Computer graphics-based original certificate identification method

Abstract

The present application relates to the field of image processing, and in particular to a computer graphics-based original certificate identification method, the method comprising the steps: setting different tags for a certificate image and a non-certificate image among historical acquired images, and denoising the historical acquired images so as to generate a certificate data set; according to an edge detection algorithm, obtaining acquired image edges of the denoised historical acquired images, the acquired image edges comprising an acquired image strong edge and an acquired image weak edge; calculating the similarity between the image edges and preset template image edges, and setting a loss function of a preset neural network model; using the certificate data set to train the neural network model and evaluating same so as to obtain an optimal model; and, in response to an image to be identified being acquired, generating an identification result. The present application uses the trained neural network model to identify whether an input image is an original certificate and achieves the effect of improving identification accuracy.

Description

A method for identifying original documents based on computer graphics Technical Field

[0001] The present application relates to the field of image processing, and in particular to a method for identifying original certificates based on computer graphics. Background Art

[0002] With the rapid development of artificial intelligence technology, artificial intelligence technology is being applied in more and more fields, such as using artificial intelligence methods to identify original certificates. In daily life, there are often scenarios where users need to take photos of their certificates and upload them for certification. When the business end receives the certificate photos uploaded by the user, it needs to determine whether the uploaded photos are original certificates. Photos that are not originals cannot pass the verification and handle business. Business is handled in real time, and it is necessary to quickly identify whether the certificates are originals.

[0003] In the related art, the patent document with patent publication number CN104573647B discloses a method and device for identifying illegal identity documents. The method obtains the deviation value of each pixel point through the mean square deviation of the RGB (Red, Green, Blue) components of each pixel point in the photo of the person on the identity document to be identified, and determines whether it is the original based on the size of the deviation value.

[0004] The related art method recognizes the original certificate by setting a threshold value, which is highly subjective and has a low recognition accuracy.

[0005] Summary of the Invention

[0006] In order to identify whether an input image is an original certificate through a trained neural network model and improve recognition accuracy, this application provides a method for identifying original certificates based on computer graphics, which adopts the following technical solutions:

[0007] A method for identifying original documents based on computer graphics, comprising the steps of:

[0008] The certificate images in the historical collected images are set with a first label, and the non-certificate images in the historical images are set with a second label, and the historical collected images are denoised to generate a certificate data set; according to the edge detection algorithm, the collected image edges of the denoised historical collected images are obtained, and the collected image edges include strong edges of the collected image and weak edges of the collected image; the similarity between the image edges and the preset template image edges is calculated, and the loss function of the preset neural network model is set, and the template image edges include strong edges of the template image and weak edges of the template image; the neural network model is trained and evaluated using the certificate data set to obtain an optimal model; and in response to the image to be recognized being acquired, a recognition result is generated.

[0009] Optionally, according to an edge detection algorithm, the edge of the certificate image of the historical collected image after denoising is obtained, including the steps of: converting the historical collected image after denoising into a grayscale image; performing edge detection on the grayscale image to obtain a strong edge collected image, a weak edge collected image, a strong edge of the collected image, and a weak edge of the collected image; calculating the Hessian matrix of each edge pixel point in the strong edge of the collected image and the weak edge of the collected image; calculating the eigenvalues ​​and eigenvectors of the Hessian matrix to obtain the maximum eigenvalue sequence of the strong edge of the collected image; calculating the distance between the strong edge of the collected image and the strong edge of the template image to obtain a first distance; calculating the distance between the weak edge of the collected image and the weak edge of the template image to obtain a second distance.

[0010] Optionally, the calculation of the first distance includes the steps of: calculating the distance between the strong edge of the acquired image and the strong edge of the template image based on the maximum eigenvalue sequence of the strong edge of the acquired image and the maximum eigenvalue sequence of the strong edge of the preset template image, and the distance calculation formula is: Among them, D ij Indicates the distance between the i-th pixel point on the edge of the captured image and the j-th pixel point on the edge of the template image. Indicates the maximum eigenvalue of the Hessian matrix of the i-th pixel point in the strong edge pixel value sequence of the collected image, The maximum eigenvalue of the Hessian matrix of the jth pixel point in the template image strong edge pixel value sequence, pv i The eigenvector representing the maximum eigenvalue of the Hessian matrix of the ith pixel in the sequence of strong edge pixel values ​​of the collected image, pv j The eigenvector corresponding to the maximum eigenvalue of the Hessian matrix of the jth pixel point in the template image strong edge pixel value sequence is represented; the dynamic regularization method is used to select each step distance, and the selection formula for each step distance is: d = min{D i+1，j , D i，j+1 , D i+1，j+1}, where d is the pixel distance, take D i+1，j , D i，j+1 , D i+1，j+1 The minimum value of D i+1，j D is the distance between the i+1th pixel point of the strong edge of the acquisition image and the jth pixel point of the strong edge of the template image. i，j+1 D is the distance between the i-th pixel point of the strong edge of the acquisition image and the j+1-th pixel point of the strong edge of the template image. i+1，j+1 is the distance between the i+1th pixel point of the strong edge of the captured image and the j+1th pixel point of the strong edge of the template image; the first distance between the strong edge of the captured image and the strong edge of the template image is obtained by adding up the selected distances of each step.

