An intelligent identification-based insurance policy structured storage method, medium and system

Abstract

The application provides a policy structured storage method, medium and system based on intelligent identification, belongs to the technical field of policy structured storage, and comprises the following steps: S10, acquiring a policy image; S20, pre-processing the policy image to obtain a first image; S30, performing character recognition on the first image to obtain a text set; S40, performing structured processing on the obtained text set according to a pre-set contract form database to obtain a policy object; and S50, storing the policy object in a database in a key-value pair mode. The method, medium and system can improve the accuracy of policy information extraction, realize the structured expression of current policy content, support text direction correction and the recognition of different personnel handwriting fonts, and can solve the problems of slow recognition speed, low precision and inaccurate result matching of the current OCR in the policy recognition system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of policy structured storage, and particularly relates to a policy structured storage method, medium and system based on intelligent identification. BACKGROUND

[0002] Optical Character Recognition (OCR) refers to that an electronic device checks printed characters, converts characters in a paper document into a black-and-white dot matrix image file in an optical manner, and converts the characters in the image into a text format through an identification system, so as to provide the characters for further editing and processing by a word processing software. A policy, that is, a written proof of an insurance contract signed by an insurer and an insured, must clearly and completely record the rights and obligations of the insurance parties, mainly including the names of the insurer and the insured, the insurance subject, the insurance amount, the insurance premium, the insurance period, the scope of compensation or payment, and the like. With the continuous improvement of informationization and intelligentization of insurance and insurance IT systems, how to store, identify and analyze policy information for users and recommend perfect insurance content for users has become a standard configuration of the IT system. At present, in the policy identification system, the OCR has the problems of slow recognition speed, low precision, and inaccurate result matching. SUMMARY

[0003] Therefore, the application provides a policy structured storage method, medium and system based on intelligent identification, which can solve the problems of slow recognition speed, low precision and inaccurate result matching of the OCR in the policy identification system.

[0004] The application is implemented as follows:

[0005] The first aspect of the application provides a policy structured storage method based on intelligent identification, which comprises the following steps:

[0006] S10, acquiring a policy image;

[0007] S20, pre-processing the policy image to obtain a first image;

[0008] S30, performing character recognition on the first image to obtain a text set;

[0009] S40, performing structured processing on the obtained text set according to a pre-set contract form database to obtain a policy object;

[0010] S50, storing the policy object in a database in a key-value pair mode.

[0011] The step of performing structured processing on the obtained text set according to the pre-set contract form database to obtain the policy object specifically comprises the following steps:

[0012] On the basis of the above technical solutions, the policy structured storage method based on intelligent identification can be further improved as follows:

[0013] The step of preprocessing the policy image specifically includes grayscale processing, median filtering and binaryzation processing.

[0014] The step of performing text recognition on the first image to obtain the text set specifically includes:

[0015] S31, segmenting the first image to obtain a plurality of segmented images to form a segmented image set;

[0016] S32, performing text detection on each segmented image in the segmented image set to obtain a text box corresponding to each segmented image, denoted as a segmented image text box;

[0017] S33, performing text direction detection on each segmented image text box to obtain the direction of the segmented image text box;

[0018] S34, performing text recognition on the segmented image text box according to the detection direction to obtain segmented image text;

[0019] S35, combining the obtained segmented image text into a text set according to the order of the segmented images.

[0020] Further, the step of performing text recognition on the segmented image text box according to the detection direction to obtain segmented image text specifically includes:

[0021] Step 1, direction labeling is performed on the segmented image text box;

[0022] Step 2, the direction of all segmented image text boxes is adjusted to be vertical upward;

[0023] Step 3, all text images in the segmented image text box are obtained to obtain a text image set;

[0024] Step 4, each text in the text image set is split into a stroke set containing a plurality of strokes;

[0025] Step 5, each type of stroke in the stroke set is aggregated to obtain a representative stroke of each type of stroke;

[0026] Step 6, the representative stroke of each type of stroke is used to select a text recognition database corresponding to a font with the highest matching degree as a target text recognition database in a preset multi-font text recognition database;

[0027] Step 7, fine-tuning the pre-trained character recognition model using the target character recognition database to obtain a target character recognition model;

[0028] Step 8, using the target character recognition model to recognize the character image set to obtain all the characters in the segmentation image text box as the segmentation image text.

