target loss to obtain the multispectral denoising network.

4. The method according to claim 3, characterized in that, The step of calculating the target loss based on the noisy policy image, the reference spectral features, the predicted spectral features, the clean policy image, and the restored policy image includes: Feature extraction is performed on the restored policy image to obtain candidate spectral features; Perform a Fourier transform on the clean policy image to obtain the first frequency domain features, and perform a Fourier transform on the restored policy image to obtain the second frequency domain features. Generate a text mask for the noisy policy image, perform mask calculation on the clean policy image based on the text mask to obtain a first text feature, and perform mask calculation on the restored policy image based on the text mask to obtain a second text feature; The target loss is obtained by calculating the loss based on the reference spectral features, the predicted spectral features, the clean policy image, the restored policy image, the candidate spectral features, the first frequency domain features, the second frequency domain features, the first text features, and the second text features.

5. The method according to claim 4, characterized in that, The step of calculating the target loss based on the reference spectral features, the predicted spectral features, the clean policy image, the restored policy image, the candidate spectral features, the first frequency domain features, the second frequency domain features, the first text features, and the second text features includes: The loss is calculated based on the reference spectral features and the predicted spectral features to obtain the first loss; The loss is calculated based on the clean policy image and the restored policy image to obtain the second loss; The third loss is obtained by calculating the loss based on the candidate spectral features and the reference spectral features; The fourth loss is obtained by calculating the loss based on the first frequency domain feature and the second frequency domain feature; Based on the first and second text features, a loss is calculated to obtain the fifth loss; The target loss is obtained by weighted summation of the first loss, the second loss, the third loss, the fourth loss, and the fifth loss.

6. The method according to claim 2, characterized in that, The step of performing surface correction on the denoised image to obtain a reference image includes: The denoised image is then subjected to perspective correction to obtain an initial image; The initial image is subjected to skeleton extraction to obtain the text line skeleton; Depth estimation is performed based on the text line skeleton to obtain the depth map of the denoised image; The reference image is obtained by performing coordinate mapping on the depth map.

7. The method according to any one of claims 1 to 6, characterized in that, The policy entry data includes policy entry fields and field entry values ​​for those fields. The reference policy data includes reference policy fields and reference field values ​​for those fields. The process involves comparing the policy entry data and the reference policy data using the invoked policy recognition model to obtain a data comparison result, including: The policy recognition model is used to compare the consistency of the policy entry fields and the reference policy fields to obtain the field comparison results. The policy recognition model is used to compare the value entered in the field with the value of the reference field to obtain the value comparison result. The field comparison results and the value comparison results are integrated to obtain the data comparison results.

8. An underwriting device, characterized in that, The device includes: The acquisition module is used to acquire the policy image and policy entry data of the target policy; The verification module is used to perform data integrity verification on the policy image and the policy entry data, and obtain an integrity verification result; wherein, the integrity verification result is used to indicate whether the policy image and the policy entry data are complete or incomplete; The request module is used to obtain the image identifier of the policy image and the data identifier of the policy entry data according to the integrity verification result, and generate a policy recognition request for the target policy according to the image identifier and the data identifier. The recognition module is used to call a preset policy recognition model according to the policy recognition request, and to perform text recognition on the policy image through the called policy recognition model to obtain reference policy data; The comparison module is used to perform a data consistency comparison between the policy entry data and the reference policy data using the invoked policy recognition model, and obtain a data comparison result; wherein, the data comparison result is used to indicate whether the policy entry data and the reference policy data are consistent or inconsistent; The underwriting module is used to underwrite the target policy based on the data comparison results.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images