Paper surface rough texture anti-counterfeiting method

Abstract

The application provides a paper surface rough texture-based anti-counterfeiting method, and specifically comprises the following steps: planning an effective texture area on a paper product surface; installing a micro microscope lens on a mobile phone camera part to collect area texture sample images; using a feature point algorithm to extract sample texture feature point descriptors in an App software; using a support vector regression algorithm to estimate a robust mapping function between the feature point descriptors, using the robust mapping function to extract matching feature point pairs, and completing image registration; establishing a plane image conversion matrix based on the obtained matching feature point pairs, realizing sample similar area superposition fusion; and evaluating the difference of the similar areas by means of surface topography difference parameters; the texture anti-counterfeiting principle is simple and has extremely low reproducibility, the anti-counterfeiting equipment is simple and portable and has low price, the feature registration algorithm and the similar area evaluation method are intelligent and reliable; from equipment manufacturing, sample collection to anti-counterfeiting identification process, the method is convenient, intelligent and self-service.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of paper products and commodity anti-counterfeiting, in particular to a paper surface rough texture anti-counterfeiting method. BACKGROUND

[0002] Paper products or commodities often face the problem of counterfeit products in the process of use and circulation, such as fake contracts, fake invoices, fake wills and fake commodities, fake packaging, etc. Counterfeit documents hinder social safety and economic order, and counterfeit commodities infringe the legitimate rights and interests of manufacturers and consumers, with direct and indirect consequences difficult to estimate. The use of anti-counterfeiting technology aims to effectively identify fake documents and fake commodities, and to maintain social and economic order, public safety and the legitimate rights and interests of citizens.

[0003] Current paper product anti-counterfeiting technologies mainly include two-dimensional codes, laser holographic technology, watermarked paper, chemical materials and radio frequency (RFID) anti-counterfeiting technology. The semantic information of two-dimensional codes is easy to interpret and widely used, but it is also easy to be copied. Laser holographic technology, watermarked paper and chemical materials mainly rely on anti-counterfeiting materials and complex marks, which have high production costs, complex technology and are not conducive to large-scale promotion. Radio frequency (RFID) anti-counterfeiting technology requires relatively expensive equipment, which is more difficult for ordinary consumers to identify. Another proposed surface texture anti-counterfeiting method is based on expensive microscope equipment, has low intelligence and poor self-identification. SUMMARY

[0004] The purpose of the present application is to provide a convenient intelligent surface texture anti-counterfeiting method. The method plans a texture area on the surface of paper products as an anti-counterfeiting information area, uses a mobile phone with a microscope lens as an information collector to collect surface rough texture samples, and finally develops an APP software based on a feature registration algorithm and a region evaluation algorithm to realize intelligent identification. The APP software uses a feature point extraction algorithm to extract feature point descriptors, estimates a robust mapping function between the feature point descriptors based on a support vector regression algorithm (SVR), extracts matching feature point pairs using the robust mapping function, and completes image registration. The method of extracting matching feature point pairs using the robust mapping function model has good repeatability and can be used as a quantitative judgment basis. A more reliable qualitative judgment method is to perform image fusion based on the matching feature point pairs, evaluate the differences in similar regions, and use the surface topography difference parameters as a qualitative judgment basis.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A paper surface rough texture anti-counterfeiting method, specifically comprising the following steps:

[0007] S1, planning an effective texture area on the surface of paper products, which can be marked with a box, a circle or a dot, or the approximate position can be specified;

[0008] S2, install a microscopic lens on the camera of the mobile phone, and acquire the sample image of the region texture in step S1 by using the device;

[0009] S3, extract the feature point descriptor of the image texture in step S2 by using the SIFT algorithm;

[0010] S4, estimate the robust mapping function between the feature point descriptors in step S3 by using the support vector regression (SVR) algorithm, extract the matching feature point pair by using the robust mapping function, and complete the image registration;

[0011] S5, establish the plane image conversion matrix based on the matching feature point pair obtained in step S4, and realize the superposition and fusion of the similar regions of the sample;

[0012] S6, evaluate the difference of the similar regions by using the surface topography difference parameter;

[0013] In steps S1 and S2, the mobile phone with the added microscopic lens is used as a portable device, the surface rough texture is used as a feature, and in steps S3-S6, the APP software is used to write the recognition algorithm, so as to realize the sample collection, qualitative identification and anti-counterfeiting of paper products.

