CN102289665A - Printed file identifying method based on powdered ink stack texture analysis - Google Patents

Abstract

The invention provides a printed file identifying method based on powdered ink stack texture analysis. At first, printed character stroke images on printed files are respectively collected through a high-precision microscopic amplifying system; small powdered ink stack blocks are extracted from the printed stroke images so that a normalized powdered ink texture picture is formed; the texture features of the printed files are extracted in a way that the texture features of the normalized texture image are extracted through a local binary pattern, a gray level concurrence matrix, a Fourier spectrum, a Gabor technology and a Wavelet technology; and the text features of the extracted printed files are classified into different patterns through a principle component analysis and a linear discriminant analysis, thus a user can judge whether the two printed files are printed and generated through the same printer or not. Through the printed file distinguishing method, the automatic printed file identification which is based on computer image analysis and has independent texts is realized and the accuracy can reach above 89 percent.

Description

Pile up the print file discrimination method of texture analysis based on ink powder

Technical field

The present invention relates to the true and false of print file is carried out authentication method.Relate in particular to a kind of print file discrimination method and belong to public safety, the examination of material evidence field based on ink powder accumulation texture analysis.

Background technology

One passes through the expert of public safety department to detect the printed material composition, searches the printing flaw, and classic methods such as extraction paper feed impression are compared to print file and collected evidence.Along with the development of computer science, computer based print file forensic technologies becomes new research focus again, comprises the size of analyzing printable character, prints print quality features such as edge roughness; Analyze the gray level co-occurrence matrixes feature of printable character image; Extraction comprises the watermark feature of type information etc.

Though traditional prints file evidence collecting method evidence is more abundant, specific aim is too strong, and range of application is narrow, can't differentiate most common print file.As chemistry and optical means, can't distinguish the identical print file of printed material composition at ink powder or ink composition check; Print the watermark comparative approach only at the high-grade printer prints file that is partially submerged into watermark, most of common print file do not have embed watermark.Existing computer approach is analyzed the size and the details print quality feature of printable character, and the feature that these existing forensic technologies are extracted is mostly simpler, and dimension is lower, and its accuracy rate, the scope of application and practicality also can't reach the standard that supports judicial foundation.Generally speaking, the research of computer printout file evidence obtaining is in the starting stage, and achievement is limited, and accuracy is also far away from practicality.

The identical characters of different printer prints does not almost have difference under naked eyes, be difficult to distinguish, but under multiplying arrangement is observed, manifested character shape, and the character edge roughness is printed multiple differences such as ink powder bulk density.The present invention by relatively printing ink powder bulk density feature, prints file judicial expertise and evidence obtaining exactly.

Summary of the invention

The present invention is an object with modal laser printing file in the public safety case, with computer vision and mode identification technology, provides a kind of print file discrimination method of piling up texture analysis based on ink powder.The present invention is judging that whether print file are the check discriminating problem of identical print file, solve as the classification problem in the pattern-recognition.

A kind of print file discrimination method of piling up texture analysis based on ink powder of the present invention, comprise the print file image acquisition, print file pre-service structure ink powder is piled up texture maps, texture maps feature extraction and eigentransformation, print document clustering and classification by feature at last, identification flow block diagram such as Fig. 2.

It specifically is the printable character stroke pattern picture that utilizes earlier on the micro-amplification system collection of the high precision print file; From print the stroke pattern picture, extract ink powder accumulation fritter then and construct normalization ink powder texture maps; The normalization ink powder is piled up the texture maps texture feature extraction; Whether utilize principal component analysis (PCA) and linear discriminant analysis to carry out pattern classification to the textural characteristics that extracts, be that same printer is printed generation to judge two parts of print file.

Of the present invention based in the ink powder accumulation texture analysis discriminating print file method, printable character stroke pattern picture on the print file that collect will extract normalization through pre-service and print ink powder accumulation texture image, pre-treatment step comprises: earlier by the stable ink powder overlay area at image segmentation positioning printing stroke center, extract 16 * 16 little ink powder again and pile up picture to be spliced into size be 256 * 256 normalization texture image from these zones; The normalization texture image that constructs is carried out the local gray level equilibrium, realize the background equilibrium of texture image; Obtain the gray scale texture image of normalized 256 * 256 pixels at last.

