Method for testing and predicting the type of a4 printing paper based on three-dimensional fluorescence spectrum and use

Abstract

The present application belongs to the field of file inspection in judicial expertise, and particularly relates to a method and use for inspecting and predicting the type of A4 printing paper based on three-dimensional fluorescence spectrum. The present application constructs an identification method for analyzing the type of A4 printing paper based on three-dimensional fluorescence spectrum, a method for comparing and inspecting the type of A4 printing paper, and an analysis method for predicting the type of A4 printing paper. The present application solves the problems of insufficient distinguishing accuracy, complicated inspection method, and small coverage of existing A4 paper type identification methods, and realizes the problems of accurate and lossless and quantitative analysis of the type of A4 printing paper.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of file inspection in judicial expertise, and particularly relates to a method for inspecting and predicting the type of A4 printing paper based on three-dimensional fluorescence spectrum and use. BACKGROUND

[0002] The identification of the type of A4 printing paper used for physical evidence in a case has always been the focus of file inspection work, but due to the existence of a large number of A4 printing paper brands and models on the market and the similar production processes of various manufacturers, the accurate positioning of the printing paper brand and model has not been solved. In judicial expertise, the demand for mining relevant information of A4 printing paper as a common carrier of file physical evidence in a case is increasing, but the related analysis method can only determine the approximate type range, or the identification process will damage the file material and affect its legal effect. The root cause of the above problems is that there is no effective identification method for the type of A4 printing paper.

[0003] The fluorescent whitening agent contained in A4 printing paper has the advantages of simple molecular structure, good stability, light resistance, thermal stability and washing resistance, and the papermaking process can achieve uniform distribution of the fluorescent whitening agent on the surface of the paper, so the fluorescence properties of A4 printing paper are very stable.

[0004] The advantage of the present application is that the three-dimensional fluorescence spectrum inspection can achieve non-destructive inspection of the file, which is an efficient and convenient comparison technology, and can accurately identify the brand and model of A4 printing paper. SUMMARY

[0005] Based on the problems of the prior art, the present application provides a method and use for inspecting and predicting the type of A4 printing paper based on three-dimensional fluorescence spectrum, aiming to solve the problems of insufficient distinguishing accuracy, complicated inspection method and small coverage of existing A4 paper type identification methods.

[0006] The present application constructs an identification method for analyzing the type of A4 printing paper based on three-dimensional fluorescence spectrum, a method for comparing and inspecting the type of A4 printing paper, and an analysis method for predicting the type of A4 printing paper. The problem of accurately and non-destructively quantitatively analyzing the type of A4 printing paper is solved.

[0007] The object of the present application is achieved by the following technical solutions:

[0008] The identification method for analyzing the type of A4 printing paper based on three-dimensional fluorescence spectrum specifically comprises the following steps:

[0009] Step one: determine the three-dimensional fluorescence of each paper sample under the same conditions, and convert the obtained three-dimensional fluorescence data into an Excel file in xlsx format to eliminate the first and second Rayleigh scattering of the three-dimensional fluorescence spectrum;

[0010] Step two: intercept the part of the spectrum covering all emission peaks after eliminating Rayleigh scattering in step one, arrange the data into a two-dimensional matrix composed of fluorescence intensity data by brand and model in increasing order of excitation wavelength, standardize the matrix data, perform principal component analysis, select principal components that can contain the main information of the spectrum, and standardize the results;

[0011] Step three: on the basis of the standardization results in step two, use local outlier factor analysis to remove outliers for each category of samples;

[0012] Step four: perform principal component analysis on all samples of each category obtained after screening in step three to reduce the dimension of the data, select principal components that can contain the main information of the spectrum, and standardize the principal component results data;

[0013] Step five: input the results after standardization in step four into the support vector machine algorithm to build a recognition model, and pretreat the test samples in the same way as the model input to obtain the prediction results.