[0014] In step S3, the SIFT algorithm is used to extract the feature point descriptor, which specifically includes the following contents:

[0015] ① define a two-dimensional image scale space , wherein represents the convolution operation in the x and y directions, , is a scale variable Gaussian function, is a spatial coordinate, is a scale coordinate;

[0016] ② make Gaussian blur of different scales on the sample image, and obtain a series of images with different sizes by continuously downsampling the original image, and arrange these images from large to small and from bottom to top to form a tower structure; the structure is divided into O groups, and each group has S layers, that is, a Gaussian pyramid is formed;

[0017] ③ subtract two adjacent images in the same group of the Gaussian pyramid to obtain a Gaussian difference pyramid;

[0018] ④ search each pixel point in the Gaussian difference pyramid and compare it with the surrounding 26 points to determine the feature point;

[0019] V. The curvature at the feature points is calculated by using the Hessian matrix to eliminate unstable edge response points;

[0020] VI. For the determined feature points, the eigenvectors thereof are calculated, and the specific steps are as follows:

[0021] VII. According to the obtained eigenvectors, the Euclidean distance method is used to perform initial matching of the feature points.

[0022] The step S4 of estimating the robust mapping function between the feature point descriptors based on the support vector regression algorithm (SVR) specifically includes the following contents:

[0023] I. Robust mapping function model

[0024] According to the homography assumption of images, two images and should have a one-to-one mapping relationship between the corresponding points, and can be expressed by a homography function; therefore, there are two homography functions that can map a point in one image to the corresponding point in the other image

[0025] ;

[0026] If can at least satisfy one of the two homography functions, i.e. can be mapped to by the function , and vice versa, then can be regarded as a pair of corresponding points of the images; based on the above homography function concept, the mapping functions and can be regarded as two vector-valued functions

[0027] ,

[0028] ;

[0029] wherein, and ( ) are general scalar functions;

[0030] Suppose is the initial matching pair set

[0031] ;

[0032] The projection of in the subspace is defined as

[0033] ;

[0034] For any , if satisfy the mapping function , will satisfy , similarly, , and can be calculated from the corresponding homologous subspace , and ;

[0035] 2. Support Vector Regression Algorithm (SVR)

[0036] Consider a data set , where represents a D-dimensional input vector, represents the target output, i.e. the class label, and the nonlinear relationship between the input and output can be described by the following regression function

[0037] (1)

[0038] where represents the prediction function, is a nonlinear mapping function, and are two coefficients to be determined, and are estimated by minimizing the regularized function

[0039] ,

[0040] (2)

[0041] In equation (2), the first term is the regularization term, and the second term describes the empirical error, which is estimated by insensitive loss function , the parameter C is the weighting parameter between the first and second terms in the equation, and the parameter can be regarded as equivalent to the distance size of the approximate accuracy in the training data; by introducing the slack variables and , the following constraint condition represented by the following equation can be introduced

[0042] ;

[0043] The positive and negative distance sizes are represented by and , respectively, and outside the area of , the slack variables take non-zero values, and the final goal of the optimization method is to find the function , which has a maximum deviation of from the actual function for all training set data, and is as flat as possible;

[0044] Equation (1) can be further expressed in the following more general form

[0045] (3)

[0046] where, , is a Lagrange multiplier term, satisfying The data points are support vectors, which can be used to construct the decision function; the kernel function can map the input space to a higher dimensional space, where the most commonly used kernel function for SVR is the radial basis kernel function, which plays an important role in both classification and regression prediction;

[0047] (4)

[0048] In equation (4), is the kernel parameter; there is only one hyperparameter to be determined in the radial basis kernel function, which makes it have good performance in predicting nonlinear functions in SVR;

[0049] ③Support vector regression algorithm can be used to estimate and scalar function

[0050] The specific mapping function iteration process is as follows:

[0051] a) From the initial feature point pair set , estimate by the SVR method, denoted as ;

[0052] b) Calculate the mapping residual of the corresponding point pair based on the current estimated value;

[0053] b) Calculate the mapping residual of the corresponding point pair based on the current estimated value;

[0054] c) According to the size of , use the robust weight function to determine the weight of each corresponding point pair ;