[0029] Step 1, the step of direction labeling the segmentation image text box, specifically:

[0030] Step 2, the step of adjusting the direction of all segmentation image text boxes to be vertical upward, specifically:

[0031] Step 3, the step of obtaining all the character images in the segmentation image text box to obtain a character image set, specifically:

[0032] The multi-font character recognition database specifically includes:

[0033] Further, the step of splitting each character in the character image set into a stroke set containing multiple strokes, specifically includes:

[0034] Using a thinning algorithm to thin each character image to obtain a single-pixel width stroke;

[0035] By tracking the single-pixel width stroke, the stroke set corresponding to the character image is obtained.

[0036] Further, the step of aggregating each type of stroke in the stroke set to obtain a representative stroke of each type of stroke, specifically includes:

[0037] Define a stroke feature descriptor to measure the shape, direction, and other information of the stroke;

[0038] K-Means clustering is performed on all stroke sets to obtain the clustering center as the representative stroke.

[0039] Further, the step of using the representative stroke of each type of stroke to select the character recognition database corresponding to the font with the highest matching degree with the representative stroke of each type of stroke in the preset multi-font character recognition database as the target character recognition database, specifically includes:

[0040] Construct a multi-font character recognition database, each database containing a specific font character set and corresponding stroke splitting information;

[0041] Input a text image to extract key strokes and obtain a representative stroke of each type of stroke;

[0042] Calculate the matching degree of the representative stroke and the representative stroke of each preset database.

[0043] The matching degrees of each database are summarized and counted, and the database with the highest matching degree is selected as the target database for subsequent character recognition.

[0044] Specifically:

[0045] A multi-font character recognition database is constructed, each database containing a character set of a specific font and corresponding stroke splitting information, for each font, representative characters are selected, stroke splitting is performed, and a representative stroke set of the font is obtained;

[0046] A text image is input, the key strokes are extracted according to the foregoing steps, and the representative strokes of each type of stroke are obtained;

[0047] The matching degree of each preset database is calculated with the representative strokes and the representative strokes of the font. The matching degree calculation method can use Hausdorff distance, etc.

[0048] The matching degrees of each database are summarized and counted, and the database with the highest matching degree is selected as the target database for subsequent character recognition.

[0049] Further, the step of fine-tuning the pre-trained character recognition model using the target character recognition database to obtain a target character recognition model, specifically includes:

[0050] Load the pre-trained character recognition model;

[0051] Construct a training data set for the target font;

[0052] Fine-tune the pre-trained character recognition model using the training data set;

[0053] Save the fine-tuned model for subsequent insurance policy character recognition.

[0054] The second aspect of the application provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions, and the program instructions are used to execute the above-mentioned insurance policy structured storage method based on intelligent identification.

[0055] The third aspect of the application provides an insurance policy structured storage system based on intelligent identification, which comprises the above-mentioned computer readable storage medium.

[0056] The application discloses an insurance policy structured storage method based on intelligent identification. The method realizes accurate extraction and structured representation of insurance policy image information through intelligent image processing and text recognition technology, converts complex insurance policy text into structured objects, and stores the structured objects in a database in the form of key-value pairs, so that the insurance policy information is easy to search, count and reuse, and has significant technical effects and progress.

[0057] Compared with the current commonly used manual input of policy information method, the present application has the following remarkable effects:

[0058] 1. Improve the accuracy of policy information extraction

[0059] The current manual entry of policy information has the phenomenon of missing and misrecording, which directly affects the effect of subsequent use of policy data. The present application can automatically identify all text information in the policy image through image segmentation, character recognition, key information extraction and other technical means, and the accuracy can reach more than 90%. Especially for different types of policies, different structured templates are designed, and the key fields are extracted combined with rule method, which can greatly improve the accuracy of information extraction.

[0060] 2. Realize the structured expression of policy content

[0061] The present application converts natural language policy text into structured object data, and stores each policy information in the form of standardized key-value pair. This greatly facilitates the retrieval and statistical analysis of policy data. Users can quickly find the required data through SQL statements, and can also flexibly perform multi-dimensional analysis, which is difficult to achieve from unstructured text.