[0011] Optionally, the similarity between the edge of the captured image and the edge of a preset template image is calculated, and the loss function of the preset neural network model is set, including the steps of: counting the number of pixel points with pixel values ​​greater than zero in the strong edge captured image to obtain the number of strong edges in the captured image; counting the number of pixel points with pixel values ​​greater than zero in the weak edge captured image to obtain the number of weak edges in the captured image; and calculating the similarity between the edge of the captured image and the edge of the template image, the calculation formula is: Wherein, ρ is the similarity, Q s To collect the number of strong edges in the image, Q w To collect the number of weak edges in the image, TQ s is the number of strong edges in the template image, TQ w is the number of weak edges in the template image, D s is the distance between the strong edge of the acquisition image and the strong edge of the template image, D w is the distance between the weak edge of the collected image and the weak edge of the template image; set the loss function, the calculation formula is: LOSS i is the loss function, y' represents the probability of the neural network model predicting a certain category, y i represents the label of the i-th image in the license dataset, ρ i Indicates the penalty factor for the classification error of the i-th sample.

[0012] Optionally, a neural network model is trained and evaluated using the license data set to obtain an optimal model. The evaluation index calculation formula of the neural network model is: Among them, P is the precision rate of the neural network model, R is the recall rate of the neural network model, and F is the evaluation index of the neural network model.

[0013] This application has the following technical effects:

[0014] 1. By collecting the eigenvalues ​​and eigenvectors of the Hessian matrix of the strong edge of the image and the strong edge of the template image, the first distance is calculated using the dynamic regularization method. Similarly, the second distance of the Hessian matrix of the weak edge of the collected image and the weak edge of the template image is calculated. Based on the first distance and the second distance, the similarity of the edges in the collected image and the template image is calculated. The loss function of the neural network model is constructed based on the similarity, and the neural network model is trained. The trained model is used to identify the newly input image, so as to achieve the purpose of identifying whether the input image is the original certificate through the trained neural network model, thereby improving the recognition accuracy.

[0015] 2. Calculate the Hessian matrix of the edge in the strong edge map of the captured image, obtain the Hessian matrix of each edge pixel, calculate the eigenvalue of the Hessian matrix corresponding to each edge pixel and the eigenvector corresponding to the eigenvalue, the largest eigenvalue represents the curvature intensity of the edge point, and the eigenvector corresponding to the largest eigenvalue represents the direction of the curvature. Because the curvature and curvature direction of the image edge indicate the characteristics of the image edge, the distance of the image edge features can reflect the distance between the two image edges. When calculating the distance between the two image edges, since the number of pixels on the two image edges is different, this application uses a dynamic regularization method to calculate the distance between the two image edges. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and other objects, features and advantages of the exemplary embodiments of the present application will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0017] FIG1 is a method flow chart of steps S1 to S5 in a method for original certificate recognition based on computer graphics according to an embodiment of the present application.

[0018] FIG2 is a method flow chart of steps S20 - S25 in a method for original certificate recognition based on computer graphics according to an embodiment of the present application.

[0019] FIG3 is a method flow chart of steps S30 - S33 in a method for original certificate recognition based on computer graphics according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0021] It should be understood that when the terms "first," "second," etc. are used in the claims, specification, and drawings of this application, they are only used to distinguish different objects, rather than to describe a specific order. The terms "comprise" and "comprising" used in the specification and claims of this application indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or combinations thereof.

[0022] The present application embodiment discloses a method for identifying original certificates based on computer graphics. The specific scenario is: using image processing to identify original certificates. Referring to FIG1 , the method includes the following steps:

[0023] S1: Set the certificate images in the historical collected images as the first label, set the non-certificate images in the historical collected images as the second label, perform denoising on the historical collected images, and generate a certificate dataset.