[0055] d) Update the weight and re-estimate the mapping function ;

[0056] Loop iteration steps (b) to (d) until the iteration termination condition is met, and obtain the mapping function and the final feature point pair set;

[0057] When predicting the above vector-valued function, three parameters , and are determined, whose optimal values depend on the initial matching points and the optimization problem itself; the kernel parameter represents the radius of the kernel function, and the regularization parameter is the penalty parameter, which balances the flatness of the final solution and the training error; in practice, increasing the value of can minimize the error, but a value that is too large will lose the generalization performance of the model, i.e., the function estimation accuracy, and increase the complexity of the model; the grid search method and cross-validation are used to select the optimal parameters to avoid overfitting problems; in addition, the size of in the insensitive loss function will also affect the performance of SVR, and the number of support vectors used to construct the regression function is negatively correlated with In multiple matching experiments, the matching algorithm proposed in this paper, the value does not strongly depend on the initial matching points, and finally =0.25 is selected;

[0058] ④ Outlier rejection, complete image registration

[0059] After obtaining the scalar function from SVR, the residual error is calculated according to the mapping function

[0060] (5)

[0061] The robust weight function assigns different weights to each point through the mapping residual error obtained by equation (5) , the weight in the middle part of the data is relatively large, and then decreases to zero at a certain point on both sides; the Tukey double weight estimation function is as follows

[0062] ;

[0063] where =4.4478 × median , for =1,..., , i.e., the residual error , so the outliers with small weights will be rejected;

[0064] At the same time, the iterative estimation of the mapping function needs to determine the termination condition

[0065] ;

[0066] and respectively represent the size of the variance value before and after the rejection of outliers in one iteration; the threshold value is set , the iteration estimation process is terminated; in addition, the mean square error visually displays the consistency of the corresponding points and the estimated mapping function; therefore, the threshold value is set , the iteration is terminated when the condition is met and are set to 75 and 0.3, respectively.

[0067] The sample similar region superposition fusion in the step S5 specifically includes the following contents:

[0068] The mapping function obtained based on the support vector regression algorithm is a high-dimensional space mapping function; in order to more intuitively describe the actual position mapping relationship between images, a three-dimensional space homography matrix can be established based on the matching point pairs H , and the coordinate transformation relationship thereof can be generally described as:

[0069] ;

[0070] wherein H contains 8 degrees of freedom, and since the image is limited to a plane, the homography matrix is further simplified as:

[0071] ,

[0072] ;

[0073] The matrix needs 4 pairs of matching points to calculate a unique solution through singular value decomposition (SVD), that is, to obtain 8 unknown necessary parameters of the matrix; the RanSAC algorithm is introduced to iteratively filter out matching points with large errors, and then the matrix is solved;

[0074] The above feature point matching method has good robustness to resolution, brightness, and camera parameter influence;

[0075] The transformation matrix H is used, that is, the two sample images are fused.

[0076] The difference of the similar region in the step S6 specifically includes the following contents

[0077] For the evaluation of the similar region, cross-correlation operation can be used to achieve, but the cross-correlation operation can be affected by the invalid region, such as the cross-correlation value of the region with an amplitude of 0 is 100%; in order to overcome this problem, the surface topography difference parameter is used to describe the relative difference of the similar region; let the similar regions be and ,​​ , then

[0078] ;

[0079] wherein, is the surface metrology general parameter: root mean square roughness; similar area and The greater the difference, The greater the value, and vice versa; when and completely equal, then =0, that is, no difference.

[0080] Compared with the prior art, the present application has the following advantages:

[0081] The present application is based on the paper surface rough texture anti-counterfeiting method, the surface texture anti-counterfeiting principle is simple, random, and the reproducibility is extremely low; based on the mobile phone with microscope lens, the sample can be collected quickly and conveniently; in the mobile phone APP, the feature point matching fusion program and the similar area evaluation program are written, which meets the requirements of paper product identification. The anti-counterfeiting material and the identification process are simple and easy to operate, which greatly reduces the cost and is convenient for wide application. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 is the flow chart of the preferred embodiment of the paper surface rough texture anti-counterfeiting method of the present application;

[0083] Figure 2 is a self-help portable device used in the paper surface rough texture anti-counterfeiting method of the present application, that is, a mobile phone with a microscope lens example;