[0062] 3. Support text direction correction

[0063] The text direction of the policy image may be inclined or horizontally arranged, and direct recognition of the text content will bring great difficulty. The present application can realize the unified arrangement of the text box through text direction correction, effectively improving the adaptability of the subsequent recognition algorithm.

[0064] 4. Recognition of different personnel handwriting fonts

[0065] Because everyone's handwriting is different, the stroke features of different fonts, direct application of a unified recognition model often cannot achieve ideal results. The present application uses the scheme of matching the strokes in the text image with the standard font library to realize accurate positioning of the recognition model corresponding to the text font, thereby significantly improving the recognition accuracy.

[0066] In summary, the present application has the innovation points of image recognition, information extraction, structured conversion, etc., which can greatly improve the intelligent degree of policy processing, reduce the manual workload, and make the policy data more convenient for statistics, analysis, sharing and management. This has important significance for promoting the digital transformation of insurance business. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor under the premise of the drawings.

[0068] Figure 1 A flowchart of a policy structured storage method based on intelligent identification provided by the present application is shown.

[0069] Figure 2 A flowchart of the step of performing text recognition on the first image to obtain a text set is shown.

[0070] Figure 3 A DeeplabV3+ model diagram is shown. DETAILED DESCRIPTION

[0071] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application.

[0072] In addition, the terms "first" and "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0073] As shown in Figure 1 A flowchart of a policy structured storage method based on intelligent identification provided by the first aspect of the present application is shown. The method includes the following steps:

[0074] S10, obtaining a policy image;

[0075] S20, pre-processing the policy image to obtain a first image;

[0076] S30, performing text recognition on the first image to obtain a text set;

[0077] S40, performing structured processing on the obtained text set according to a pre-set contract form database to obtain a policy object;

[0078] S50, storing the policy object in the database in the form of key-value pairs.

[0079] In step S10, the method of obtaining a policy image includes mobile phone shooting, user selection of picture uploading, etc. In step S10, the method of obtaining a policy image includes mobile phone shooting, user selection of picture uploading, etc.

[0080] Step S20, pre-process the insurance policy image to obtain a first image:

[0081] Before pre-processing the insurance policy image, first, the insurance policy image needs to be cropped to remove irrelevant content and only keep the main content of the insurance policy. Then the following pre-processing is performed:

[0082] 1) Image graying: convert the RGB image to a grayscale image to simplify subsequent processing and reduce computational complexity.

[0083] 2) Median filtering: used to eliminate image noise points and make the image smooth.

[0084] 3) Image binarization: convert the grayscale image to a black and white binary image to facilitate subsequent recognition. Common binarization methods include Otsu method, adaptive threshold method, etc. Here, the Otsu method is used.

[0085] Since the image graying, median filtering, and binarization processing are very common existing technologies, they can also be processed using conventional methods instead of the above formulas. The specific implementation can refer to the article "Image Preprocessing" in Baidu Encyclopedia or the book "Image Processing, Analysis and Machine Vision (3rd Edition)", published by Tsinghua University Press, translated by Ai Haizhou, Su Yanchao, etc.

[0086] In step S40, since the insurance policy is a standard contract, each standard contract corresponds to a form, and the key of this form can be pre-stored in the form database. According to the obtained text set, the text matching method is used to obtain the value corresponding to each key of the form, forming a key-value pair as the insurance policy object.

[0087] In extremely special cases (this case is extremely rare, because insurance companies have been using normal insurance policies for many years, even if new insurance policies are used, they are all set by insurance companies in advance, and the keys of the forms corresponding to the insurance policies can be pre-set when the insurance policies are set), template matching and rule extraction methods can also be considered to obtain the insurance policy object, and the specific steps are as follows:

[0088] Step 41, construct a contract form knowledge base

[0089] First, collect different types of insurance policy templates, label the templates, obtain form structure information, and construct a contract form knowledge base. The knowledge base stores the mapping relationship between insurance policy types and corresponding template structures. The template structure is represented by hierarchical labeling, for example:

[0090] Insurance policy type A

[0091] |--Insurance policy number

[0092] |--Insured person information

[0093] | Name

[0094] | ID

[0095] | Insurance Company

[0096] | Effective Date

[0097] Step 42, Extracting Structural Information Based on Template Matching

[0098] For the input policy text, first determine the policy type and find the corresponding template structure in the knowledge base. Then, use string matching algorithms to extract text segments item by item based on the template structure. An example algorithm is as follows:

