CN104732548A - Print file identification method based on texture synthesis - Google Patents

Abstract

The invention provides a print file identification method based on texture synthesis in order to solve the problem of identification under the condition that same characters between print files lack even no same character exists between the print files. Each character image of a print file is acquired by utilizing an overall high magnification method, an incomplete character image of the print file is converted into a complete texture image through image inpainting, interferences of different character morphological structures are eliminated accordingly, and therefore feature fusion can be performed on all the images synthesized by one printer, and feature description of the printer is more accurate. Four statistical features of a gray-level co-occurrence matrix of the synthesized images are calculated, namely, the contrast degree, consistency, self-correlation and entropy, a feature mean value of all the images serves as feature description of one printer, and feature distance between two printers serves as a basis for matching classification. By the adoption of the print file identification method based on texture synthesis, same characters do not need to exist between the files of the printers, the application range of the print file identification method is widened and accuracy is improved.

Description

A kind of print file discrimination method based on textures synthesis

Technical field

Technical field assert by the source machine that the present invention relates to printer document, and propose a kind of print file discrimination method based on textures synthesis, object is the discriminating one to one in order to realize two parts of printed documents.

Background technology

Because the parameter configuration of every platform printer when producing and the loss situation in using are all unique, therefore the document printed is not identical yet, this is also the objective base that print file can distinguish qualification, and so-called print file differentiate to be exactly that qualification two parts of printed documents are whether from the technology of same printer.

Universal along with computing machine and printer, print file also replace the main media that manual documentation becomes beared information gradually, meanwhile, the criminal offences such as relevant print file are distorted, forgery are also increasing, and therefore increasing people drops into the Related Research Domain that print file are differentiated.It is all have when same word at two parts of printed documents just can accomplish that current print file are differentiated mainly, because when kinds of characters does coupling discriminating, the impact of the impact meeting override printer feature of its morphosis, cannot obtain correct conclusion; Secondly, in order to the accuracy obtained, identical characters multiplicity can not very little, because it is more stable to repeat more printer characteristics that it shows.And above said these 2 are difficult to be met in actual applications, especially more difficultly in Chinese printed document is differentiated to accomplish, therefore the present invention proposes a kind of print file discrimination method based on textures synthesis, to solve the problem cannot carrying out when not having identical characters to mate or identical characters repeats little differentiating.

Summary of the invention

For the print file solved in the rare situation of printed document identical characters differentiate problem, the present invention utilizes the method for textures synthesis, by the blank parts reparation in the single character picture of printed document, obtain complete texture image thus eliminate the morphosis impact between kinds of characters.The texture information of file and picture carries the characteristic of its printer, and whether two parts of printed documents are from same printer to utilize textural characteristics to differentiate.

The present invention adopts following technical scheme:

Based on a print file discrimination method for textures synthesis, specifically comprise the following steps:

(1) integral image magnification at high multiple scanning system is utilized to gather the single character picture of printed document;

(2) by texture synthesis method, the white space in the single character picture of incompleteness is repaired as complete texture image;

(3) calculate the gray level co-occurrence matrixes of the complete texture image after repairing, and extract contrast, consistance, auto-correlation and entropy four statistical natures;

(4) normalized and Fusion Features are done to carried feature, calculate the diagnostic characteristics of printed document, and using the characteristic mean of all images of a printed document as the feature of affiliated printer;

(5) classified by sorter, differentiate that whether two parts of printed documents are from same printer.

In described step (5), calculate the characteristic distance of two parts of printed documents, when distance is less than empirical value, judge that two parts of documents are from same printer.

Described empirical value adopts the method for Bayesian Estimation to calculate: the printed document sample first taking 100 printers, calculate the range distribution between same printer document and the range distribution between different printer document respectively, minimum for criterion with Type Ⅰ Ⅱ error rate sum, find empirical value T.

Described texture synthesis method comprise strain-based design Method and Process restorative procedure and all can be the method for complete texture by incomplete texture repair.