[0084] Figure 3 is a rough texture sample 1 of a copy paper surface collected by a mobile phone microscope lens in the paper surface rough texture anti-counterfeiting method of the present application;

[0085] Figure 4 is a rough texture sample 2 of a copy paper surface collected by a mobile phone microscope lens in the paper surface rough texture anti-counterfeiting method of the present application;

[0086] Figure 5 is a sample feature point matching result graph in the paper surface rough texture anti-counterfeiting method of the present application;

[0087] Figure 6 is a sample similar area fusion superposition area graph in the paper surface rough texture anti-counterfeiting method of the present application;

[0088] Figure 7 is a support vector regression algorithm parameter schematic diagram. DETAILED DESCRIPTION

[0089] To clarify the technical problems, technical solutions, implementation processes and performance displays, the following embodiments are further described in detail. It should be understood that the specific embodiments described here are only used to explain the present application, and are not intended to limit the present application. Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference signs in the drawings represent functionally identical or similar elements. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

[0090] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations.

[0091] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the specific embodiments below. Those skilled in the art should understand that the present disclosure can also be implemented without some specific details. In some examples, methods, means, elements and circuits that are well known to those skilled in the art are not described in detail, in order to highlight the main ideas of the present disclosure. Embodiment 1

[0092] As shown in Figure 1 A paper surface rough texture anti-counterfeiting method, specifically comprising the following steps:

[0093] S1, planning an effective texture area on the surface of a paper product, which can be marked with a box, a circular frame or a dot, or can be specified as an approximate position;

[0094] S2, installing a microscopic microscope lens on the camera of a mobile phone, and using the device to obtain a texture sample image of the area in step S1;

[0095] S3, using SIFT algorithm to extract the feature point descriptor of the image texture in step S2;

[0096] S4, using Support Vector Regression (SVR) algorithm to estimate the robust mapping function between the feature point descriptors in step S3, and using the robust mapping function to extract the matching feature point pairs to complete image registration;

[0097] S5, establishing a plane image conversion matrix based on the matching feature point pairs obtained in step S4, and realizing sample similar area superposition fusion;

[0098] S6, evaluating the difference of the similar area by means of surface topography difference parameters;

[0099] The mobile phone is used as a portable device in steps S1 and S2, and the surface roughness is used as a feature, and the APP software is used to write an identification algorithm in steps S3-S6, so as to realize paper sample collection, qualitative identification and anti-counterfeiting.

[0100] The step S1 specifically comprises:

[0101] ①A circle frame, a square frame or a dot can be printed, or only a region is specified;

[0102] ②The region is used as an identification texture, and the area of the region is about (1-5) mm 2 That's all.

[0103] The step S2 specifically comprises:

[0104] ①A micro microscope lens is installed on the camera of the mobile phone, and the pixel resolution is higher than 10 μm, as shown in Figure 2 .

[0105] ②The micro device is used to collect a clear image of the texture of the effective region, and the image has at least 1000*1000 pixels, as shown in Figure 3 and Figure 4 .

[0106] The SIFT algorithm is used to extract feature point descriptors in the step S3, and the step specifically comprises the following contents:

[0107] ①A two-dimensional image scale space is defined , wherein represents convolution operations in x and y directions, , is a scale variable Gaussian function, is a spatial coordinate, is a scale coordinate;

[0108] ②The sample image is blurred by a Gaussian function with different scales, and the original image is continuously down-sampled to obtain a series of images with different sizes, and the images are arranged in a tower structure from large to small and from bottom to top; the structure is divided into O groups, and each group has S layers, that is, a Gaussian pyramid is formed;

[0109] ③The images in the same group of the Gaussian pyramid are subtracted two by two to obtain a Gaussian difference pyramid;

[0110] ④Each pixel point in the Gaussian difference pyramid is searched and compared with 26 points around the pixel point to determine a feature point, and the determination is specifically as follows:

[0111] Traverse each pixel point on the Gaussian difference pyramid, and compare it with its 8 adjacent points of the same scale and 9*2 adjacent points of the upper and lower scales, if it is a maximum or minimum point, it is temporarily identified as a feature point;

[0112] ⑤The curvature at the feature point is calculated by the Hessian matrix to eliminate unstable edge response points, and the specific steps are as follows:

