A method and apparatus for automatically auditing proof materials

Abstract

The application provides a method and device for automatically auditing certification materials, which can automatically analyze and judge whether the certification materials uploaded by a user are standard, greatly improves the overall auditing efficiency, and has high accuracy. The application utilizes a template configuration function to configure templates of different certification materials and a similarity threshold value; according to a similarity algorithm, the similarity between the certification materials uploaded by the user and the templates is calculated according to different templates uploaded by the user, and the picture data that does not meet the requirements can be pre-filtered. The application trains the data of the certification materials of different modules by collecting different template training data by using a tool, then trains and improves an adaboost algorithm, obtains a picture auditing algorithm, and uses the data set obtained by the manual review module as training samples again, recursively trains the auditing algorithm, and improves the algorithm accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data processing, and relates to computer software, databases and data modeling, in particular to a method and device for automatically auditing proof materials. BACKGROUND

[0002] With the large-scale application of software technology, the uploading, inputting and storage of large-scale data currently cannot be supported without software systems. For any system that requires users to upload proof materials, it is often necessary to manually audit the proof materials uploaded by the users through the background to determine whether they meet the requirements of the system. However, with the gradual increase in the business volume of the system, a large amount of manpower is required to audit the proof materials, and the work intensity may be too great to cause auditing errors, which will result in the gradual decrease in the overall efficiency of the system and even the delay in the operation of the social key link.

[0003] The proof materials uploaded by the users often contain stamped seal information such as housing proof, educational background certificate, degree certificate, household register, driver's license, social security payment information, labor contract, birth certificate, medical report and the like. Most systems require the users to upload the proof materials after taking photos of them. However, there is currently a lack of accurate and effective automatic auditing scheme for the materials. Based on the unique properties of the proof materials, the present application proposes a method and device for automatically auditing the proof materials. SUMMARY

[0004] To solve the above problems, the present application uses database and data modeling technology to construct a method and device for automatically auditing proof materials.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows:

[0006] A method for automatically auditing proof materials, comprising the following steps:

[0007] Step 1: configuring a proof material template required to be audited by the uploading system, specifically comprising:

[0008] Step 1.1, collecting the types of proof materials required to be audited by the current business system;

[0009] Step 1.2, selecting proof materials that meet the conditions from the proof materials collected in step 1.1, and making a proof material template therefrom;

[0010] Step 1.3, marking whether the proof material template made in step 1.2 needs to be stamped with a seal and the type of the proof material, and setting different code values for different types;

[0011] Step 2: storing the code values corresponding to the types of the proof materials uploaded by the users;

[0012] Step 3: Calculate the similarity of the user uploaded proof materials and the template materials configured in step 1; including:

[0013] Step 3.1, according to the user uploaded proof material code value, filter out the template for similarity calculation through the proof material type uploaded in step 2;

[0014] Step 3.2, calculate the similarity of the current user proof material and the template through the similarity matching algorithm, including the following process: first, crop the pictures in the user uploaded materials to make them consistent with the picture size in the template; And the cropped pictures are processed by gray scale; Calculate the histogram data of the picture, count the probability distribution of the same pixel points, and calculate the similarity of the two picture histograms according to the correlation calculation formula;

[0015] Step 4: Filter the proof materials with low similarity in step 3, judge whether the similarity calculated in step 3 exceeds the set threshold, if not, directly determine that the current proof material does not meet the requirements, and return the result information directly; If it is equal to or exceeds the set threshold, it enters step 5;

[0016] Step 5: Automatically review the proof materials filtered in step 4; Identify the proof materials exceeding the threshold in step 4 and extract the seal information in the materials through the pre-trained improved adaboost algorithm, including:

[0017] Step 5.1, use the internet and the system itself to collect proof materials, and manually annotate the training data required by the improved adaboost algorithm;

[0018] Step 5.2, use the manually annotated training data set to perform improved adaboost training, and obtain the trained classifier as the picture review algorithm;

[0019] Step 5.3, for the proof materials of step 4, use the picture review algorithm to give specific identification results, that is, whether the current proof material contains seal information, if it contains, it directly enters step 6 for review, if it does not contain, it directly returns the result information;

[0020] Step 6: Quickly check whether the proof material picture containing the seal reviewed by the review algorithm in step 5 has errors, if the manual review passes, return the result information to the specific business system through the picture review result module; When the review does not pass, notify the specific business system of the result;

[0021] Step 7: Return the result to the business system through the picture review result module in the review layer, thereby completing the review of the proof materials and improving the review efficiency.

