Electronic certificate forgery detection method and application in certificate monitoring

Abstract

The application discloses an electronic certificate forgery detection method, which comprises the following steps: performing median filtering and sharpening operation processing on obtained electronic certificate data, constructing a residual set containing sharpening residual error, sharpening-median filtering residual error and median filtering residual error by using the processed electronic certificate; performing discrete cosine conversion on the residual set to obtain the features of frequency domain variation before and after capturing the original image; using an autoregressive model and a Markov model to extract multiple residual features in the spatial domain to evaluate the detail difference between the electronic certificates before and after median filtering; finally, integrating the features from the spatial domain and the frequency domain to construct a feature set; inputting the extracted feature set into a support vector machine for training to obtain a median filtering detector to perform a detection task. The application solves the deficiencies of traditional electronic certificate forgery detection in low-resolution and high-compression image processing, has a high detection rate, and can be well applied in certificate monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of multimedia content security and license monitoring and counterfeiting, and particularly to an electronic license counterfeiting detection method based on double-domain multiple residual median filtering forensics and application in license monitoring. BACKGROUND

[0002] With the rapid progress of technology, users without professional training can easily use image convenience software to produce counterfeit images; electronic licenses, as a digital form of personal identity information, qualification certification and other legal documents, have gradually become an indispensable part of modern society. However, the popularity of electronic licenses also exposes them to the risk of counterfeiting, making them a target for criminals. These counterfeit electronic licenses, if misused, can have an undue impact on the judicial and financial fields, and also have a significant negative impact on normal market activities. Counterfeiters often use post-processing operations to conceal the modifications they have made, and techniques such as median filtering, Gaussian filtering and mean filtering can significantly affect the detection effect of counterfeit images, especially median filtering.

[0003] Median filtering is a widely used nonlinear operation for image noise removal, which improves the quality of the filtered image through smoothing processing, while the noise characteristics of the counterfeit area are also weakened compared to the original image, resulting in partial blurring of the edge information. For this reason, median filtering is often used as a counter-forensic means to hide potential forensic clues by blurring or eliminating subtle features in digital images. Therefore, research on detection techniques based on median filtering is of great significance for enhancing the counterfeit detection capabilities of electronic licenses.

[0004] The patent document with publication number CN114757877A discloses a deep counterfeiting detection method based on frequency domain filtering residual, belonging to the technical field of image classification, including a preprocessing module and a classification module, comprising the following steps: obtaining the low-frequency information map of the image by performing Haar wavelet transform on the image; performing residual operation on the gray image of the original image and the low-frequency information map to obtain the medium-high frequency information residual map of the original image; in the classification module, the original image and the medium-high frequency information residual map are spliced and then input into a convolutional neural network for classification processing. The characteristics of the medium-high frequency domain of the image are obtained through Haar wavelet transform and residual, so as to mine the counterfeit traces in the RGB domain that are weakened or contaminated due to image compression in the frequency domain, thereby improving the precision of the detection method when detecting compressed images. The RGB image and the high-frequency residual image are spliced together as the input of the convolutional neural network, fully utilizing the rich semantic information of the RGB domain and the detailed texture information of the medium-high frequency domain, so that the detection method has good robustness to image compression.

[0005] The patent document with the publication number CN105868778A discloses a median filter detection method based on frequency domain fluctuation characteristics: discrete Fourier transform processing is performed on the obtained original image to obtain a frequency domain image of the original image; the frequency domain image is processed to obtain a description point capable of reflecting the frequency domain fluctuation characteristics of the original image, and the obtained image after processing is divided into N annular regions, and then the total number of description points in each annular region is counted, that is, N-dimensional detection features are constructed; the extracted N-dimensional detection features are input into a support vector machine for training to obtain a support vector machine median filter detector, and then whether the original image has undergone median filtering is detected through the support vector machine median filter detector.

[0006] The existing methods are mostly concentrated on the utilization of spatial domain or frequency domain features, and ignore the mutual correlation between them, and perform poorly when processing small size images or low quality JPEG images.

[0007] Therefore, it is an urgent problem to provide an electronic certificate forgery detection method which can be applied to the certificate monitoring activity, especially an electronic certificate forgery detection method based on dual-domain multiple residual median filtering forensics. SUMMARY

[0008] In view of the problem that the median filter detection performs poorly on small size images or low quality JPEG images, the present application provides an electronic certificate forgery detection method based on dual-domain multiple residual median filtering forensics, which aims to combine spatial domain and frequency domain features, and simultaneously introduces multiple residual analysis to improve the accuracy and robustness of forged image detection, thereby providing more effective protection for the credibility of electronic certificates and providing help for electronic certificate monitoring and authenticity inspection.