[0113] S351, the image in the Gaussian difference pyramid produces strong edge response, and the unstable edge response points need to be eliminated, which have a larger principal curvature in the direction across the edge, and a smaller principal curvature in the direction perpendicular to the edge; the principal curvature can be calculated by a 2*2 Hessian matrix H is obtained:

[0114]

[0115] In the formula, is a differential operator;

[0116] S352, let be the maximum eigenvalue, be the minimum eigenvalue; the sum of them is calculated by the trace of the matrix, and the product of them is calculated by the determinant of the matrix: H H

[0117] S353, let , then

[0118] The value of the formula is the smallest when the two eigenvalues are equal, and the larger the value is, the larger the ratio of the two eigenvalues is, which further indicates that the gradient value of the feature point in one direction is larger, and the gradient value in the other direction is smaller, which is the case of edge response; therefore, in order to eliminate the edge response points, only is needed; set a proper value; the feature points satisfying the above formula are retained, and the feature points not satisfying the above formula are eliminated;

[0119] ⑥For the determined feature points, the characteristic vectors are calculated, and the specific steps are as follows:

[0120] S361, the gradient direction distribution characteristics of the neighborhood pixels of the feature points are used to specify the direction parameters for each feature point,

[0121]

[0122] The above formula is the modulus and direction of the gradient at

[0123] ​​​​​​​S362、After the gradient calculation of the feature point is completed, the gradient direction and amplitude of the pixels in the neighborhood are counted using a histogram; the gradient direction histogram divides the range of 0°-360° into 36 columns, with each column being 10°; finally, the peak direction of the histogram is taken as the main direction of the feature point, and the directions reaching 80% of the peak value are taken as the auxiliary directions;

[0124] S363、The gradient of each pixel in the 4x4=16 window around the feature point is calculated, and the weight of the pixels far from the center is reduced using a Gaussian descending function, and finally a 128-dimensional feature description vector is formed;

[0125] ⑦According to the obtained feature vector, the initial matching of the feature points is performed using the Euclidean distance method, and the specific steps are as follows:

[0126] S371、Take a feature point in the image, and find the first two feature points in the image to be registered that have the closest Euclidean distance to the feature description vector of the feature point;

[0127] S372、Among the two feature points, if the ratio of the closest distance to the second closest distance is less than a certain proportion threshold (0.6-0.9), the pair of matching points is accepted.

[0128] The step S4 of estimating the robust mapping function between the feature point descriptors based on the support vector regression algorithm (SVR) specifically includes the following contents:

[0129] ① Robust mapping function model

[0130] According to the homography assumption of the image, two images and should have a one-to-one mapping relationship between the corresponding points, and can be expressed by a homography function; therefore, there are two homography functions that can map a point in one image to the corresponding point in the other image

[0131] ;

[0132] If can at least satisfy one of the two homography functions, i.e. can be mapped to by the function , and vice versa, then can be regarded as a pair of corresponding points of the image; based on the above homography function concept, the mapping functions and can be regarded as two vector-valued functions constituting

[0133] ,

[0134] ;

[0135] wherein, and ( is a general scalar function;

[0136] Assume is the initial matching pair set

[0137]

[0138] It is defined as follows the projection of the subspace

[0139]

[0140] For any , if satisfies the mapping function , then will satisfy , and , and can be calculated from the corresponding homologous subspace , and ;

[0141] 2. Support vector regression algorithm (SVR)

[0142] Consider the data set , where represents a D-dimensional input vector, represents the target output, i.e., the class label, and the nonlinear relationship between input and output can be described by the following regression function

[0143] (1)

[0144] where represents the prediction function, is a nonlinear mapping function, and are two coefficients to be determined, and are estimated by minimizing the regular function

[0145] ,

[0146] (2)

[0147] In equation (2), the first term is the regularization term, and the second term describes the empirical error, which is estimated by insensitive loss function , and the parameter C is the weighting parameter between the first and second terms in the equation, and the parameter ​​​The distance size of the approximation accuracy in the training data is regarded as equivalent; the slack variable is introduced and The constraint condition expressed by the following equation can be derived

[0148] ;