[0024] Specifically, the historically collected images to be identified are collected. Historically collected images include certificate images and non-certificate images. Certificate images are original images of certificates, and the rest of the images are non-certificate images. In this application, non-original certificate images are also considered non-certificate images. Each historically collected image is given a label: 1 for a certificate image and 0 for a non-certificate image. After denoising the historically collected images using Gaussian filtering, a certificate dataset is obtained.

[0025] S2: According to the edge detection algorithm, the edges of the collected images of the denoised historical collected images are obtained. The collected image edges include strong edges and weak edges of the collected images. Referring to Figure 2, step S2 includes steps S20-S25:

[0026] S20: Convert the denoised historical image into a grayscale image.

[0027] S21: Perform edge detection on the grayscale image to obtain a strong edge acquisition image, a weak edge acquisition image, a strong edge of the acquisition image, and a weak edge of the acquisition image.

[0028] The Canny edge detection algorithm is used to extract image edge features from grayscale images. The Canny edge detection algorithm extracts strong edges and weak edges of grayscale images. In order to make the edge continuity better, if there is a point in the eight neighborhoods of the weak edge that is a strong edge, the weak edge is considered to be a strong edge.

[0029] The pixel values ​​of strong edge pixels are set to 1, and the pixel values ​​of other pixels are set to 0 to obtain the strong edge 0 / 1 map of the grayscale image. The strong edge 0 / 1 map of the grayscale image is multiplied by the grayscale image to obtain the strong edge acquisition image, and then the strong edge of the acquisition image is obtained. The strong edge of the acquisition image refers to the obvious edge feature or boundary in the strong edge acquisition image. Similarly, the weak edge acquisition image and the weak edge of the acquisition image can be obtained.

[0030] S22: Calculate the Hessian matrix of each edge pixel in the strong edge of the acquired image and the weak edge of the acquired image.

[0031] S23: Calculate the eigenvalues ​​and eigenvectors of the Hessian matrix to obtain the maximum eigenvalue sequence of the strong edge of the acquired image.

[0032] S24: Calculate the distance between the strong edge of the captured image and the strong edge of the template image to obtain a first distance.

[0033] Calculate the Hessian matrix of the strong edge of the captured image, obtain the Hessian matrix of each edge pixel in the captured strong edge image, calculate the eigenvalue of the Hessian matrix of each edge pixel and the eigenvector corresponding to the largest eigenvalue, the largest eigenvalue represents the curvature intensity of the edge pixel, and the eigenvector corresponding to the largest eigenvalue represents the direction of the curvature. Because the curvature and curvature direction of the edge indicate the characteristics of the edge, the distance of the edge feature can be used to reflect the distance between the two edges. Therefore, when calculating the distance between the strong edge of the captured image and the strong edge of the template image, since the number of pixels of the strong edge of the captured image and the strong edge of the template image is different, the edge distance cannot be directly calculated using Euclidean distance, etc. Therefore, the present application uses a dynamic regularization method to calculate the distance between the strong edge of the captured image and the strong edge of the template image. Referring to Figure 3, the calculation process includes steps S240-S242, which are as follows:

[0034] S240: Calculating the distance between the strong edge of the acquired image and the strong edge of the template image according to the maximum eigenvalue sequence of the strong edge of the acquired image and the preset maximum eigenvalue sequence of the strong edge of the template image.

[0035] For example, the maximum eigenvalue sequence corresponding to the strong edge of the captured image is (A, B, C, D, E), and the eigenvector corresponding to (A, B, C, D, E) is (a, b, c, d, e). The maximum eigenvalue sequence corresponding to the strong edge of the template image is (U, V, W, X, Y, Z), and the eigenvector corresponding to (U, V, W, X, Y, Z) is (u, v, w, x, y, z), as shown in Table 1:

[0036] Table 1:

[0037] D in the table ij Represents the distance between the i-th pixel point on the edge of the captured image and the j-th pixel point on the edge of the template image. Find the shortest distance from AU to EZ in the table, which is the distance between the strong edge of the captured image and the strong edge of the template image.

[0038] The formula for calculating each distance is:

[0039] Among them, D ij Indicates the distance between the i-th pixel point on the edge of the captured image and the j-th pixel point on the edge of the template image. Represents the maximum eigenvalue of the Hessian matrix of the i-th pixel point in the sequence of strong edge pixel values ​​in the collected image. Represents the maximum eigenvalue of the Hessian matrix of the jth pixel in the template image's strong edge pixel value sequence. iThe eigenvector representing the maximum eigenvalue of the Hessian matrix of the ith pixel in the sequence of strong edge pixel values ​​in the collected image. j The eigenvector corresponding to the maximum eigenvalue of the Hessian matrix of the j-th pixel point in the template image strong edge pixel value sequence.