[0022] Further, the user uploaded proof materials in step 2 are stored through the http api method.

[0023] Furthermore, the specific process of improving the AdaBoost algorithm in step 5 is as follows:

[0024] 1. Given a training sample set S = {(x1, y1), ..., (x... m ,y m )}, weak separator space H, where x i Let ∈X be a sample vector, y = ±1 be the class label, and m be the total number of samples; initialize the sample probability distribution, D1(i) = 1 / m, i = 1,...,m;

[0025] 2. Iterate T times:

[0026] (1) h in the weak classifier space H:

[0027] a. Divide the sample space X into n equal parts, i.e., X1, X2, ..., Xn. n

[0028] b. The probability distribution of the training samples is D t ,calculate

[0029]

[0030] c. Output of the weak classifier in the above partitioning

[0031]

[0032] in It is a dynamic positive constant;

[0033] d. Calculate the normalization factor

[0034]

[0035] (2) Select an h from the weak classifier space. t Minimize Z

[0036]

[0037] (3) Update the weights of each training sample to satisfy the normalization of the probability distribution.

[0038]

[0039] Z t As the normalization factor, it makes

[0040] 3. The final strong classifier H is

[0041]

[0042] wherein b is a given threshold value, and the default value is 0.

[0043] Further, the picture data that fails the audit in step 6 is fed back to the training data set as training samples for the next model training.

[0044] The application also provides a device for automatically auditing proof materials, which is used to implement the method for automatically auditing proof materials, and comprises a configuration layer, an access layer, a filtering layer, an audit layer, and a review layer.

[0045] The configuration layer comprises a template uploading module and a template configuration module, the template uploading module is used to upload templates, the template configuration module is used to configure different encoding values for distinguishing template types according to different templates, and a similarity threshold value is configured.

[0046] The access layer comprises a picture access interface module, which is used to store the proof materials uploaded by users in a manner of http api and inform the current proof material type.

[0047] The filtering layer comprises a picture similarity calculation template and a picture filtering module, the picture similarity calculation template is used to calculate the similarity between the proof materials uploaded by users and the template materials configured by the configuration layer, and the filtering module is used to filter the proof materials with low similarity and send the proof materials with high similarity to the audit layer.

[0048] The audit layer comprises a picture audit algorithm module, which is used to recognize the proof materials exceeding the threshold value filtered by the filtering module and extract the seal information in the materials by using the improved adaboost algorithm trained in advance.

[0049] The review layer comprises a picture artificial review module and a picture audit result module, the artificial review module is used to review whether the proof materials containing seals audited by the audit layer are wrong, and the picture audit result module is used to output the results, and the picture data that fails the audit is also fed back to the audit layer as training samples for the next model training.

[0050] The application has the following beneficial effects:

[0051] 1. The application can automatically analyze and judge whether the proof materials uploaded by users are standard, greatly improves the overall audit efficiency, and has high accuracy.

[0052] 2. The application configures the templates of different proof materials and the similarity threshold value by using the template configuration function, calculates the similarity between the proof materials uploaded by users and the templates according to different templates uploaded by users according to the similarity algorithm, and can pre-filter the picture data that does not meet the requirements.

[0053] 3. The present application trains the data of the proof materials of different modules by collecting different template training data and using the tool to mark, then trains the improved adaboost algorithm, obtains the picture auditing algorithm, and uses the data set obtained by the artificial review module as the training sample again, recursively trains the auditing algorithm, and improves the algorithm accuracy.

[0054] 4. The present application overcomes the defect that the smoothing factor in the traditional continuous adaboost algorithm is a fixed value, so that the smoothing effect on all features is the same, which leads to the over-learning phenomenon of the continuous adaboost algorithm during training, improves the adaboost algorithm, changes the fixed value of the smoothing factor to a dynamic platform factor, and the smoothing factor of different training samples corresponds to different values. The improved adaboost algorithm greatly suppresses the over-learning phenomenon and greatly improves the recognition rate. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The automatic proof material auditing method provided by the present application is shown in the flowchart.