[0149] As shown in Figure 7 , the described positive and negative distance sizes are represented by and respectively, outside the area of , the slack variable adopts a non-zero value, and the final goal of the optimization method is to find the function , whose maximum deviation from the actual function for all training set data is , and at the same time, as flat as possible;

[0150] Equation (1) can be further expressed in the following more general form

[0151] (3)

[0152] wherein , is the Lagrange multiplier term, satisfying The data points satisfying are support vectors, which can be used to construct the decision function; the kernel function can map the input space to a higher dimensional space, where the most commonly used kernel function for SVR is the radial basis kernel function, which plays an important role in classification and regression prediction;

[0153] (4)

[0154] In equation (4), is the kernel parameter; there is only one hyperparameter to be determined in the radial basis kernel function, which makes it have good performance in predicting nonlinear functions in SVR;

[0155] ③ The support vector regression algorithm can be used to estimate the scalar function in

[0156] The specific mapping function iteration process is as follows:

[0157] a) From the initial feature point pair set , estimate by the SVR method, denoted as ;

[0158] b) Calculate the mapping residual of the corresponding point pair based on the current estimated value;

[0159] b) Calculate the mapping residual of each corresponding point pair based on the current estimate ;

[0160] c) Determine the weight of each corresponding point pair according to the size of the mapping residual ;

[0161] d) Update the weight and re-estimate the mapping function ;

[0162] Loop the iteration steps (b) to (d) until the iteration termination condition is satisfied, and obtain the mapping function and the final set of corresponding point pairs

[0163] When the SVR predicts the above vector-valued function, three parameters , and are determined, and their optimal values depend on the initial matching point pairs and the optimization problem itself; the kernel parameter represents the radius of the kernel function, and the regularization parameter is a penalty parameter used to balance the flatness of the final solution and the training error; in practice, increasing the value of can minimize the error, but if the value of is too large, the generalization performance of the model, i.e., the accuracy of the function estimate, will be lost, and the complexity of the model will be increased; the grid search method and cross-validation are used to select the optimal parameter using the LIBSVM library to avoid overfitting problems; in addition, the size of in the insensitive loss function will also affect the performance of the SVR, and the number of support vectors used to construct the regression function is negatively related to ; in multiple matching experiments, the matching algorithm proposed in this paper, the value of does not strongly depend on the initial matching point pairs, and the final value of =0.25 is selected

[0164] ④Outlier rejection, complete image registration

[0165] After obtaining the scalar function from the SVR, the mapping residual is calculated according to the mapping function

[0166] (5)

[0167] The robust weight function assigns different weights to each point pair based on the mapping residual obtained by equation (5) , and the weight in the middle part of the data is relatively large, and then decreases to zero at a certain point on both sides; the Tukey double weight estimation function is as follows ​

[0168] ;

[0169] wherein, =4.4478 × median , and =1,..., , is the residual , the outliers with smaller weights will be removed;

[0170] Meanwhile, the iterative estimation of the mapping function needs to determine the termination condition

[0171] ;

[0172] and respectively represent the size of the variance value before and after removing outliers in one iteration; set the threshold , when , the iterative estimation process is terminated; in addition, the mean square error directly shows the consistency of the corresponding points and the estimated mapping function; therefore, set the threshold , when , the iteration is terminated; and are set to 75 and 0.3, respectively.

[0173] The image registration result, that is, the robust feature point pair extraction result is shown in Figure 5 .

[0174] The sample similar region superposition fusion in the step S5 specifically includes the following contents:

[0175] The mapping function obtained based on the support vector regression algorithm is a high-dimensional space mapping function; for more intuitive description of the actual position mapping relationship between images, a three-dimensional space homography matrix can be established based on the matching point pairs H , and the coordinate transformation relationship can be generally described as:

[0176] ;

[0177] wherein H contains 8 degrees of freedom, and since the image is limited to a plane, the homography matrix is further simplified as:

[0178] ,

[0179] ;

[0180] The matrix needs 4 pairs of matching points to calculate a unique solution by singular value decomposition (SVD), i.e. to obtain 8 unknown necessary parameters of the matrix; the RanSAC algorithm is introduced to iteratively filter out matching points with large errors, and then the matrix is solved;

[0181] The feature point matching method has good robustness to resolution, brightness and camera parameter influence;

[0182] The transformation matrix is used H , so that the two sample images can be fused, and the result is shown in Figure 6 .