[0009] The purpose of the present application is achieved as follows:

[0010] The electronic certificate forgery detection method based on dual-domain multiple residual median filtering forensics provided by the present application comprises the following steps:

[0011] Step 1): after the obtained electronic certificate image data is converted into a gray scale image, median filtering and sharpening operation processing are performed;

[0012] Step 2): a residual set containing sharpening residual, sharpening-median filtering residual and median filtering residual is constructed by using the processed image;

[0013] Step 3): discrete cosine conversion is performed on the residual set constructed in step 1) to obtain the features capturing the frequency domain changes before and after the original image; multiple residual features in the spatial domain are extracted by using an autoregressive model and a Markov model to evaluate the detail differences between the unaltered image and the median filtered image;

[0014] Step 4): Integrate the features from the spatial domain and frequency domain to construct a feature set and input it into a support vector machine for training to obtain a median filter detector to perform the detection task.

[0015] Step 2, which involves constructing the residual set, includes:

[0016] For the image after median filtering (I) MF The median filtered residual (I) is obtained by subtracting the original image (I) from the median filtered residual (I). MFR The formula is as follows:

[0017] I MFR =I MF -I

[0018] For the image after sharpening (I) S The sharpened residual (I) is obtained by subtracting the original image (I) from the sharpened image (I). SR The formula is as follows:

[0019] I SR =I S -I

[0020] For sharpening residual (I) SR ) and median filter residual (I MFR The difference is used to obtain the sharpened-median filter residual (I). SMR This is equivalent to the image after median filtering (I). MF ) and the image after sharpening (I S To calculate the difference, use the following formula:

[0021] I SMR =I SR -I MFR =I S -I MF ;

[0022] Step 3) involves obtaining the frequency domain features, including:

[0023] First, perform discrete cosine transformation on the three types of residuals obtained in step 1, and extract the maximum, minimum, average and variance of the resulting discrete cosine transformation coefficient matrix as features.

[0024] Step 3) describes the features extracted by the autoregressive model, including:

[0025] Before inputting the three residuals in the residual set into the autoregressive model, the data dimension of the image residual is first expanded to prepare for subsequent transposition and vectorization processing. Next, the augmented image is transposed, and the original image and the transposed image are vectorized and merged into a new vector, which is then used as input to the autoregressive model. Finally, the top ten orders of the autoregressive model output are taken as features.

[0026] The Markov model feature extraction of step 3) includes:

[0027] The first-order difference matrix in eight directions is used to calculate the changes between adjacent pixels, and a first-order difference array D is defined h (x,y) is as follows:

[0028] D h (x,y) = I(x+Δx,y+Δy) - I(x,y)

[0029] where (x,y) represents the pixel position in the image, and Δx and Δy represent the horizontal and vertical offsets, respectively; these offsets are taken from the set {(-1,-1),(-1,0),(-1,1),(0,-1),(0,1),(1,-1),(1,0),(1,1)}, corresponding to eight different directions, and the first-order difference matrix is subdivided into left, middle, and right matrices as inputs to the Markov model, with a threshold of 1; the results in the vertical, horizontal, and diagonal directions are divided into two groups to calculate the average values, as follows:

[0030] F1 = (F (-1,-1) +F (1,1) +F (-1,1) +F (1,-1) ) / 4

[0031] F2 = (F (-1,0) +F (1,0) +F (0,-1) +F (0,1) ) / 4

[0032] The average of the three residuals is calculated, and a Markov feature with a dimension of 108 is finally formed;

[0033] The training classification of step 4) includes:

[0034] The acquired image data is divided into 20 mutually exclusive subsets, of which 19 subsets are used as the training set, and the remaining subsets are used as the test set in each validation, and this process is repeated 20 times to obtain the average performance indicators of the algorithm on the entire data set. Meanwhile, a grid search is performed on the key parameters, the search range of the penalty factor C and the kernel function parameter γ is defined, and the cross-validation method is used to ensure the fairness of the performance evaluation.

[0035] The license forgery detection method is applied to electronic license monitoring.

[0036] Active beneficial effects: Based on the statistical differences of image residuals in different domains, the present application proposes the concepts of sharpening residuals and sharpening-medium filtering residuals, and forms a residual set together with medium filtering residuals. The multi-model strategy further enhances the feature extraction effects in the frequency domain and the spatial domain. Using the multi-model strategy, discrete cosine transform, autoregressive model and Markov model are used for feature extraction in the frequency domain and the spatial domain respectively, and the frequency domain features and the spatial domain features are combined to more sensitively capture the differences between the medium filtering processed image and the original image, thereby improving the accuracy of medium filtering detection. Specifically, multiple residual DCT statistical features are extracted in the frequency domain to capture detailed differences between the unchanged image and the medium filtering image. In the spatial domain, autoregressive model and Markov model are combined to extract features from multiple residuals. Finally, these features are combined into a feature set for classification. When facing small size images or low quality JPEG images, it can still maintain a high accuracy. At the same time, when facing other operations, it can also have a good discrimination ability for medium filtering operation and has strong robustness. This method has good performance on small size images or low quality JPEG images. It can provide help for electronic license monitoring and monitoring inspection in market activities, and has good application. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The detection feature extraction flowchart of the present application. DETAILED DESCRIPTION