[0099] Let the text collection be T = t1, t2, …, t n , and the template structure be M = m1, m2, …, m k . Perform the following matching:

[0100] For each template item m i :

[0101] - Construct a matching string set S i , containing all possible matching texts;

[0102] - For each text segment t j , calculate the matching degree with S i : match(t j , S i );

[0103] - Select the text segment with the highest matching degree from T with S i as the extraction result corresponding to m n ;

[0104] The matching degree calculation uses Jaccard similarity:

[0105]

[0106] Repeat the above process until all items in the template M are matched, and finally obtain the structured result.

[0107] Step 43, Extracting Structural Information Based on Rules

[0108] Use regular expressions or syntax rules to perform structured analysis on the text. An example rule is as follows:

[0109] Policy number rule: starts with the letter "BH" followed by 10 digits;

[0110] If there is a string like "BH123456789" in the text, it can be extracted as the policy number;

[0111] At the same time, a series of regular rules of address, date, amount, etc. are defined, and structured information is extracted from the text according to the rules;

[0112] Comprehensive based on template matching and rule extraction method, can accurately identify the structure of the policy, get structured policy object.

[0113] The method of step S50 is a commonly used method, and the preferred database is Redis, Amazon DynamoDB, Microsoft Azure Cosmos DB, Memcached, etc.

[0114] In the above technical solution, the step of pre-processing the policy image specifically includes: gray processing, median filtering and binaryzation processing.

[0115] As shown in the above technical solution, the step of performing text recognition on the first image to obtain a text set specifically includes: Figure 2

[0116] S31, segmenting the first image to obtain a plurality of segmented images to form a segmented image set;

[0117] S32, performing text detection on each segmented image in the segmented image set to obtain a text box corresponding to each segmented image, denoted as a segmented image text box;

[0118] S33, performing text direction detection on each segmented image text box to obtain the direction of the segmented image text box;

[0119] S34, performing text recognition on the segmented image text box according to the detection direction to obtain a segmented image text;

[0120] S35, combining the obtained segmented image text into a text set according to the order of the segmented images.

[0121] Step S31, segmenting the first image to obtain a plurality of segmented images to form a segmented image set:

[0122] 1) Tilt correction

[0123] Due to the shooting angle and other reasons, the input image may have a certain tilt. In order to improve the segmentation effect, tilt correction is needed first.

[0124] Use Hough transform to detect the main straight line direction of the image, and calculate the tilt angle θ.

[0125] 2) Segmentation line detection ​

[0126] Detect the segmentation line in the image using a probabilistic Chinese character Stein model (PSM).

[0127] 3) Image cutting

[0128] Cut the image horizontally into a plurality of sub-images based on the detected segmentation line. Then, cut vertically based on connected domain analysis. Finally, obtain a set of segmented images {I1, I2, …, In}. n

[0129] Step S32, perform text detection on each segmented image in the set of segmented images to obtain a text box corresponding to each segmented image, denoted as segmented image text box:

[0130] 1) Preprocessing

[0131] - Grayscale, binarization

[0132] - Eliminate interference lines

[0133] - Vertical and horizontal projection, determine the text box boundary

[0134] 2) Text box detection

[0135] Scan the image, and form a connected component by connecting the pixel points. Output the bounding rectangle of all connected components, which is the detected text box.

[0136] 3) MSN text detection

[0137] The text detection method borrows the idea of the MSER algorithm, and takes the extreme stable region in the image as the text region.

[0138] wherein, assume that the pixel point set in the text box is R = {r1, r2, …, r n}, and the extreme stable region determination criterion is:

[0139]

[0140] wherein, Δ(r i ) represents the range of R change caused by changing the threshold value with r i as the reference point, and T y is the threshold value, generally taking a value of 85%. The R satisfying the above formula is the MSER, and the bounding rectangle thereof is the detected text box.

[0141] Integrate the above methods to detect the text box of each segmented image.

[0142] Step S33, perform text box direction detection on each segmented image text box to obtain the direction of the segmented image text box:

[0143] 1) Edge detection ​

[0144] Edge detection is performed on the text box using a Sobel operator to obtain an edge image.