[0014] Further, the three-dimensional fluorescence spectrum of a single A4 paper measured in step one has a detection condition of an excitation wavelength range of 280-420 nm, a step of 10 nm, an emission wavelength range of 300-600 nm, a slit width of 10 nm / 10 nm, and determination of 10 parallel samples; the method for eliminating Rayleigh scattering is the Delaunay triangular interpolation method.

[0015] Further, the range of step two is excitation wavelength 280-420 nm and emission wavelength 370-570 nm; the obtained two-dimensional matrix (n x 1515), n is the number of samples; the principal component analysis results are selected as PC1-PC10, each selected category can cover more than 95% information, and the implementation is realized through Origin software.

[0016] Further, the local outlier factor analysis method used in step three is to judge outliers in the feature space with 7 surrounding points, and to remove outliers with a lof value of 1.2 as the standard, which is realized through Rstudio software.

[0017] Further, the kernel function kernel in the support vector machine algorithm in step five is selected as linear kernel, and the gamma value is selected as 0.1, which is realized through Rstudio software.

[0018] The method for comparing and testing A4 printing paper categories specifically includes the following steps:

[0019] Step one: Collect the three-dimensional fluorescence spectrum data of n (n is greater than 30) different types of A4 printing paper under the same conditions, randomly sample m (m is greater than 15) samples in parallel, and eliminate the first and second Rayleigh scattering of the three-dimensional fluorescence spectrum; the obtained spectrum data is analyzed by principal component analysis according to the type and model, and the results are standardized, and the outliers are removed by local anomaly factor analysis; the samples after screening are made into three-dimensional fluorescence contour maps, the contour map data is enhanced by using the rotation method, and the contour map is standardized;

[0020] Step two: Take the contour map obtained after standardization in step one as the training sample, build a deep learning network model for training, obtain the recognition model, and remove the full connection layer;

[0021] Step three: Measure the three-dimensional fluorescence contour maps of the samples A, B, … under the same conditions, and standardize them according to step one;

[0022] Step four: Use the recognition model obtained in step two to extract the features of the three-dimensional fluorescence contour maps of the samples A, B, …, and get a tensor, respectively denoted as feat1, feat2, …, and calculate the cosine similarity as the standard for comparison and verification.

[0023] Further, the three-dimensional fluorescence of a single A4 paper is measured in step one, and the detection conditions are as follows: excitation wavelength range is 280-420 nm, step is 10 nm, emission wavelength range is 300-600 nm, and slit width is 10 nm / 10 nm; the Rayleigh scattering elimination method is Delaunay triangular interpolation method.

[0024] Further, the principal component analysis results should cover 95-98% of the information, and PC1-PCx (x is greater than 3) is selected, which is realized by Origin software; the local anomaly factor method is used to judge the outliers in the feature space with 7 surrounding points, and the outliers are removed with lof value 1.2 as the standard, which is realized by Rstudio software.

[0025] Further, the contour map in step one is made by MATLAB software contourf function.

[0026] Further, the data enhancement method in step one is to rotate the sample by 90 degrees, 180 degrees and 270 degrees clockwise respectively, that is, one sample gets four pictures.

[0027] Further, the picture standardization processing method in step one uses the transforms module in PyTorch, uses the normalize function to standardize the data, converts from PIL Image format to PyTorch tensor format, and completes the picture preprocessing.

[0028] Further, in step two, a deep learning neural network is built for training, preferably a pre-trained Resnet, Densenet, InceptionNet, vgg, etc. neural network and replace the fully connected layer, epoch is selected to be above 50.

[0029] Further, in step four, the cosine calculation formula is as follows:

[0030]

[0031] Wherein, feat1 and feat2 represent the feature vectors of the sample to be measured, feat1·feat2 represents the inner product of the two vectors;

[0032] ||feat1||2、||feat1||2 represents the vector two norm.

[0033] Further, in step four, the cosine similarity is used as the standard for evaluating whether it is the same kind of paper, which is 0.9.