[0038] In order to more clearly illustrate the object, technical scheme and advantages of the present application, the technical scheme in the embodiments of the present application will be described in detail below. The embodiments are only part of the embodiments of the present application, not all. Based on these embodiments of the present application, all other embodiments that can be obtained by those skilled in the art without creative work shall be included in the protection scope of the present application. The specific embodiments of the present application will be described in detail below with reference to the embodiment figures.

[0039] Embodiment 1

[0040] As shown in the figure, the embodiment of the present application provides an electronic license forgery detection method based on dual-domain multiple residual medium filtering forensics, which includes the following steps: Figure 1 Step 1): After converting the obtained electronic license image data into a gray image, perform medium filtering and sharpening operation processing;

[0041]

[0042] ​Step 2): constructing a residual set containing sharpening residual, sharpening- median filtering residual and median filtering residual using the processed image, the processing including:

[0043] The acquired image is first converted into a gray image, and then the gray image is subjected to median filtering and sharpening operation processing. The median filtering processed image (I MF ) is subtracted from the original image (I) to obtain the median filtering residual (I MFR ), and the formula is as follows:

[0044] I MFR = I MF - I

[0045] The sharpening processed image (I S ) is subtracted from the original image (I) to obtain the sharpening residual (I SR ), and the formula is as follows:

[0046] I SR = I S - I

[0047] The sharpening residual (I SR ) is subtracted from the median filtering residual (I MFR ) to obtain the sharpening- median filtering residual (I SMR ), which is equivalent to subtracting the median filtering processed image (I MF ) from the sharpening processed image (I S ), and the formula is as follows:

[0048] I SMR = I SR - I MFR = I S - I MF ;

[0049] Step 3: performing discrete cosine conversion on the residual set to obtain features capturing the frequency domain changes before and after the original image, and the formula for the discrete cosine conversion is as follows:

[0050]

[0051] wherein M and N are the sizes of the matrix rows and columns, respectively; the average value, variance, maximum value and minimum value of the obtained discrete cosine transform coefficient matrix are extracted as frequency domain features;

[0052] Using the autoregressive model and the Markov model to extract multiple residual features in the spatial domain to evaluate the detail differences between the unchanged image and the median filtering image, and the specific details are as follows:

[0053] Before inputting the three residuals in the residual set into the autoregressive model, first expand the data dimension of the image residual to perform transposition and vectorization processing, then transpose the expanded image, and combine the original image with the transposed image after vectorization into a new vector, which will be used as the input of the autoregressive model. Finally, extract the top ten order results from the model output as the autoregressive features.

[0054] For the three residuals in the residual set, a first-order difference matrix in eight directions is used to calculate the changes between adjacent pixels. Define the first-order difference array D h (x,y) as follows:

[0055] D h (x,y) = I(x + Δx, y + Δy) - I(x, y)

[0056] where (x, y) represents the pixel position in the image, and Δx and Δy represent the horizontal and vertical offsets, respectively, which are taken from the set {(-1, -1), (-1, 0), (-1, 1), (0, -1), (0, 1), (1, -1), (1, 0), (1, 1)}, corresponding to eight different directions. The first-order difference matrix is subdivided into left, middle, and right matrices as inputs for the Markov model, with a threshold of 1. The results in the vertical, horizontal, and diagonal directions are divided into two groups to calculate the average values, as follows:

[0057] F1 = (F (-1,-1) +F (1,1) +F (-1,1) +F (1,-1) ) / 4

[0058] F2 = (F (-1,0) +F (1,0) +F (0,-1) +F (0,1) ) / 4

[0059] The average of the three residuals results in a Markov feature with a dimension of 108.

[0060] Step 4): Integrate the features in the spatial and frequency domains to construct a feature set and input it into the support vector machine for training to obtain a median filter detector for detection tasks.

[0061] F = [F spatial ,F frequency ]

[0062] The acquired image data is divided into 20 mutually exclusive subsets, 19 of which are used as the training set and the remaining subset is used as the test set in each validation. The process is repeated 20 times to obtain the average performance of the algorithm on the entire data set. Meanwhile, a grid search is performed on the key parameters, the search range of the penalty factor C and the kernel function parameter gamma is defined, and the cross-validation method is used to ensure the fairness of the performance evaluation. The classification performance is evaluated by the detection accuracy (ACC), and the ROC curve is used to illustrate the true positive rate (TPR) of the detector at a low false positive rate (FPR).