The little ink powder of the present invention in the image pre-service piled up the normalization texture image size of picture and splicing and all can be adjusted.

In the print file discrimination method based on ink powder accumulation texture analysis of the present invention, the extraction of print file textural characteristics is adopted local binary pattern, gray level co-occurrence matrixes, Fourier spectrum, GABOR and WAVELET technology are extracted the textural characteristics of normalization texture image.

In the method for the present invention, svm classifier device, ADABOOST sorter, and many features and multiple Classifiers Combination method are adopted in described pattern classification.

Print file discrimination method based on ink powder accumulation texture analysis of the present invention is applicable to the discriminating of common black and white laser printing file, color laser printing file and laser photography file.

Print file image acquisition of the present invention is used print character image magnified sweep system architecture such as Fig. 3.

The document image resolution that existing scanner scanning arrives does not reach the observation requirement end of than.Many minutias that really can reflect the printer identity only just can manifest under the observation of high power zoom microscope, therefore select the high magnified glass head to add ccd video camera and gather the print file image; The image enlargement factor will satisfy and shows details (40～60 times) such as ink powder accumulation form on the print file clearly; Adopt the illumination of LED lamp that sufficient light uniformly is provided.

Image pre-service among the present invention, by the background area in the removal print file image and the borderline region of printable character, the stable ink powder overlay image that extracts printable character stroke central area is constructed the texture picture of unified size.Earlier by the stable ink powder overlay area at image segmentation positioning printing stroke center, from these zones, extract 16 * 16 little ink powder again and pile up picture to be spliced into size be 256 * 256 normalization texture image; The normalization texture image that constructs is carried out the local gray level equilibrium, thereby reduced the influence of inhomogeneous illumination, realize the background equilibrium of texture image.

Fig. 4 .a-d has provided the ink powder of print file and has piled up the texture leaching process.Fig. 4 .d and Fig. 4 .f have provided the normalization texture that extracts in two different printers, as can be seen, it is obviously different that the ink powder of two printers is piled up texture: it is comparatively fine and closely woven that the printing ink powder among Fig. 4 .d is piled up texture, and the accumulation of the ink powder among Fig. 4 .f texture is comparatively coarse.

During structure normalization texture maps, the selection of sub-pictures size can influence the correct acquisition of textural characteristics, one big textures windows helps the calculating of textural characteristics, because it has comprised more pixel, can obtain textural characteristics more accurately, but the width limitations of printable character stroke the maximum possible width of textures windows; Can't catch the textural characteristics of the uniqueness of silhouette target if window is too little.So sub-pictures size (4 * 4,8 * 8,16 * 16) will be chosen according to experimental result.The present invention adopts 16 * 16th, the preferred dimensions that obtains through overtesting.

The method of texture feature extraction is a lot, the local binary pattern (LBP) of Texture classification better performances, gray level co-occurrence matrixes (GLCM), Fourier spectrum (FT) are adopted in feature extraction of the present invention and selection, GABOR, technology such as WAVELET are extracted the textural characteristics of normalization texture image.Can select after the feature extraction to improve classification accuracy with eigentransformation.

The ink powder that the present invention is based on principal component analysis (PCA) (PCA) and linear discriminant analysis (LDA) is piled up the texture discriminating

The ink powder of structure print file is piled up texture maps, behind the texture feature extraction, utilizes PCA and LDA method to carry out feature selecting.Pile up the specific algorithm that texture merges the identification of PCA+LDA tagsort based on ink powder, performing step is as follows:

1, N part print file of given different printer generations through image acquisition, pre-service and feature extraction, obtain the training set f that ink powder is piled up textural characteristics _i, i=1,2 ..., N calculates earlier covariance matrix, asks its m (individual eigenvalue of maximum characteristic of correspondence vector w of m≤N-1) then ₁ ^Pca..., This m proper vector constitutes the PCA projection matrix $W_{\mathrm{pca}}=[w_1^{\mathrm{pca}},\·\·\·,w_i^{\mathrm{pca}},\·\·\·,w_m^{\mathrm{pca}}].$

2, utilize PCA projection matrix W _PcaWith training set n-dimensional vector space conversion is the m dimension MEF space and the best feature MEFs that describes of acquisition of dimensionality reduction.Be y _i=(y ₁ ⁱ, y ₂ ⁱ..., y _m ⁱ) ^T=W _Pca ^T(x _i-m _i-m ₀), i=1 wherein, 2 ..., N.