[0183] The step S6 of evaluating the difference of the similar regions specifically includes the following contents

[0184] For the evaluation of the similar regions, cross-correlation operation can be used, but the cross-correlation operation can be affected by invalid regions, such as the cross-correlation value of regions with amplitude of 0 is 100%; to overcome this problem, the surface topography difference parameter is used to describe the relative difference of the similar regions; let the similar regions be and , , then

[0185] ;

[0186] wherein, is a surface measurement general parameter: root mean square roughness; the greater the difference between the similar regions and , the greater the value, and vice versa; when and are completely equal, , then =0, i.e. no difference.

[0187] The basic principle, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application, and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A paper surface roughness texture anti-counterfeiting method, characterized in that, Specifically comprising the following steps: S1, planning an effective texture area on the surface of a paper product, marking with a box, a circle frame or a dot, or specifying an approximate position; S2, installing a microscopic lens on the camera of a mobile phone, and using the device to obtain a texture sample image of the area in step S1; S3, extracting a feature point descriptor of the image texture in step S2 by using a SIFT algorithm; S4, estimating a robust mapping function between the feature point descriptors in step S3 by using a support vector regression (SVR) algorithm, extracting a matched feature point pair by using the robust mapping function, and completing image registration; S5, establishing a plane image conversion matrix based on the matched feature point pair obtained in step S4, and realizing sample similar area superposition fusion; S6, evaluating the difference of the similar area by means of a surface topography difference parameter; In steps S1 and S2, a microscopic lens is added to a mobile phone as a portable device, surface roughness texture is used as a feature, and in steps S3-S6, an APP software is used to write an identification algorithm to realize paper product sample collection, qualitative identification and anti-counterfeiting; The step S6 of evaluating the difference of the similar regions specifically includes the following: using a surface topography difference parameter to describe the relative difference of the similar regions; let the similar regions be and , then ； Wherein, General parameters for surface metrology: root mean square roughness; similar areas The greater the difference, The greater the value, the smaller the value; when The greater the difference, The greater the difference, Completely equal, Then =0, that is, no difference.

2. The paper surface roughness texture anti-counterfeit method according to claim 1, characterized in that, In step S3, the SIFT algorithm is used to extract a feature point descriptor, which specifically includes the following contents: wherein represents a convolution operation in x and y directions, , is a scale-variant Gaussian function, is a spatial coordinate, is a scale coordinate;​ ② The original image is continuously down-sampled to obtain a series of images with different sizes by making different scale Gaussian blurring to the sample image. These images are arranged in a tower structure from large to small, from bottom to top. This structure is divided into O groups S layers, that is, a Gaussian pyramid is formed; ③The images in the same group of the Gaussian pyramid are subtracted two by two to obtain a Gaussian difference pyramid; ④Each pixel point in the Gaussian difference pyramid is searched and compared with its surrounding 26 points to determine a feature point; ⑤The curvature at the feature point is calculated by using a Hessian matrix to eliminate unstable edge response points; ⑥For the determined feature point, a feature vector is calculated; ⑦According to the obtained feature vector, a Euclidean distance method is used for initial matching of the feature points.