Described texture synthesis method adopts the texture repairing method based on sample block, first will mark the region to be repaired of image:

First, utilize Otsu algorithm to do binary conversion treatment to original image, tentatively distinguish texture region and background area;

Secondly, remove the noise in background area by medium filtering, but medium filtering cannot remove the noise of texture region marginal portion than comparatively dense, so need to operate by dilation erosion the pure property ensureing texture region;

Finally, the background area after dilation erosion is modified as pure white in the corresponding position of original image, and the white portion obtained is region to be repaired.

Described texture synthesis method utilizes in source region and fills region to be filled to be filled piece of match block the most similar, if region to be repaired is Ω, and the edge in known region to be Φ, δ Ω be region to be repaired; Key step is as follows:

(2.1) priority calculates, and determines repairing order: priority computing method have employed two indices: the degree of confidence C (p) of pixel and the data item D (p) of edge strength;

P (p)=C (p) * D (p) (1) wherein,

$\left\{\begin{array}{c}C\left(p\right)=\frac{{\mathrm{\Σ}}_{q\∈{\mathrm{\Ψ}}_p^{\mathrm{\Φ}}}C\left(q\right)}{\mathrm{Area}\left({\mathrm{\Ψ}}_p\right)}\\ D\left(p\right)=\frac{|{\▿I}_p^{\⊥}}{\mathrm{\α}}+\mathrm{\ϵ}\end{array}\right.---\left(2\right)$

Wherein, Ψ _pfor the object block to be repaired centered by p point, with be respectively Ψ _pin adhere to the part in Ω and Φ region separately, represent the isophote vector at p point place, n _pfor unit normal vector; α is normalized parameter, and ε is disturbance constant; The pixel around C (p) larger explanation p point with high confidence level is more, and D (p) larger explanation p point place is the joint of known region isophote and zone boundary to be repaired; The place that P (p) value is high should preferentially be repaired;

(2.2) point setting preferred value the highest as p, according to object block Ψ _pin the texture of part searches for best matching blocks in known region Φ, and the block that namely matching distance is minimum, is designated as

(2.3) will in with corresponding partial replication is to Ψ _pin, and upgrade the edge δ Ω in region to be repaired;

(2.4) more than, 3 continuous cyclings of step, until all regions to be repaired are all filled, complete texture repair.

In described step (3), contrast, consistance, auto-correlation and entropy four statistical nature computing formula are respectively:

Contrast: $\mathrm{CON}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i-j)}^{2}P(i,j)$

Consistance: $\mathrm{Homo}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}P(i,j){\log }_{2}P(i,j)$

Auto-correlation: $\mathrm{Cor}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\frac{\mathrm{ijP}(i,j)-{\mathrm{\μ}}_x{\mathrm{\μ}}_y}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$

Wherein:

${\mathrm{\μ}}_x=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\mathrm{iP}(i-j),{\mathrm{\μ}}_y=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\mathrm{jP}(i,j)$

${\mathrm{\σ}}_x=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i-{\mathrm{\μ}}_x)}^{2}P(i,j),{\mathrm{\σ}}_y=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i,{\mathrm{\μ}}_y)}^{2}P(i,j)$

Entropy: $\mathrm{Ee}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{P}^2(i-j).$

Compared with prior art, the print file discrimination method based on textures synthesis that the present invention proposes, does not need there is identical characters between printer document, improves the scope of application and the accuracy rate of print file discrimination method.

Accompanying drawing explanation

Fig. 1 is that printed document gathers character picture example;

Fig. 2 is textures synthesis image shows; The complete texture image of (a) original image (b) image (c) to be repaired;

Fig. 3 is Criminisi algorithm schematic diagram.

Embodiment

Below in conjunction with drawings and Examples, the invention will be further described.

1. integral image magnification at high multiple gathers and Character segmentation

Be amplified the printed document image after 150 times by integral image magnification at high multiple scanning system, then obtain by Character segmentation the single character picture enriching texture information, as shown in Figure 1.