[0034] The analysis method for predicting the type of A4 printing paper specifically comprises the following steps:

[0035] Step one: Collecting three-dimensional fluorescence spectrum data of n (n greater than 50) different types of A4 printing paper under the same conditions, randomly taking m (m greater than 15) samples for parallel measurement, and eliminating the first and second Rayleigh scattering of the three-dimensional fluorescence spectrum; The obtained spectrum data is analyzed by principal component analysis according to brand and model, and the results are standardized, and the outlier samples are removed by local anomaly factor analysis; After screening, the sample is made into a three-dimensional fluorescence contour map, the contour map data is enhanced by rotating, and the contour map is standardized;

[0036] Step two: Taking the standardized picture obtained in step one as the training sample, building a deep learning network model for training, and saving the obtained model;

[0037] Step three: Under the same conditions, measure the three-dimensional fluorescence contour map of the sample paper to be measured, and standardize it according to the method in step one;

[0038] Step four: Input the sample picture obtained by standardizing in step three into the saved model in step two, and get the model prediction result.

[0039] Further, in step one, the measured three-dimensional fluorescence of a single A4 paper is detected under the condition that the excitation wavelength range is 280-420 nm, the step is 10 nm, the emission wavelength range is 300-600 nm, and the slit width is 10 nm / 10 nm; The method for eliminating Rayleigh scattering is Delaunay triangular interpolation.

[0040] Further, the principal component analysis result should cover 95-98% of the information, PC1-PCx (x is greater than 3) are selected, and the Origin software is used to realize; the local anomaly factor method is used to judge the outlier point in the feature space with the surrounding 7 points, and the outlier sample is removed with the lof value 1.2 as the standard, and the Rstudio software is used to realize.

[0041] Further, the contour plot in step one is made by the contourf function of MATLAB software.

[0042] Further, the data enhancement method in step one is to rotate 90 degrees, 180 degrees and 270 degrees clockwise respectively, that is, one sample gets four pictures respectively.

[0043] Further, the picture standardization processing method in step one uses the transforms module in PyTorch, uses the normalize function to standardize the data, converts from the PIL Image format to the PyTorch tensor format, and completes the picture preprocessing.

[0044] Further, in step two, a deep learning neural network is built for training, Resnet, Densenet, InceptionNet, Vgg and other neural networks are selected, the training set and test set ratio is 9:1, and the model with training accuracy of 90% and above is selected.

[0045] Further, in step four, the specific probability of the prediction result should be given.

[0046] Compared with the prior art, the beneficial effects of the present application are.

[0047] A fast, accurate and non-destructive A4 printing paper type identification method is provided. First, the three-dimensional fluorescence spectrum of the fluorescent whitening agent contained in the A4 paper is taken as the distinguishing point, which is universal for different types of A4 printing paper. Second, compared with the traditional data feature extraction method, the method of extracting three-dimensional fluorescence spectrum features by using the pre-trained model parameters of different brands of A4 printing paper is more targeted and has high discrimination. In addition, the A4 paper brand comparison test method proposed by the present application judges whether the suspicious paper and the normal paper are the same type of paper by calculating the cosine similarity, and has a wide application range. Finally, the present application combines the neural network model to predict and analyze the unknown paper, which can realize non-destructive, quantitative and probabilistic analysis of paper types. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 Example one SVM model training result;

[0049] Figure 2 Resulting eigenvectors and cosine similarity of three pages of paper in Example Two;

[0050] Figure 3 Three-dimensional fluorescence contour plot. DETAILED DESCRIPTION

[0051] The technical solutions of the present application will be described below in connection with the embodiments, which are only some of the embodiments of the present application, but not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] Example One.

[0053] The identification method based on three-dimensional fluorescence spectrum analysis is used to identify the types of A4 printing paper.

[0054] Five types of A4 printing paper are known, and the type information is shown in Table 1. The type identification is performed by using the present application, and the specific implementation steps are as follows:

[0055] Step 1: Randomly take 10 pieces of each of the five types of A4 paper, and measure 50 pieces of three-dimensional fluorescence data under the condition that the excitation wavelength range is 280-420 nm, the step length is 10 nm, the emission wavelength range is 300-600 nm, and the slit width is 10 nm / 10 nm; take the file to be detected, and measure its three-dimensional fluorescence data under the same condition; and eliminate the first and second Rayleigh scattering of the three-dimensional fluorescence spectrum of a total of 51 samples by using the Delaunay triangular interpolation method.