3, calculate by the best feature y that describes of training set ₁..., y _i..., y _NScatter matrix S in the class that constitutes _wAnd scatter matrix S between class _b, calculate corresponding matrix then

K eigenvalue of maximum characteristic of correspondence vector w ₁ ^F1d...,

Constitute linear (LDA) projection matrix of differentiating of Fisher by this k eigenvalue of maximum characteristic of correspondence vector $W_{\mathrm{fld}}=[w_1^{\mathrm{fld}},\·\·\·,$ $w_i^{\mathrm{fld}},\·\·\·,w_k^{\mathrm{fld}}].$

4, utilize LDA projection matrix W _F1dWith the MEF space conversion is the k dimension MDF space of dimensionality reduction, obtains corresponding optimal classification feature MDFs.Be z _i=(z ₁ ⁱ, z ₂ ⁱ..., z _k ⁱ) ^T=W _F1d ^Ty _i, i=1 wherein, 2 ..., N.Optimal classification proper vector z according to these corresponding texture image training sets ₁..., z _i..., z _NForm c class print file ink powder textural characteristics database.

5, ink powder that certain is to be identified is piled up textural characteristics x and is carried out the best that the PCA projective transformation obtains in the m dimension MEF space and describe feature, y=(y ₁, y ₂..., y _m) ^T=W _Pca ^T(x-m-m ₀).

6, the best of m dimension being described proper vector y carries out the LDA projective transformation and obtains optimal classification feature z=(z in the k dimension MDF space ₁, z ₂..., z _k) ^T=W _F1d ^TY.

7, utilize the Euclidean distance sorter, calculate the characteristic matching distance that ink powder to be identified is piled up the institute's ink powder accumulation texture classes in texture and the training set (property data base), and draw the discriminating conclusion with threshold ratio.When two parts of print file characteristic matching distance values less than threshold value, then be judged to same printer and print generate; If the characteristic matching distance value, then is judged to different printer prints greater than threshold value and generates.

For avoiding the existing digit evidence collecting method to rely on the narrower situation of content of text applicable surface strongly, employing of the present invention and print text content are irrelevant, and the printing ink powder that extensively is present in the print file of various quality is piled up textural characteristics.This category feature extensively is present in the text printing file of various quality, has therefore also avoided the problem that is difficult to extract as printing flaw feature occurs at random.Be present in all print file owing to print the ink powder accumulation, and irrelevant with the print text content, therefore method of the present invention has text independence and advantage of wide range of application.The present invention has realized that based on computer image analysis text robotization print file are independently differentiated, more than the rate of accuracy reached to 89%.Can use for the judicial expertise and the evidence obtaining of print file.

Description of drawings

Fig. 1 printer prints character difference

(a) different printer details are printed ink powder accumulation difference (b), and the identical characters profile of 4 different printer prints relatively.

Fig. 2 piles up the print file discrimination process of texture analysis based on ink powder

Fig. 3 print character image magnified sweep system architecture

Among the figure: 1-objective table, 2-two dimension motorized precision translation stage (x direction, y direction), 3-bracing frame, 4-up-down adjustment screw rod, 5-CCD video camera, the varifocal enlarging lens of 6-, 7-LED light source, 8-paper lay down location.

Fig. 4 normalization texture maps manufacturing process

Among the figure: a-prints the stroke micro-image; B-extracts 16 * 16 ink powder covering fritter from image; C-extracts the normalization texture maps that ink powder covers little block structure from Fig. 4 a; Effect behind d-Fig. 4 c gray balance; The normalization texture maps of another printer of e-; Effect behind f-Fig. 4 e gray balance.

The print file identification experiment result of two kinds of texture methods of Fig. 5

Embodiment

Embodiment 1

Adopt 40 printers to test the identification result of piling up texture analysis based on ink powder.To every printer, gather two printable character pictures on the different files respectively, the character picture on file is as training sample (No. four black matrix Chinese characters), and the character picture on another file (No. three black matrix Chinese characters) is used to test identification result.Gather 25 of printable character images on every print file, can construct the normalization texture image about 15 width of cloth.