2. textures synthesis

If to single character image zooming-out feature, so its eigenwert must contain the texture of image and morphosis two kinds of information of character.In printer is differentiated, the texture information of image is the important evidence differentiated, but due to the interference of character morphosis, the feature between different character pictures can not be directly used in coupling classification.In order to eliminate this interference, the present invention proposes to utilize the method for textures synthesis to repair the character picture of printed document.Texture synthesis method is various, and we have selected the method based on sample block that effect is proposed by Criminisi preferably, because the method effect is fine, so do not do any change.

For the texture repairing method based on sample block, first to mark the region to be repaired of image: first utilize Otsu algorithm to do binary conversion treatment to original image, tentatively distinguish texture region and background area, secondly the noise in background area is removed by medium filtering, but medium filtering cannot remove the noise of texture region marginal portion than comparatively dense, so need to operate by dilation erosion the pure property ensureing texture region, finally the background area after dilation erosion is modified as pure white in the corresponding position of original image, the white portion obtained is region to be repaired, as shown in (b) in Fig. 2.By calculating the priority of degree of confidence item and the data item decision repairing of repairing edge, then the region that search and multiblock to be repaired mate most in the known region of image, finally this matching area is copied to multiblock to be repaired, obtain complete texture image, in Fig. 2 shown in (c).

3. feature extraction

Classical gray level co-occurrence matrixes (GLCM) algorithm is directly quoted in texture image feature extraction, get horizontal and vertical directions, the scanning of the corresponding printer of difference and page orientation, step-length respectively gets 20 steps, obtain the gray level co-occurrence matrixes P of image, the size of matrix is K × K, and in matrix, each element representation is P (i, j), 4 conventional statistical natures of GLCM are finally calculated as characteristics of image:

(1) contrast: $\mathrm{CON}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i-j)}^{2}P(i,j)$

(2) consistance: $\mathrm{Homo}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}P(i,j){\log }_{2}P(i,j)$

(3) auto-correlation: $\mathrm{Cor}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\frac{\mathrm{ijP}(i,j)-{\mathrm{\μ}}_x{\mathrm{\μ}}_y}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$ Wherein:

${\mathrm{\μ}}_x=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\mathrm{iP}(i-j),{\mathrm{\μ}}_y=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\mathrm{jP}(i,j)$

${\mathrm{\σ}}_x=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i-{\mathrm{\μ}}_x)}^{2}P(i,j),{\mathrm{\σ}}_y=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i,{\mathrm{\μ}}_y)}^{2}P(i,j)$

(4) entropy: $\mathrm{Ee}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{P}^2(i-j)$

So often open feature totally 2 × 20 × 4=160 dimension of image zooming-out, be designated as x ^l, l=1,2 ..., 160.

4. Fusion Features

Because 4 statistical nature codomains of GLCM are different, first will do normalized respectively, for contrast metric, normalization formula is as follows:

$\left\{\begin{array}{c}{y}^l=\frac{x^l-x_{\min }}{x_{\max }-x_{\min }},l=\mathrm{1,2},...,40\\ x_{\max }=\max \left(x^l\right),x_{\min }=\min \left(x^l\right)\end{array}\right.---\left(1\right)$

All the other 3 statistical natures process all in this way, finally obtain new eigenwert y ^l, l=1,2 ..., 160.Remember that the new feature value that a printer M opens image is l=1,2 ..., 160, m=1,2 ..., M.Complete texture after synthesis is no longer by character form structure influence, and all characters of so a printed document can be regarded as same character, and therefore, the source machine feature of this printed document can be described as the characteristic mean that all M open image:

$\begin{array}{cc}{\mathrm{Fea}}^l=\frac{1}{M}\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{y}_m^l& l=\mathrm{1,2},...,160\end{array}---\left(2\right)$

5. sorter

Obtained the feature interpretation of a printed document by (2) formula, the matching distance between the source machine of two parts of printed documents is defined as their features with between city block distance:

$\mathrm{Dist}=\frac{\underset{l=1}{\overset{160}{\mathrm{\Σ}}}|{\mathrm{Fea}}_1^l-{\mathrm{Fea}}_2^l|}{\sqrt{\underset{l=1}{\overset{160}{\mathrm{\Σ}}}{\mathrm{Fea}}_1^l\underset{l=1}{\overset{160}{\mathrm{\Σ}}}{\mathrm{Fea}}_2^l}}---\left(3\right)$

Then adopting the method for threshold classification to adjust the distance to differentiate, judging that when being less than threshold value T two parts of printed documents are from same printer, judge from different machine when being greater than T.For obtaining this threshold value, we adopt the method for Bayesian Estimation.First take the printed document sample of 100 printers, calculate the range distribution between same printer document and the range distribution between different printer document respectively, minimum for criterion with Type Ⅰ Ⅱ error rate sum, find empirical value T.