[0145] 2) Calculate the edge direction histogram

[0146] 3) Direction determination

[0147] The direction with the most number of histograms is selected as the text direction σ.

[0148] Step S34, according to the segmented image text box in the detection direction, text recognition is performed to obtain the segmented image text:

[0149] 1) Angle correction

[0150] The previously obtained text direction σ is used to perform image rotation correction to obtain an angle-corrected image.

[0151] 2) Text recognition

[0152] An identification model based on CNN-LSTM-CTC is used, and the specific steps are:

[0153] (1) Use CNN to extract text image features to obtain a feature sequence F = (f1, f2, …, f n );

[0154] (2) Use LSTM to predict the feature sequence to obtain a probability matrix P = (p1, p2, …, p m );

[0155] (3) Perform CTC Loss optimization to obtain the most likely output text label = (l1, l2, …, l k );

[0156] Wherein, the CTC Loss function is:

[0157] L CTC = -log(p(label|F));

[0158] Through model training, the recognition of the segmented image text can be completed to obtain the text result.

[0159] Step S35, the segmented image text obtained is combined into a text set according to the order of the segmented image.

[0160] The segmented image text is directly spliced in order to obtain a complete insurance policy text set, and the text recognition process is completed.

[0161] Optionally, in the above image segmentation, each user has their own shooting habits, and the insurance policy pictures received by the server are irregular. Through the annotation of the training insurance policy samples, an insurance policy instance segmentation model based on DeeplabV3+ is trained. The image segmentation step further includes:

[0162] Step 31.1, collect the insurance policy strongly related pictures in Baidu pictures by searching the insurance policy keywords; superimpose the existing insurance policy data in the system to synthesize the training data set, and manually annotate the insurance policy instance segmentation data set.

[0163] Step 31.2, construct an instance segmentation model based on DeeplabV3+, put the data set obtained in step 31.1 into the DeeplabV3+ model according to the training set and the validation set in the ratio of 9:1, and train to obtain an instance segmentation model of the insurance policy.

[0164] Step 31.3, the inference result obtained by the model in step 31.2 is processed to obtain the minimum bounding rectangle of the prediction result by the minAreaRect method of Opencv, and affine transformation is performed on it to obtain the corrected insurance policy image.

[0165] Wherein, Figure 3 is a DeeplabV3+ model diagram, the picture is processed by a deep learning network of encoding (Encoder) and decoding (Decoder) to complete the instance segmentation of the insurance policy, wherein DCNN represents the backbone network for extracting picture features, FP represents the feature map, Concat represents feature fusion, Upsample represents up sampling, and Prediction represents the prediction map

[0166] Optionally, for step S32, the current text detection algorithm can be roughly divided into regression-based method and segmentation method. The present patent is based on the segmentation method, which constructs a deep neural network model to output the text segmentation result (probability map, each pixel is the probability of being a positive sample) and threshold result (threshold map, used to determine the foreground and background) of the picture, and combines the probability map and the threshold map to convert into a detection result. Text detection includes the following steps:

[0167] Step S32.1, annotate the ocr training data set. The training data in 31.1 is annotated with text boxes to obtain a text detection training set and a text recognition training set.

[0168] Step S32.2, generating a probability map. A 1 / 4 feature map F is obtained through a Feature Pyramid Networks (FPN) network structure, a Probability map (P) and a Threshold map (T) are obtained through F, and a Binary map (B) is obtained through P and T. P, T and B are supervised and trained during training, and P and B use the same supervision signal (label).

[0169] Step S32.3, generating a text box. P is binarized using a fixed threshold of 0.2 to obtain B; a shrunk text region is obtained from B; and the shrunk text region is expanded by an offset coefficient D' of the Vatti clipping algorithm to obtain a final text box. The Vatti clipping algorithm is the underlying implementation principle of many geometry libraries, such as clipper2, which is implemented based on the Vatti clipping algorithm.

[0170] Optionally, the following is another embodiment of step S34.