[0056] Step 2: Intercept the data of the excitation wavelength 280-420 nm and the emission wavelength 370-570 nm of the 51 samples in Step 1, arrange 10 samples of each type in increasing order of excitation wavelength, obtain a two-dimensional matrix (10x1515) composed of fluorescence intensity data, and then normalize the two-dimensional matrix data and perform principal component analysis, select PC1-PC10, and normalize the analysis results.

[0057] Step 3: Input the standardized PC1-PC10 in Step 2 into the lof algorithm to calculate the lof values of 10 samples of each type as shown in Table 2, and discard 3 outlying samples with lof values greater than 1.2.

[0058] Step 4: Re-measure the 3 outlying samples as supplements, and perform principal component analysis on all the type samples obtained after Step 3 to realize data dimension reduction, select PC1-PC10, and normalize the data.

[0059] Step 5: Use the 50 samples obtained in Step 4 to build an svm support vector machine model through Rstudio software, the kernel function kernel is preferably linear kernel, the gamma value is selected as 0.1, and the model training result is as shown in Figure 1 The to-be-tested samples PC1-PC10 are input into the model, and are finally identified as the Asia-Pacific Sunbo brand bright series A4 printing paper.

[0060] Table 1 A4 paper brand information table used in Example 1.

[0061] Brand Longcheng Deli Baowang Yataisenbo Tianzhang Paper Industry Model Copy paper Coral sea Color paper Bright Lehuo Tianzhang

[0062] Table 2 lof value of each brand in Example 1.

[0063]

[0064] Example 2.

[0065] The method for comparing and testing A4 printing paper types is adopted.

[0066] In March 2023, a suspicious contract composed of three A4 printing papers was received, and the police suspected that the middle page (the second page) was replaced. The identification requirement is to determine the type of A4 printing paper in the suspicious contract.

[0067] The specific implementation operation steps are as follows:

[0068] Step 1: Take 30 samples from the sample library of China Criminal Police College, extract 450 three-dimensional fluorescence contour maps, and rotate the 450 pictures clockwise by 90 degrees, 180 degrees and 270 degrees respectively. After data enhancement, a total of 1800 pictures are obtained; use the transforms module in PyTorch to standardize the data, use the normalize function, the parameters are mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], adjust the image size to 224x224, and convert the image from PIL Image format to PyTorch tensor format. Complete the standardization processing of the picture.

[0069] Step 2: Take the 1800 pictures obtained after standardization in Step 1 as training samples, select Resnet50 to build a deep learning network model for training; divide the training set and the test set according to the ratio of 9:1, the batch_size is 64, the lr is set to 0.001, and the total number of epochs is set to 50. The maximum accuracy of the model is 0.919 in the 46th round, and the model is saved. Save the obtained model and remove the fully connected layer;

[0070] Step 3: Measure the three-dimensional fluorescence contour maps of three test sample A4 papers under the same conditions respectively, and standardize them according to the method in Step 1.

[0071] Step 4: Use the model parameters obtained in Step 2 to extract the features of the three-dimensional fluorescence contour maps of the three test sample A4 papers, and obtain three tensors, denoted as feat1, feat2, and feat3 respectively. Calculate the cosine similarity between feat2 and feat1, feat3 respectively according to the cosine calculation formula, which is the probability of the same kind of paper. The results are shown in Figure 2 , where the cosine similarity of feat1 and feat3 is 0.933; the cosine similarity of feat1 and feat2 is 0.563; and the cosine similarity of feat2 and feat3 is 0.725. It can be judged that the first and third A4 printed papers are of the same kind, and the paper used in the second page is replaced and does not belong to the same kind.

[0072] Example Three.

[0073] An analysis method for predicting the type of A4 printed paper is adopted.