6. test

Set up 30 printers, the database of every platform printer 80 Chinese characters, wherein half image makees sample, and half image is tested.Utilize Criminisi based on the texture synthesis method of sample block, obtain the complete texture of each character picture, then calculate their GLCM feature, finally obtain the distance between two printers, and itself and threshold value T are carried out comparison-of-pair sorting.Experimental result accuracy, false rejection rate FRR, false acceptance rate FAR tri-standards are evaluated, as shown in table 1.

Table 1 is identification result one to one

Texture synthesis method	Accuracy	Error rate	FRR	FAR
					Criminisi	98％	2％	2.22％	0

Below Criminisi texture synthesis method is elaborated:

Criminisi texture synthesis method utilizes in source region and fills region to be filled to be filled piece of match block the most similar.As shown in Figure 3, if region to be repaired is Ω, the edge in known region to be Φ, δ Ω be region to be repaired.Algorithm key step is as follows:

(1) priority calculates, and determines repairing order.Priority computing method have employed two indices: the degree of confidence C (p) of pixel and the data item D (p) of edge strength.

P(p)＝C(p)*D(p) (2)

Wherein,

$\left\{\begin{array}{c}C\left(p\right)=\frac{{\mathrm{\Σ}}_{q\∈{\mathrm{\Ψ}}_p^{\mathrm{\Φ}}}C\left(q\right)}{\mathrm{Area}\left({\mathrm{\Ψ}}_p\right)}\\ D\left(p\right)=\frac{|{\▿I}_p^{\⊥}}{\mathrm{\α}}+\mathrm{\ϵ}\end{array}\right.---\left(2\right)$

Ψ _pfor the object block to be repaired centered by p point, with be respectively Ψ _pin adhere to the part in Ω and Φ region separately, represent the isophote vector at p point place, n _pfor unit normal vector.α is normalized parameter, and ε is disturbance constant.The pixel around C (p) larger explanation p point with high confidence level is more, and D (p) larger explanation p point place is the joint of known region isophote and zone boundary to be repaired.The place that P (p) value is high should preferentially be repaired.

(2) point setting preferred value the highest as p, according to object block Ψ _pin the texture of part searches for best matching blocks in known region Φ, and the block that namely matching distance is minimum, is designated as

(3) will in with corresponding partial replication is to Ψ _pin, and upgrade the edge δ Ω in region to be repaired.

Above 3 continuous cyclings of step, until all regions to be repaired are all filled, complete texture repair.

Claims (7)

1., based on a print file discrimination method for textures synthesis, it is characterized in that: specifically comprise the following steps:

(1) integral image magnification at high multiple scanning system is utilized to gather the single character picture of printed document;

(2) by texture synthesis method, the white space in the single character picture of incompleteness is repaired as complete texture image;

(3) calculate the gray level co-occurrence matrixes of the complete texture image after repairing, and extract contrast, consistance, auto-correlation and entropy four statistical natures;

(4) normalized and Fusion Features are done to carried feature, calculate the diagnostic characteristics of printed document, and using the characteristic mean of all images of a printed document as the feature of affiliated printer;

(5) classified by sorter, differentiate that whether two parts of printed documents are from same printer.

2. a kind of print file discrimination method based on textures synthesis according to claim 1, it is characterized in that: in described step (5), calculate the characteristic distance of two parts of printed documents, when distance is less than empirical value, judge that two parts of documents are from same printer.