[0056] Figure 2 The similarity matching algorithm is shown in the schematic diagram.

[0057] Figure 3 The review process diagram of the review layer is shown.

[0058] Figure 4 The artificial marking data in the improved adaboost algorithm is shown in the schematic diagram.

[0059] Figure 5 The picture auditing algorithm flowchart is shown.

[0060] Figure 6 The picture review flowchart is shown.

[0061] Figure 7 The automatic proof auditing device architecture provided by the present application is shown in the flowchart. DETAILED DESCRIPTION

[0062] The technical solutions provided by the present application will be described in detail below in combination with specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application. In addition, the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0063] The present application provides an automatic proof material auditing method, the main process of which is shown in Figure 1 as follows:

[0064] Step 1: configure the template of the proof materials that need to be audited by the uploading system through the configuration layer, including the following:

[0065] Step 1.1, collect which proof materials need to be audited by the current business system;

[0066] Step 1.2, for the paper proof materials collected in step 1.1, make them into picture format and upload them to the system,

[0067] Step 1.3, mark whether the proof material template made in step 1.2 needs to be stamped and the type of the proof material, and set different coding values for different types, which are directly applied to the similarity comparison in the following steps;

[0068] Step 2: store the proof materials uploaded by the user through the access interface module of the access layer, mainly through the httpapi method to store the proof materials uploaded by the user, and the code value corresponding to the current proof material needs to be uploaded to distinguish the type of the proof material,

[0069] Step 3: calculate the similarity of the proof materials uploaded by the user and the template materials configured in step 1 through the picture similarity in the filtering layer; the specific details are as follows:

[0070] Step 3.1, according to the type of the proof materials uploaded in step 2, filter out the templates for similarity calculation according to the code value of the proof materials uploaded by the user;

[0071] Step 3.2, calculate the similarity of the current user's proof materials and the template through the similarity matching algorithm, and the specific calculation is referred to Figure 2 , including the following process: first, crop the pictures in the user's uploaded materials to make them consistent with the picture size in the template; and then perform grayscale processing on the cropped pictures, and the pixels of the processed pictures are between [0-225]; calculate the histogram data of the pictures, and count the probability distribution of the same pixel points; according to the correlation calculation formula, calculate the similarity of the two picture histograms.

[0072] Step 4: filter the proof materials with low similarity in step 3 using the picture filtering module in the filtering layer, and determine whether the similarity calculated in step 3 exceeds the set threshold value; if not, it is directly determined that the current proof material does not meet the requirements, and directly enters step 6, and the result information is returned through the audit result module of the review layer, and the overall process is as follows Figure 3 If it is equal to or exceeds the set threshold value, it enters step 5;

[0073] Step 5: automatically audit the proof materials filtered in step 4 using the picture audit algorithm module; this step identifies the proof materials exceeding the threshold value in step 4 and extracts the seal information in the materials by using the improved adaboost algorithm trained in advance, and the specific details are as follows:

[0074] Step 5.1: Using the internet and supporting materials collected by the system itself, manually annotate the training data needed to improve the AdaBoost algorithm. Specific annotated data is as follows: Figure 4 As shown, this is where the material is stamped (the original image was in color, but due to patent submission requirements, a gray image is now provided).

[0075] Step 5.2: Using the manually labeled training dataset, perform improved AdaBoost training and obtain the trained classifier as the image review algorithm. The specific process of improving the AdaBoost algorithm is as follows:

[0076] 1. Given a training sample set S = {(x1, y1), ..., (x... m ,y m )}, weak separator space H, where x i Let ∈X be a sample vector, y = ±1 be the class label, and m be the total number of samples. Initialize the sample probability distribution, D1(i) = 1 / m, i = 1, ..., m.

[0077] 2. Iterate T times

[0078] (1) h in the weak classifier space H:

[0079] a. Divide the sample space X into n equal parts, i.e., X1, X2, ..., Xn. n

[0080] b. The probability distribution of the training samples is D t ,calculate

[0081]

[0082] p represents the probability value;

[0083] c. Output of the weak classifier in the above partitioning.