3. The paper surface roughness texture anti-counterfeit method according to claim 1, characterized in that, In step S4, the robust mapping function between the feature point descriptors is estimated based on a support vector regression algorithm (SVR), which specifically includes the following contents: ①Robust mapping function model According to the homography assumption of the images, two images and should have one-to-one mapping relationship between their corresponding points, which is expressed by a homography function; therefore, there are two homography functions to map a point in one image to the corresponding point in another image ； If satisfy both homographic functions, i.e. are mapped by the function to and vice versa, then are considered as a pair of corresponding points of the image; based on the above homographic function concept, the mapping functions and are considered as two vector-valued functions, respectively ， ， wherein and (a) is a general scalar function; ) is a general scalar function; Assume For initial matching pairs set ; It is in The projection of a subspace is defined as follows ; For any , if satisfies the mapping function , then will satisfy , and similarly, , and are calculated from the corresponding homologous subspaces , and , respectively. ②Support vector regression algorithm (SVR) Consider a dataset where, denotes a D-dimensional input vector, denotes the target output, i.e., a class label, the non-linear relationship between the input and the output is described by the following regression function （1） wherein, represents a prediction function, is a non-linear mapping function, and are two coefficients to be determined and are estimated by minimizing a regular function ， （2） In equation (2), the first term is a regularization term, and the second term describes empirical error, which is given by Robust loss function The estimation is performed with the parameter C being a weighting parameter between the first and second terms in the equation, while the parameter is considered to be equivalent to the distance size of the approximation accuracy in the training data; a relaxation variable and are introduced to introduce the constraint condition expressed in the following equation ； The positive and negative distance sizes are represented by and respectively. Outside the region , the slack variable takes non-zero values, and the ultimate goal of the optimization method is to find a function that has a maximum deviation from the actual function of for all training set data, and is as flat as possible. Formula (1) is further expressed in the following more general form （3） wherein, and is a Lagrange multiplier term, satisfying The data points are support vectors, which are used to construct the decision function; the kernel function maps the input space to a higher dimensional space, where the most commonly used kernel function for SVR is the radial basis kernel function, which plays an important role in both classification and regression prediction. （4） In formula (4), is a kernel parameter; there is only one hyperparameter to be determined in the radial basis kernel function, which makes it have good performance in predicting nonlinear functions in SVR; The support vector regression algorithm is used to estimate the scalar function in and ​ The specific mapping function iteration process is as follows: a) from an initial set of feature point pairs , denoted by , is estimated by the SVR method ; b) calculating a mapping residual for the corresponding point pair based on the current estimate ;​ c) determining a weight value for each corresponding point pair according to the size of the difference between the two points using a robust weight function ; d) updating the weights and re-estimating the mapping function ; Loop iteration steps (b) to (d) until the iteration termination condition is met, and the mapping function and the final feature point pair set are obtained; In predicting the vector-valued function, three parameters , and are determined, whose optimal values depend on the initial matching pairs and the optimization problem itself; the kernel parameter represents the radius of the kernel function, and the regularization parameter is a penalty parameter to balance the flatness of the final solution and the training error; in practice, increasing the value of can minimize the error, but a value too large will lose the generalization performance of the model, i.e., the function estimation accuracy, and increase the complexity of the model; the grid search method and cross-validation are used to select the optimal parameters using the LIBSVM library to avoid overfitting problems; in addition, the size of in the insensitive loss function will also affect the performance of SVR, and the number of support vectors used to construct the regression function is negatively correlated with In multiple matching experiments, the matching algorithm proposed in this paper, the value does not strongly depend on the initial matching pairs, and the final = 0.25 is selected. ④Outlier elimination, completing image registration After obtaining the scalar function from the SVR, the mapping function calculating the residual （5） The mapping residual of robust weight function is obtained by equation (5), which assigns different weights to each point The weight is relatively large in the middle of data, and then decreases to zero at some point on both sides. The Tukey double weight estimation function is as follows ； wherein, = 4.4478 x median , and = 1,..., , i.e. the residual , then the outliers with smaller weights will be removed; At the same time, the iteration estimation of the mapping function needs to determine the termination condition ； and These represent the variance values ​​before and after removing outliers in a single iteration; a threshold is set. ,when When the time is reached, the iterative estimation process terminates; in addition, the mean square error... This visually demonstrates the consistency between the corresponding points and the estimated mapping function; therefore, a threshold is set. When satisfied When the iteration terminates; and Set them to 75 and 0.3 respectively.

4. The paper surface roughness texture anti-counterfeit method according to claim 1, characterized in that, In step S5, the sample similar area superposition fusion is realized, which specifically includes the following contents: The mapping function obtained based on the support vector regression algorithm is a high-dimensional space mapping function; in order to more directly describe the mapping relationship between actual positions of images, a three-dimensional space homography matrix is established based on the matching point pairs H The coordinate transformation relationship is described as: ； wherein H With 8 degrees of freedom, since the image is confined to a plane, the homography matrix is further simplified to: ， ； The matrix needs 4 pairs of matched points to calculate a unique solution by singular value decomposition (SVD), that is, to obtain 8 unknown necessary parameters of the matrix; The RanSAC algorithm is introduced to iteratively filter out the matched point pairs with large errors, and then the matrix is solved; The above feature point pair matching method has good robustness to resolution, brightness and camera parameter influence. Utilizing transformation matrix H i.e. two sample images are fused.

Images