3. a kind of print file discrimination method based on textures synthesis according to claim 2, it is characterized in that: described empirical value adopts the method for Bayesian Estimation to calculate: the printed document sample first taking 100 printers, calculate the range distribution between same printer document and the range distribution between different printer document respectively, minimum for criterion with Type Ⅰ Ⅱ error rate sum, find empirical value T.

4. a kind of print file discrimination method based on textures synthesis according to claim 1, is characterized in that: described texture synthesis method comprise strain-based design Method and Process restorative procedure and all can be the method for complete texture by incomplete texture repair.

5. a kind of print file discrimination method based on textures synthesis according to claim 4, is characterized in that: described texture synthesis method adopts the texture repairing method based on sample block, first will mark the region to be repaired of image:

First, utilize Otsu algorithm to do binary conversion treatment to original image, tentatively distinguish texture region and background area;

Secondly, remove the noise in background area by medium filtering, but medium filtering cannot remove the noise of texture region marginal portion than comparatively dense, so need to operate by dilation erosion the pure property ensureing texture region;

Finally, the background area after dilation erosion is modified as pure white in the corresponding position of original image, and the white portion obtained is region to be repaired.

6. a kind of print file discrimination method based on textures synthesis according to claim 4, it is characterized in that: described texture synthesis method utilizes in source region and fills region to be filled to be filled piece of match block the most similar, if region to be repaired is Ω, the edge in known region to be Φ, δ Ω be region to be repaired; Key step is as follows:

(2.1) priority calculates, and determines repairing order: priority computing method have employed two indices: the degree of confidence C (p) of pixel and the data item D (p) of edge strength;

P(p)＝C(p)*D(p) (1)

Wherein,

$\left\{\begin{array}{c}C\left(p\right)=\frac{{\mathrm{\Σ}}_{q\∈{\mathrm{\ψ}}_p^{\mathrm{\Φ}}}C\left(q\right)}{\mathrm{Area}\left({\mathrm{\ψ}}_p\right)}\\ D\left(p\right)=\frac{|\▿I_p^{\⊥}\·n_p|}{\mathrm{\α}}+\mathrm{\ϵ}\end{array}\right.---\left(2\right)$

Wherein, Ψ _pfor the object block to be repaired centered by p point, with be respectively Ψ _pin adhere to the part in Ω and Φ region separately, represent the isophote vector at p point place, n _pfor unit normal vector; α is normalized parameter, and ε is disturbance constant; The pixel around C (p) larger explanation p point with high confidence level is more, and D (p) larger explanation p point place is the joint of known region isophote and zone boundary to be repaired; The place that P (p) value is high should preferentially be repaired;

(2.2) point setting preferred value the highest as p, according to object block Ψ _pin the texture of part searches for best matching blocks in known region Φ, and the block that namely matching distance is minimum, is designated as

(2.3) will in with corresponding partial replication is to Ψ _pin, and upgrade the edge δ Ω in region to be repaired;

(2.4) more than, 3 continuous cyclings of step, until all regions to be repaired are all filled, complete texture repair.

7. a kind of print file discrimination method based on textures synthesis according to claim 1, is characterized in that: in described step (3), and contrast, consistance, auto-correlation and entropy four statistical nature computing formula are respectively:

Contrast: $\mathrm{CON}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i-j)}^{2}P(i,j)$

Consistance: $\mathrm{Homo}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}P(i,j){\log }_{2}P(i,j)$

Auto-correlation: $\mathrm{Cor}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\frac{\mathrm{ijP}(i,j)-{\mathrm{\μ}}_x{\mathrm{\μ}}_y}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$

Wherein:

${\mathrm{\μ}}_x=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\mathrm{iP}(i,j),{\mathrm{\μ}}_y=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}\mathrm{jP}(i,j)$

${\mathrm{\σ}}_x=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(i-{\mathrm{\μ}}_x)}^{2}P(i,j),{\mathrm{\σ}}_y=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{(j-{\mathrm{\μ}}_y)}^{2}P(i,j)$

Entropy: $\mathrm{Ee}=\underset{i=0}{\overset{K-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{K-1}{\mathrm{\Σ}}}{P}^2(i,j).$