[0074] In May 2023, A4 printed paper samples were found in a “counterfeit” case. The police wanted to narrow down the scope of investigation by identifying the type of paper. The specific implementation steps are as follows: (here is compared with the sample library of China Criminal Police College to determine the matching degree of the type of paper).

[0075] Step 1: Take 750 three-dimensional fluorescence contour maps of 50 types of paper from the sample library of China Criminal Police College (one contour map is shown in Figure 3 ), rotate each of the 750 pictures by 90 degrees, 180 degrees, and 270 degrees clockwise, and obtain a total of 3000 pictures after data augmentation; use the transforms module in PyTorch to standardize the data, use the normalize function with parameters mean = [0.5, 0.5, 0.5] and std = [0.5, 0.5, 0.5], adjust the image size to 224x224, and convert the image from PIL Image format to PyTorch tensor format to complete the image standardization.

[0076] Step 2: Use the 3000 standardized pictures obtained in Step 1 as training samples, and select Resnet101 to build a deep learning network model for training; divide the training set and test set in the ratio of 9:1, set the batch_size to 64, set the lr to 0.001, and set the total number of epochs to 100. The model accuracy reaches a maximum value of 0.926 at the 83rd round, and the model is saved.

[0077] Step 3: The three-dimensional fluorescence contour map of the test sample paper is measured under the same conditions, and is normalized in the manner of step 1;

[0078] Step 4: The test sample picture obtained after the pretreatment of step 3 is input into the traceability model based on the Resnet101 model trained by 3000 pictures of 50 categories in the sample library, to obtain the top three categories with matching probabilities and corresponding probabilities. The model prediction result is shown in Table 3, wherein the corresponding category of Class18 in the sample library is Haizhieu, and the matching rate reaches 94.09%, and it can be confirmed that the grammage is 70 g / m 3 .

[0079] Table 3: Matching results of the test sample and the sample library in Example 3.

[0080] Top three categories with the highest probability Matching rate Class 18 94.09％ Class 2 5.79％ Class 2 0.04％

Claims

1. A method for identifying the type of A4 printing paper based on three-dimensional fluorescence spectrum analysis, characterized in that, Specifically comprising the following steps: Step one: measure the three-dimensional fluorescence of each paper sample under the same conditions, convert the obtained three-dimensional fluorescence data into an Excel file in xlsx format, measure 10 parallel samples; eliminate the first and second Rayleigh scattering of the three-dimensional fluorescence spectrum; Step two: intercept the part of the spectrum after eliminating Rayleigh scattering in step one, arrange the data into a two-dimensional matrix composed of fluorescence intensity data in increasing order of excitation wavelength for each brand and model, standardize the matrix data, perform principal component analysis, select the principal components containing the main spectral information, and standardize the results; Step three: on the basis of the standardized results in step two, use local outlier factor analysis to remove outliers from each category of samples; Step four: perform principal component analysis on all samples of each category obtained after screening in step three to reduce the dimension of the data, select the principal components containing the main spectral information, and standardize the principal component result data; Step five: input the results after standardization in step four into the support vector machine algorithm to build a recognition model, and pre-process the test samples in the same way as the model input to obtain the prediction results.

2. The method of identifying the kind of A4 printing paper based on three-dimensional fluorescence spectrum analysis according to claim 1, characterized in that, The range of interception in step two is excitation wavelength 280-420 nm and emission wavelength 370-570 nm; the obtained two-dimensional matrix (n x 1515), n is the number of samples; the principal component analysis results are PC1-PC10, which are realized by Origin software; in the support vector machine algorithm in step five, the kernel function kernel is selected as linear kernel, and the gamma value is selected as 0.1, which is realized by Rstudio software.