[0171] Further, in the above technical solution, the step of detecting the segmented image text box according to the direction to obtain the segmented image text, specifically includes:

[0172] Step 1, direction labeling is performed on the segmented image text box;

[0173] Step 2, the direction of all segmented image text boxes is adjusted to be vertical upward;

[0174] Step 3, all text images in the segmented image text box are obtained to obtain a text image set;

[0175] Step 4, each text in the text image set is split into a stroke set containing multiple strokes;

[0176] Step 5, each type of stroke in the stroke set is aggregated to obtain a representative stroke of each type of stroke;

[0177] Step 6, the representative stroke of each type of stroke is used to select a text recognition database corresponding to the font with the highest matching degree as a target text recognition database in a preset multi-font text recognition database;

[0178] Step 7, the target text recognition database is used to fine-tune a pre-trained text recognition model to obtain a target text recognition model;

[0179] Step 8, recognizing the character image set by using the target character recognition model to obtain all the characters in the segmented image text box as the segmented image text.

[0180] In the above steps, steps 1-2 are conventional operations and are not explained in detail.

[0181] Step 3, obtaining all the character images in the segmented image text box, the specific steps of obtaining the character image set include:

[0182] 1) Perform image binarization to obtain a black and white image.

[0183] 2) Extract all character images by searching for character connections. The search methods include:

[0184] - DFS (Depth First Search): Extract character connection areas based on depth-first search;

[0185] - Contour Tracing: Extract character areas based on contour tracing;

[0186] 3) Trim the character image to remove the surrounding white space boundary;

[0187] 4) Combine the extracted character images to obtain the character image set {I1, I2, …, I n};

[0188] Step 4, the specific steps of splitting each character in the character image set into a stroke set containing multiple strokes include:

[0189] 1) Thinning each character image using a thinning algorithm to obtain single-pixel width strokes;

[0190] 2) Obtain the stroke set corresponding to the character image by tracing the single-pixel width strokes.

[0191] Further, in the above technical solution, the step of splitting each character in the character image set into a stroke set containing multiple strokes includes:

[0192] Thinning each character image using a thinning algorithm to obtain single-pixel width strokes;

[0193] Obtain the stroke set corresponding to the character image by tracing the single-pixel width strokes.

[0194] The thinning algorithm here can use layer-by-layer stripping thinning or distance transformation-based thinning method.

[0195] Further, in the above technical solution, the step of aggregating each type of stroke in the stroke set to obtain a representative stroke of each type of stroke includes:

[0196] Define stroke feature descriptors to measure the shape, orientation, etc. of strokes.

[0197] K-Means clustering on all stroke set, the clustering center as the representative stroke.

[0198] That is, the specific implementation of step 5 is:

[0199] 1) Define stroke feature descriptors to measure the shape, orientation, etc. of strokes.

[0200] For example, distance-based feature descriptors: select n key points in the stroke, calculate the Euclidean distance between adjacent key points, and form a feature vector.

[0201] 2) K-Means clustering on all stroke set: the goal is to divide multiple strokes into multiple classes to minimize intra-cluster variance.

[0202] Further, in the above technical solution, the step of using the representative stroke of each type of stroke to select the font corresponding to the highest matching degree of the representative stroke of each type of stroke from the pre-set multi-font character recognition database as the target character recognition database, specifically includes:

[0203] Construct a multi-font character recognition database, each database containing a specific font character set and corresponding stroke splitting information;

[0204] Input a text image to extract key strokes and get a representative stroke for each type of stroke;

[0205] Calculate the matching degree of the representative stroke and the representative stroke of each pre-set database;

[0206] Summarize and count the matching degrees of each database, and select the highest matching database as the target database for subsequent character recognition.

[0207] Further, in the above technical solution, the step of using the target character recognition database to fine-tune the pre-trained character recognition model to obtain the target character recognition model, specifically includes:

[0208] Load the pre-trained character recognition model;

[0209] Construct a training data set for the target font;

[0210] Fine-tune the pre-trained character recognition model using the training data set;

[0211] Save the fine-tuned model for subsequent policy character recognition.

[0212] The specific implementation of step 7 is as follows:

[0213] 1) Load a pre-trained character recognition model, which can adopt a CNN+RNN+CTC architecture.

[0214] 2) Construct a training data set for the target font:

[0215] - For each target font character, extract its stroke information, including stroke order, stroke classification, etc.

[0216] - Map the target font strokes to the stroke structure and order in the insurance policy image. Here, the stroke extraction method in step 4 can be referred to.

[0217] - Generate character images for training.