[0040] It's PV j and pv i The cosine of the angle between the two eigenvectors is the projection of the eigenvalue onto the eigenvector. This calculation of the distance between two edges—that is, the distance between the i-th pixel on the edge of the captured image and the j-th pixel on the edge of the template image—takes into account not only the eigenvalue distance but also the eigenvector distance. This ensures that the closer the strong edges of the captured image and the template image are, the closer the distance. Similarly, the distance between weak edges of the captured image and the template image can be calculated.

[0041] S241: Use dynamic regularization method to select the distance of each step.

[0042] The formula for selecting the distance of each step is: d=min{D i+1，j , D i，j+1 , D i+1，j+1}

[0043] Among them, d is the pixel distance, take D i+1，j , D i，j+1 , D i+1，j+1 The minimum value of D i+1，j D is the distance between the i+1th pixel point of the strong edge of the acquisition image and the jth pixel point of the strong edge of the template image. i，j+1 D is the distance between the i-th pixel point of the strong edge of the acquisition image and the j+1-th pixel point of the strong edge of the template image. i+1，j+1 is the distance between the i+1th pixel point of the strong edge of the acquisition image and the j+1th pixel point of the strong edge of the template image.

[0044] S242: Add the selected distances of each step to obtain a first distance between the strong edge of the captured image and the strong edge of the template image.

[0045] S25: Calculate the distance between the weak edge of the captured image and the weak edge of the template image to obtain a second distance.

[0046] The calculation method of the second distance is the same as the first distance. The strong edge of the acquired image is replaced by the weak edge of the acquired image, and the strong edge of the template image is replaced by the weak edge of the template image. It will not be repeated here to obtain the distance between the weak edge of the acquired image and the weak edge of the template image.

[0047] S3: Calculate the similarity between the edge of the image and the edge of the preset template image, set the loss function of the preset neural network model, and the edge of the template image includes the strong edge of the template image and the weak edge of the template image.

[0048] S30: Count the number of pixels with pixel values ​​greater than zero in the strong edge acquisition image to obtain the number of strong edges in the acquisition image; S31: Count the number of pixels with pixel values ​​greater than zero in the weak edge acquisition image to obtain the number of weak edges in the acquisition image.

[0049] S32: Calculate the comprehensive distance between the edge of the captured image and the edge of the template image, that is, the similarity. The calculation formula is:

[0050] Among them, ρ is similarity, Q s To collect the number of strong edges in the image, Q w To collect the number of weak edges in the image, TQ s is the number of strong edges in the template image, TQ w is the number of weak edges in the template image, D s is the distance between the strong edge of the acquisition image and the strong edge of the template image, D w is the distance between the weak edge of the captured image and the weak edge of the template image;

[0051] S33: Set the loss function. The calculation formula of the loss function is:

[0052] LOSS i is the loss function, y' represents the probability of the neural network model predicting a certain category, y i represents the label of the i-th image in the license dataset, ρ i Represents the penalty factor for the classification error of the i-th sample. ρ i The smaller the value, the larger the loss function is, which means that the penalty for misclassifying images with small similarities between the edges of the captured image and the edges of the template image is greater.

[0053] S4: Use the license dataset to train and evaluate the neural network model to obtain the optimal model.

[0054] The license data set is randomly divided into training set and test set in a ratio of 4:1. The above loss function is used to optimize the neural network model by gradient descent. When the model training times reaches the set maximum number of training times or the loss is less than the preset threshold, the model training stops and the model evaluation is performed.

[0055] The calculation formula for the evaluation index of the neural network model is:

[0056] Where P is the precision of the neural network model, R is the recall of the neural network model, and F is the evaluation metric of the neural network model. Precision measures how many samples predicted as positive by the model are actually positive, while recall measures the proportion of actual positive samples correctly predicted by the model. The optimal model is selected based on the model evaluation metric F.

[0057] S5: In response to the image to be recognized being acquired, a recognition result is generated.

[0058] Take an image of the object to be detected, input it into the optimal model, and output the recognition result of whether the photo is the original certificate.

[0059] Although this specification has shown and described a number of embodiments of the present application, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, variations, and alternatives without departing from the spirit and scope of the present application. It should be understood that in practicing the present application, various alternatives to the embodiments of the present application described herein may be employed.