From the normalization texture image, after the texture feature extraction, utilize the PCA+LDA method to carry out feature selecting, obtain the proper vector behind the dimensionality reduction, adopt the weighted euclidean distance sorter to differentiate.The proper vector of unknown print file ink powder texture is compared with the print file of the known sample that has trained, the weighted euclidean distance WED of its proper vector that and if only and if k class sample hour, the input print file are classified as k platform printer and generate.Weighted euclidean distance calculates by following formula:

$d=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\frac{{(f_i-f_i^{\left(k\right)})}^2}{{\left({\mathrm{\δ}}_i^{\left(k\right)}\right)}^2}---\left(1\right)$

F wherein _iI feature of expression unknown sample, Average and the variance of representing i feature of k class sample respectively, N represents the feature sum that each sample extracts.

Overall accuracy and overall error rate to gray level co-occurrence matrixes (GLCM) and two kinds of characterization method print file checks of fourier spectrum (FT) are added up, and the discrimination of two kinds of methods as shown in Figure 5.Horizontal ordinate is the number of proper vector, and ordinate is a discrimination.Because two kinds of methods just can reach the maximal value of checking accuracy when 13 left and right sides discriminant vectorses, and along with the increase of discriminant vectors number, its accuracy tends towards stability, so the discrimination of preceding 20 discriminant vectors correspondences that only drawn among the figure.The best experimental result of two kinds of methods sees Table 1.

The present invention can also adopt other texture analysis characterization method and eigentransformation in special type is extracted, also can adopt the sorter of other types at minute time-like, as SVM, and ADABOOST, and various many features, multiple Classifiers Combination method.In a word, as long as the method for having used stamping ink powder pattern reason to analyze just belongs to protection scope of the present invention.

In addition, the present invention not only differentiates at common black and white laser printing file, also can be applied to color laser printing, laser photography file, and the printing of other types, the judicial inspection of printed text and discriminating.Therefore, as long as used the inspection of document, the discriminating of above-mentioned ink powder texture analysis method, just belong to protection scope of the present invention.

The experimental result of two kinds of texture methods of table 1

Method	Accuracy (%)	Error rate (%)
			GLCM	89.23	10.77
FT	88.77	11.23

Claims (6)

1. the print file discrimination method based on ink powder accumulation texture analysis is characterized in that: at first, utilize the printable character stroke pattern picture on the micro-amplification system collection of the high precision print file; From print the stroke pattern picture, extract ink powder accumulation fritter then and construct normalization ink powder texture maps; The normalization ink powder is piled up the texture maps texture feature extraction; Whether utilize principal component analysis (PCA) and linear discriminant analysis to carry out pattern classification to the textural characteristics that extracts, be that same printer is printed generation to judge two parts of print file.

2. the print file discrimination method of piling up texture analysis based on ink powder according to claim 1, it is characterized in that: the printable character stroke pattern picture on the print file that collect extracts normalization through pre-service and prints ink powder accumulation texture image, pre-treatment step comprises: earlier by the stable ink powder overlay area at image segmentation positioning printing stroke center, extract 16 * 16 little ink powder again and pile up picture to be spliced into size be 256 * 256 normalization texture image from these zones; The normalization texture image that constructs is carried out the local gray level equilibrium, realize the background equilibrium of texture image; Obtain the gray scale texture image of normalized 256 * 256 pixels at last.

3. the print file discrimination method based on ink powder accumulation texture analysis according to claim 2 is characterized in that: the little ink powder in the image pre-service is piled up the normalization texture image size of picture and splicing and all can be adjusted.

4. the print file discrimination method of piling up texture analysis based on ink powder according to claim 1, it is characterized in that: the extraction of print file textural characteristics, adopt local binary pattern, gray level co-occurrence matrixes, Fourier spectrum, GABOR and WAVELET technology are extracted the textural characteristics of normalization texture image.

5. the print file discrimination method based on ink powder accumulation texture analysis according to claim 1, it is characterized in that: svm classifier device, ADABOOST sorter, and many features and multiple Classifiers Combination method are adopted in described pattern classification.

6. the described application of piling up the print file discrimination method of texture analysis based on ink powder of claim 1, it is characterized in that: this method is applicable to the discriminating of common black and white laser printing file, color laser printing file and laser photography file.

Images