[0084]

[0085] in It is a dynamic positive number.

[0086] d. Calculate the normalization factor

[0087]

[0088] (2) Select an h from the weak classifier space. t Minimize Z

[0089]

[0090] (3) update the weight of each training sample to meet the normalization of probability distribution

[0091]

[0092] where Z t is a normalization factor, so that

[0093] 3, the final strong classifier H is

[0094]

[0095] Where b is a given threshold, the default value is 0.

[0096] Step 5.3, when the proof material of step 4 enters the picture audit algorithm module of the audit layer, the picture audit algorithm will give specific identification results, mainly whether the current proof material contains seal information, contains and directly enters step 6 review layer, does not contain and directly returns result information through the audit result module of the review layer, the overall process is as shown in Figure 5

[0097] Step 6: using the artificial review module in the review layer, quickly verifying whether the seal-containing proof material picture audited by step 5 audit algorithm has errors, if the artificial review passes, the result information is returned to the specific business system through the picture audit result module; When there is an error, the artificial review does not pass, in addition to notifying the specific business system of the result, the picture data of the audit failure is fed back to the audit layer as the training sample for next model training, so as to improve the overall audit accuracy, the overall process is as shown in Figure 6 Picture audit result module is used to output audit result, when the review is the material filtered by the filtering layer which does not exceed the threshold, then directly return error result.

[0098] Step7: through the picture audit result module in the review layer, the result is returned to the business system, so as to complete the audit of the proof material and improve the audit efficiency.

[0099] The application provides an automatic proof material auditing device for realizing the automatic proof material auditing method, and the architecture is as shown in Figure 7 , which comprises a configuration layer, an access layer, a filtering layer, an audit layer and a review layer.

[0100] The configuration layer comprises a template uploading module and a template configuration module, the template uploading is used for uploading the manually made picture template, and the template configuration mainly refers to configuring different coding values for different templates to distinguish the template types, and configuring a similarity threshold.

[0101] ​The access layer comprises a picture access interface module, which is configured to store the uploaded proof materials of the user through an http API and to inform the current proof material type.

[0102] The filtering layer comprises a picture similarity calculation template and a picture filtering module, wherein the picture similarity calculation template is configured to calculate the similarity between the uploaded proof materials of the user and the template materials configured by the configuration layer, and the filtering module is configured to filter the proof materials with low similarity and to send the proof materials with high similarity to the auditing layer.

[0103] The auditing layer comprises a picture auditing algorithm module, which is configured to recognize the proof materials exceeding the threshold value filtered by the filtering module through a pre-trained improved adaboost algorithm and to extract the seal information in the materials, and the specific process is described in the method for automatically auditing proof materials.

[0104] The review layer comprises a picture artificial review module and a picture auditing result module, wherein the artificial review module is configured to review whether the proof materials containing the seal audited by the auditing layer are incorrect, and the picture auditing result module is configured to output the result; the picture data that fails to pass the audit is fed back to the auditing layer as the training sample for the next model training.

[0105] It should be noted that the above content only illustrates the technical idea of the present application, and cannot be used to limit the protection scope of the present application. For those skilled in the art, a number of improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements fall within the protection scope of the claims of the present application.