[0060] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for identifying original documents based on computer graphics, characterized in that: Includes steps: The certificate images in the historical collected images are set with a first label, and the non-certificate images in the historical images are set with a second label, and the historical collected images are denoised to generate a certificate data set; According to the edge detection algorithm, the edge of the collected image of the historical collected image after denoising is obtained, wherein the edge of the collected image includes a strong edge of the collected image and a weak edge of the collected image; Calculating the similarity between the edge of the image and the edge of a preset template image, and setting the loss function of the preset neural network model, wherein the edge of the template image includes a strong edge of the template image and a weak edge of the template image; Using the certificate data set to train a neural network model and evaluate it to obtain an optimal model; In response to the image to be recognized being acquired, generating a recognition result; According to the edge detection algorithm, the edge of the certificate image of the historical collected image after denoising is obtained, including the steps of: Convert the denoised historically collected images into grayscale images; Perform edge detection on the grayscale image to obtain a strong edge acquisition image, a weak edge acquisition image, a strong edge of the acquisition image, and a weak edge of the acquisition image; Calculate the Hessian matrix of each edge pixel in the strong edge of the acquired image and the weak edge of the acquired image; Calculate the eigenvalues ​​and eigenvectors of the Hessian matrix to obtain the maximum eigenvalue sequence of the strong edges of the captured image; Calculate the distance between the strong edge of the captured image and the strong edge of the template image to obtain a first distance; Calculate the distance between the weak edge of the captured image and the weak edge of the template image to obtain a second distance; The calculation of the first distance includes the following steps: According to the maximum eigenvalue sequence of the strong edge of the acquired image and the maximum eigenvalue sequence of the strong edge of the preset template image, the distance between the strong edge of the acquired image and the strong edge of the template image is calculated. The distance calculation formula is: Among them, D ij Represents the distance between the i-th pixel point on the edge of the captured image and the j-th pixel point on the edge of the template image. It represents the maximum eigenvalue of the Hessian matrix of the ith pixel point in the strong edge pixel value sequence of the collected image. Represents the maximum eigenvalue of the Hessian matrix of the jth pixel point in the template image strong edge pixel value sequence, pv i The eigenvector representing the maximum eigenvalue of the Hessian matrix of the ith pixel point in the sequence of strong edge pixel values ​​of the collected image, pv j The eigenvector corresponding to the maximum eigenvalue of the Hessian matrix of the jth pixel point in the template image strong edge pixel value sequence is represented; the dynamic regularization method is used to select each step distance, and the selection formula for each step distance is: d=min{D i+1,j ,D i，j+1 ,D i+1,j+1 } Among them, d is the pixel distance, take D i+1，j , D i，j+1 , D i+1，j+1 The minimum value of D i+1，j is the distance between the i+1th pixel point of the strong edge of the acquisition image and the jth pixel point of the strong edge of the template image, D i，j+1 is the distance between the i-th pixel point of the strong edge of the acquisition image and the j+1-th pixel point of the strong edge of the template image, D i+1，j+1 The distance between the i+1th pixel point of the strong edge of the acquisition image and the j+1th pixel point of the strong edge of the template image; Adding the selected distances of each step to obtain a first distance between a strong edge of the captured image and a strong edge of the template image; Calculate the similarity between the edge of the captured image and the edge of the preset template image, and set the loss function of the preset neural network model, including the steps of: Count the number of pixels whose pixel values ​​are greater than zero in the strong edge acquisition image to obtain the number of strong edges in the acquisition image; Count the number of pixels whose pixel values ​​are greater than zero in the weak edge acquisition image to obtain the number of weak edges in the acquisition image; Calculate the similarity between the edge of the captured image and the edge of the template image. The calculation formula is: Where, ρ is the similarity, Q s To collect the number of strong edges in the image, Q w To collect the number of weak edges in the image, TQ s is the number of strong edges in the template image, TQ w is the number of weak edges in the template image, D s is the distance between the strong edge of the acquisition image and the strong edge of the template image, D w The distance between the weak edge of the acquisition image and the weak edge of the template image; Set the loss function and the calculation formula is: LOSS i is the loss function, y' represents the probability of the neural network model predicting a certain category, y i represents the label of the i-th image in the certificate dataset, ρ i represents the penalty factor for the classification error of the i-th sample, ρ i The smaller the value, the larger the loss function is, which means that the greater the penalty for misclassifying images with small similarities between the edges of the captured image and the edges of the template image.

2. The method for identifying original documents based on computer graphics according to claim 1, characterized in that: The neural network model is trained and evaluated using the license data set to obtain the optimal model. The evaluation index calculation formula of the neural network model is: Among them, P is the precision of the neural network model, R is the recall of the neural network model, and F is the evaluation index of the neural network model. Mark.

Images