3. Method for the comparison of A4 printing paper types for the verification, characterized in that, Specifically comprising the following steps: Step one: collect three-dimensional fluorescence spectrum data of n (n greater than 30) different types of A4 printing paper under the same conditions, randomly take m (m greater than 15) samples for parallel measurement, and eliminate the first and second Rayleigh scattering of the three-dimensional fluorescence spectrum; Perform principal component analysis on the obtained spectral data by type and model, and the principal component analysis results cover 95-98% of the information, take PC1-PCx (x greater than 3), realize by Origin software, and standardize the results, and use local outlier factor analysis to remove outliers; after screening, make a three-dimensional fluorescence contour map, enhance the contour map data by rotating, and standardize the contour map; Step two: use the standardized contour map obtained in step one as a training sample to build a deep learning network model for training, obtain a recognition model, and remove the fully connected layer; Step three: measure the three-dimensional fluorescence contour maps of test samples A, B, and so on under the same conditions, and standardize them according to step one; Step four: use the recognition model obtained in step two to extract the features of the three-dimensional fluorescence contour maps of test samples A, B, and so on, and obtain tensors feta1, feat2, and so on, respectively, and calculate the cosine similarity to represent the standard for comparison and verification.

4. The method for comparing the test A4 printing paper species according to claim 3, characterized in that, The step two builds a deep learning neural network for training, selects a pre-trained Resnet, Densenet, InceptionNet, and vgg neural network and replaces the fully connected layer, and the epoch is selected to be more than 50.

5. The method for comparing the test A4 printing paper species according to claim 3, characterized in that, The cosine calculation formula in step four is as follows: Wherein, feat1 and feat2 respectively represent the feature vectors of the sample to be measured, feat1·feat2 represents the inner product of two vectors; ‖feat1‖2, ‖feat1‖2 represents the vector two norm; The cosine similarity in step four uses 0.9 as the standard for evaluating whether it is the same kind of paper.

6. An analysis method for predicting a kind of A4 printing paper, characterized by, Specifically includes the following steps: Step one: under the same conditions, collect the three-dimensional fluorescence spectrum data of n (n is greater than 50) different kinds of A4 printed paper, randomly take m (m is greater than 15) samples for parallel measurement, and eliminate the first and second Rayleigh scattering of three-dimensional fluorescence spectrum; The obtained spectrum data is analyzed by principal component analysis according to brand and model, the principal component analysis result should cover 95-98% information, PC1-PCx (x is greater than 3) is selected, realized by Origin software, and the result is standardized, and the local anomaly factor analysis method is used to eliminate outlier samples; After screening, the sample is made into a three-dimensional fluorescence contour map, the contour map data is enhanced by using rotation, and the contour map is standardized; Step two: take the standardized picture obtained in step one as the training sample, build a deep learning network model for training, and save the obtained model; Step three: measure the three-dimensional fluorescence contour map of the sample paper under the same conditions, and standardize the processing according to step one; Step four: input the sample picture obtained by standardizing processing in step three into the saved model in step two, and get the model prediction result, which is a specific probability.

7. The analysis method for predicting a kind of A4 printing paper according to claim 6, characterized by, In step two, a deep learning neural network is built for training, and Resnet, Densenet, InceptionNet and Vgg neural network are selected, the training set and test set ratio is 9:1, and the model with training accuracy of 90% and above is selected.

8. The method according to any one of claims 1 to 7, characterized in that, The three-dimensional fluorescence of single A4 paper is measured in step one, the detection conditions are excitation wavelength range of 280-420nm, step length of 10nm, emission wavelength range of 300-600nm, and slit width of 10nm / 10nm; The method for eliminating Rayleigh scattering is Delaunay triangular interpolation method.

9. The method according to any one of claims 1 to 7, characterized in that, The local anomaly factor method is used to judge the outlier in the feature space with 7 surrounding points, and the outlier sample is removed with lof value of 1.2 as the standard, which is realized by Rstudio software.

10. The method according to any one of claims 3 to 7, characterized in that, The step one is made by the contourf function of MATLAB software; the data enhancement mode in the step one is respectively rotating 90 degrees, 180 degrees and 270 degrees clockwise, that is, one sample respectively obtains four pictures; the picture standardization processing method in the step one uses the transforms module in PyTorch, uses the normalize function to standardize the data, converts from the PIL Image format to the PyTorch tensor format, and completes the picture preprocessing.

Images