Claims

1. A method for automatically verifying supporting documents, characterized in that, Includes the following steps: Step 1: Configure the template for supporting documents that the upload system requires to be reviewed, specifically including: Step 1.1: Collect the categories of supporting documents that the current business system requires to be reviewed; Step 1.2: Select the supporting documents that meet the criteria from the supporting documents collected in Step 1.1, and use them to create a supporting document template; Step 1.3: Mark whether the supporting document template created in Step 1.2 needs to be stamped and the type of supporting document, and set different coding values ​​for different types; Step 2: Store the supporting documents uploaded by the user and the corresponding code values ​​for the document types; Step 3: Calculate the similarity between the user-uploaded supporting documents and the template documents configured in Step 1; including: Step 3.1: Based on the type of supporting documents uploaded in Step 2, select templates for similarity calculation according to the code values ​​of the supporting documents uploaded by the user; Step 3.2: Calculate the similarity between the current user's proof materials and the template using a similarity matching algorithm. This includes the following steps: First, crop the image from the user's uploaded materials to make it the same size as the image in the template; then, convert the cropped image to grayscale; calculate the histogram data of the image, statistically analyze the probability distribution of identical pixels, and calculate the similarity between the histograms of the two images according to the correlation calculation formula. Step 4: Filter the supporting materials with low similarity in Step 3, and determine whether the similarity calculated in Step 3 exceeds the set threshold. If not, directly determine that the current supporting material does not meet the requirements and return the result information directly; if it is equal to or exceeds the set threshold, proceed to Step 5. Step 5: Automatically review the supporting documents filtered in Step 4; identify supporting documents exceeding the threshold in Step 4 using a pre-trained improved AdaBoost algorithm and extract the seal information from the documents, including: Step 5.1: Using the internet and supporting materials collected by the system itself, manually label the training data needed to improve the AdaBoost algorithm; Step 5.2: Using the manually labeled training dataset, perform improved AdaBoost training and obtain the trained classifier as the image review algorithm; Step 5.3: For the supporting materials in Step 4, the image review algorithm is used to give the specific recognition result, that is, whether the current supporting materials contain seal information. If they do, proceed directly to Step 6 for review; otherwise, return the result information directly. Step 6: Quickly verify whether the image of the supporting document containing the seal reviewed by the image review algorithm in Step 5 has any errors. If it passes the manual review, the result information is directly returned to the specific business system through the image review result module; if it fails the review, the result is notified to the specific business system. Step 7: The results are sent back to the business system through the image review result module in the review layer, thereby completing the review of the supporting materials and improving the review efficiency.

2. The method for automatically verifying supporting documents according to claim 1, characterized in that, In step 2, the supporting documents uploaded by the user are stored via HTTP API.

3. The method for automatically verifying supporting documents according to claim 1, characterized in that, The specific process of improving the AdaBoost algorithm in step 5 is as follows: I. Given a training sample set S = {(x1, y1), ..., (x m ,y m )}, weak separator space H, where x i Let ∈X be a sample vector, y = ±1 be the class label, and m be the total number of samples; initialize the sample probability distribution, D1(i) = 1 / m i = 1, ..., m; II. Iterate T times: (1) h in the weak classifier space H: a. Divide the sample space X into n equal parts, i.e., X1, X2, ..., Xn. n b. The probability distribution of the training samples is D t ,calculate p represents the probability value; c. Output of the weak classifier in the above partitioning in It is a dynamic positive constant; d. Calculate the normalization factor (2) Select an h from the weak classifier space. t Minimize Z (3) Update the probability distribution of each training sample to satisfy the normalization of the probability distribution. Z t As the normalization factor, it makes III. The final strong classifier H is Where b is a given threshold, with a default value of 0.

4. The method for automatically verifying supporting documents according to claim 1, characterized in that, In step 6, the image data that fails the review is fed back into the training dataset as training samples for the next model training.

5. An apparatus for automatically verifying supporting documents, used to implement the method for automatically verifying supporting documents as described in any one of claims 1-4, comprising: Configuration layer, access layer, filtering layer, auditing layer, review layer; The configuration layer includes a template upload module and a template configuration module. The template upload module is used to upload templates, and the template configuration module is used to configure different encoding values ​​according to different templates to distinguish template types, and to configure similarity thresholds. The access layer includes an image access interface module, which is used to store the supporting documents uploaded by the user via an HTTP API and to indicate the type of the supporting document. The filtering layer includes an image similarity calculation template and an image filtering module. The image similarity calculation template is used to calculate the similarity between the user-uploaded supporting materials and the template materials configured in the configuration layer. The filtering module is used to filter out supporting documents with low similarity and send supporting documents with high similarity to the review layer; The review layer includes an image review algorithm module, which is used to identify proof materials that exceed the threshold filtered out by the filtering module and extract the seal information in the materials through a pre-trained improved AdaBoost algorithm. The review layer includes a manual image review module and an image review result module; The manual review module is used to review whether there are any errors in the images of supporting documents containing seals reviewed by the review layer, and the results are output by the image review result module; the image data that fails the review is also fed back to the review layer as training samples for the next model training.

Images