[0063] The license forgery detection method is applied to electronic license monitoring.

[0064] The present application is based on the statistical difference of image residual in different domains, and proposes the concepts of sharpening residual and sharpening-medium filtering residual, which together with the medium filtering residual form a residual set. The multi-model strategy is used to further enhance the feature extraction effect in the frequency domain and the spatial domain. Using the multi-model strategy, discrete cosine transform, autoregressive model and Markov model are used in the frequency domain and the spatial domain respectively for feature extraction, and the frequency domain features and the spatial domain features are combined to more sensitively capture the differences between the medium filtering processed image and the original image, thereby improving the accuracy of medium filtering detection. Specifically, multiple residual DCT statistical features are extracted in the frequency domain to capture the detailed differences between the unchanged image and the medium filtering image. In the spatial domain, autoregressive model and Markov model are combined to extract features from multiple residuals. Finally, these features are combined into a feature set for classification. When facing small size images or low quality JPEG images, it can still maintain high accuracy. At the same time, when facing other operations, it can also have good discrimination ability for medium filtering operation and has strong robustness. This method has good performance on small size images or low quality JPEG images. It can provide help for electronic license monitoring and monitoring inspection in market activities, and has good application.

[0065] The above embodiments are only used to clarify the technical solutions of the present application, and do not limit the same. Those skilled in the art can adjust the technical solutions of the embodiments, or replace some of the technical features. Without changing the core of the corresponding technical solution, many modifications can also be made, which should be included in the protection scope of the present application.

Claims

1. A method of electronic passport forgery detection, characterized in that, The method comprises the following steps: Step 1): the acquired electronic license image data is converted into a gray image, and then median filtering and sharpening operation processing are performed; Step 2): a residual set containing a sharpening residual, a sharpening-median filtering residual and a median filtering residual is constructed by using the processed image; Step 3): discrete cosine conversion is performed on the residual set constructed in step 1) to obtain features capturing the frequency domain changes before and after the original image; a plurality of residual features in the spatial domain are extracted by using an autoregressive model and a Markov model to evaluate the detail difference between the unchanged image and the median filtering image; Step 4): the features in the spatial domain and the frequency domain are integrated to construct a feature set and input into a support vector machine for training to obtain a median filtering detector to perform a detection task; Step 2) constructing the residual set includes: for the image after median filtering processing and the original image to get the median filtering residual , the formula is as follows: ， For the image after sharpening processing and the original image to obtain the sharpening residual The formula is as follows: , For sharpening residual With median filter residual Subtracting to get sharpening - median filter residual residual , which is equivalent to the image after median filter processing With the image after sharpening processing Subtracting, the formula is as follows: ； The step 3 autoregressive model feature extraction includes: Before the three kinds of residuals in the residual set are input into the autoregressive model, the residual matrix is preprocessed as follows: First, the data dimension of the image residual is expanded to prepare for subsequent transposition and vectorization processing; next, the augmented image is transposed, and the original image and the transposed image are vectorized and combined into a new vector; finally, these vectors will be used as the input of the autoregressive model; the first ten order results of the autoregressive model output are taken as the features; The step 3 Markov model feature extraction includes: First, the change between adjacent pixels is calculated using a first-order difference matrix in eight directions; define the first-order difference array The formula is as follows: ， wherein denotes a pixel position in the image, and and denote horizontal and vertical offsets, respectively; these offsets are taken from the set ， corresponding to eight different directions; the first-order difference matrix is subdivided into left, middle and right matrices as the input of the Markov model, and the threshold is set to 1; the results of the vertical, horizontal and diagonal directions are divided into two groups to calculate the average values, as follows: The average of three residuals is obtained, and finally the Markov features with dimension 108 are formed.

2. The electronic passport forgery detection method of claim 1, wherein, The step 3 frequency domain feature acquisition includes: the three kinds of residuals in the residual set are respectively processed by discrete cosine conversion, and the obtained discrete cosine conversion coefficient matrix is respectively extracted to obtain the maximum value, the minimum value, the average value and the variance as the features.

3. The electronic passport forgery detection method of claim 1, wherein, The step 4 training classification includes: the acquired image data is divided into 20 mutually exclusive subsets, of which 19 subsets are used as the training set, and the remaining subsets are used as the test set in each validation; the process is repeated 20 times to obtain the average performance indicators of the algorithm on the entire data set; at the same time, a grid search is performed on the key parameters, the search range of the penalty factor C and the kernel function parameter γ is defined, and the cross-validation method is used to ensure the fairness of the performance evaluation.

4. The electronic license forgery detection method according to any one of claims 1-3, applied to electronic license monitoring.

Images