[0218] 3) Fine-tune the model:

[0219] - Set a small learning rate and fix the CNN feature extraction layer parameters.

[0220] - Input the mapped stroke image and output the target font label.

[0221] - Through iterative training, fine-tune the RNN and CTC layers to obtain a character recognition model for the target font.

[0222] 4) Save the fine-tuned model for subsequent insurance policy character recognition.

[0223] The second aspect of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions, and the program instructions are used to execute the above-mentioned insurance policy structured storage method based on intelligent recognition.

[0224] The third aspect of the present application provides an insurance policy structured storage system based on intelligent recognition, which comprises the above-mentioned computer readable storage medium.

[0225] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A policy structuring storage method based on intelligent identification, characterized in that, The method comprises the following steps: S10, obtaining an insurance policy image; S20, preprocessing the insurance policy image to obtain a first image; S30, performing text recognition on the first image to obtain a text set; S40, performing structured processing on the obtained text set according to a pre-set contract form database to obtain an insurance policy object; S50, storing the insurance policy object in a database in a key-value pair mode; The step of performing text recognition on the first image to obtain a text set specifically comprises: S31, segmenting the first image to obtain a plurality of segmented images to form a segmented image set; S32, performing text detection on each segmented image in the segmented image set to obtain a text box corresponding to each segmented image, denoted as a segmented image text box; S33, performing text direction detection on each segmented image text box to obtain the direction of the segmented image text box; S34, performing text recognition on the segmented image text box according to the detection direction to obtain segmented image text; S35, combining the obtained segmented image text into a text set according to the order of the segmented images; The step of performing text recognition on the segmented image text box according to the detection direction to obtain segmented image text specifically comprises: Step 1, direction labeling is performed on the segmented image text box; Step 2, the direction of all segmented image text boxes is adjusted to be vertical upward; Step 3, all text images in the segmented image text box are obtained to obtain a text image set; Step 4, each text in the text image set is split into a stroke set containing a plurality of strokes; Step 5, each type of stroke in the stroke set is aggregated to obtain a representative stroke of each type of stroke; Step 6, a representative stroke of each type of stroke is used to select, in a pre-set multi-font text recognition database, a text recognition database corresponding to a font with the highest matching degree as a target text recognition database; Step 7, the target text recognition database is used to fine-tune a pre-trained text recognition model to obtain a target text recognition model; Step 8, the target text recognition model is used to recognize the text image set to obtain all text in the segmented image text box as segmented image text; The step of splitting each text in the text image set into a stroke set containing a plurality of strokes specifically comprises: Each text image is thinned using a thinning algorithm to obtain a single-pixel-width stroke; The stroke set corresponding to the text image is obtained by tracking the single-pixel-width stroke; The step of aggregating each type of stroke in the stroke set to obtain a representative stroke of each type of stroke specifically comprises: A stroke feature descriptor is defined to measure the shape and direction of the stroke; K-Means clustering is performed on all stroke sets to obtain a clustering center as a representative stroke; The step of selecting, in a pre-set multi-font text recognition database, a text recognition database corresponding to a font with the highest matching degree as a target text recognition database using a representative stroke of each type of stroke specifically comprises: A multi-font character recognition database is constructed, each database containing a character set of a specific font and corresponding stroke splitting information; A text image is input to extract key strokes and obtain representative strokes of each type of stroke; The matching degree of each preset database with the representative strokes of the font is calculated; The matching degrees of the databases are summarized and the database with the highest matching degree is selected as the target database for subsequent character recognition.

2. The method of claim 1, wherein the method is characterized by, The step of preprocessing the insurance policy image specifically includes grayscale processing, median filtering, and binaryzation processing.

3. The method of claim 2, wherein the method further comprises: The step of fine-tuning the pre-trained character recognition model using the target character recognition database to obtain a target character recognition model specifically includes: loading a pre-trained character recognition model; constructing a training data set for the target font; fine-tuning the pre-trained character recognition model using the training data set; saving the fine-tuned model for subsequent insurance policy character recognition.

4. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, and when the program instructions are run, a method for structuring and storing an insurance policy based on intelligent recognition according to any one of claims 1-3 is executed.

5. An intelligent recognition based policy structuring and storage system characterized in that, The computer readable storage medium of claim 4 is included.

Images