CN102842119A - Quick document image super-resolution method based on image matting and edge enhancement - Google Patents

Abstract

The invention discloses a quick document image super-resolution method based on image matting and document enhancement. The quick document image super-resolution method based on the image matting and the edge enhancement comprises the following steps: the image matting technology based on guide filtering is used to divide a low-resolution document image into three component parts of a foreground color, a background color and foreground transparency; a quick interpolation method and an interpolation method based on the edge enhancement are respectively used to do super resolution to the foreground color part, the background color part and the foreground transparency part; the foreground color part, and the background color part and the foreground transparency part which are after being done super solution are used to reconstruct a super-resolution document image. The quick document image super-resolution method based on the image matting and the document enhancement can quickly and effectively improve the quality of the document image which is after being done super resolution, improve the visual effect of the document image, and can solve the displaying or printing problems of a low-resolution document image or video on super-resolution electronic equipment, and has vital important practical application value.

Description

A kind of based on scratching the fast text image super-resolution method that picture and edge strengthen

Technical field

The present invention relates to a kind of text image super-resolution method, it is a kind of based on scratching the text image super-resolution method of picture with the edge enhancing to say so more specifically.

Background technology

The text image super-resolution is a branch of image super-resolution research, and in general image super-resolution method can be divided three classes: based on the method for sample learning, based on the method for reconstruct with based on the method for interpolation.Method based on sample learning solves the super-resolution problem through the corresponding relation between study high-definition picture piece/parameter and the low-resolution image piece/parameter; These class methods need consume considerable time and carry out sample training, and sample image choose oversubscription result's influence very big.Come iteration to go out final high-definition picture based on the method for reconstruct through the error between the low-resolution image that minimizes original low-resolution image and constitute by the high-definition picture down-sampling; This method need not trained, but can not guarantee that the final high-definition picture of algorithm convergence has desired visual effect ^[3]Compare these methods, image interpolation method such as bilinear interpolation and bicubic interpolation are simply quick, therefore obtain to use more widely.But simple interpolation method can cause soft edge and edge sawtooth.Therefore, some interpolation methods based on the edge have appearred recently ^[4-5], these class methods guarantee the clear and continuous of image border as far as possible in the process of carrying out interpolation, therefore can solve the fuzzy and sawtooth problem of image border to a certain extent.But, compare with simple interpolation method, the calculation consumption of these class methods is too high, is difficult to satisfy some practical application request, such as low resolution text image or video on high resolution display, show etc.Q. people such as Shan proposes based on the up-sampling of iteration and the quick super-resolution method of deblurring ^[6], still this method is lost image texture easily or artifacts occurred at the place, image border in the deblurring process, and the Rapid Realization of this algorithm need be based upon on the basis of GPU acceleration.

Summary of the invention

For solving the problems referred to above that exist on text image super-resolution method speed and the effect, the invention provides a kind of based on scratching the fast text image super-resolution method that picture and edge strengthen.The algorithm speed that the present invention proposes is fast and can when amplifying text image, effectively guarantee the sharpness at text image edge.

The technical scheme that the present invention addresses the above problem may further comprise the steps:

1) the low resolution text image is carried out binaryzation, obtain the bianry image of low resolution text image.

2) with low resolution text image and bianry image thereof as input, adopt the foreground to transparent degree image that obtains text image as algorithm of scratching based on guiding filtering;

3) according to former text image and step 2) the foreground to transparent degree image that obtains estimates foreground and background colour;

4) foreground and the background colour to the low resolution text image partly adopts interpolation algorithm to carry out SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

5) the foreground to transparent degree to the low resolution text image partly adopts edge enhancing interpolation algorithm to carry out SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

6) foreground, background colour and the combination of foreground to transparent degree that will pass through after the SUPERRESOLUTION PROCESSING FOR ACOUSTIC obtain final high resolving power text image;

In the above-mentioned fast text image super-resolution method based on stingy picture and edge enhancing, described step 1) adopts big Tianjin method to choose the threshold value of text image binaryzation.

In the above-mentioned fast text image super-resolution method based on stingy picture and edge enhancing; Described step 2) in the low resolution text image as the guiding text image; Thereby the binary image channeling conduct filtering to the low resolution text image obtains scratching picture result, i.e. the foreground to transparent degree of low resolution text image.

Above-mentioned based on scratching in the fast text image super-resolution method that picture and edge strengthen, utilize the guiding filtering algorithm to the separating of local linear model in the described step 3), the foreground to transparent degree for input, obtains foreground and background colour to original image as guiding.

In the above-mentioned fast text image super-resolution method, adopt the expansion erosion algorithm in the morphological image processing to obtain foreground and background colour in the described step 3) based on stingy picture and edge enhancing.

Above-mentioned based on scratching in the fast text image super-resolution method that picture and edge strengthen, adopt first interpolation to carry out the edge again in the described step 4) and strengthen and carry out the edge earlier and strengthen again the text image interpolation method that interpolation combines.

In the above-mentioned fast text image super-resolution method, adopt the bicubic interpolation method in the described step 5) based on stingy picture and edge enhancing.

The above-mentioned fast text image super-resolution method based on stingy picture and edge enhancing adopts bilinear interpolation method in the described step 5).

In the above-mentioned fast text image super-resolution method, adopt the arest neighbors interpolation method in the said step 5) based on stingy picture and edge enhancing.

Owing to adopt technique scheme; Technique effect of the present invention is: the present invention adopts based on the text image super-resolution method of scratching picture technology and edge enhancing and has effectively improved the sharpness of amplifying the back text image; The speed of algorithm is very fast simultaneously; Therefore; Can be used to solve problems such as low branch rate text image or video demonstration and the printing on high resolution display and printer, the subsequent treatment of text image and the application on electronic equipment thereof are had most important theories meaning and using value.

Description of drawings

Fig. 1 is a process flow diagram of the present invention.

Fig. 2 is that the text image super-resolution result of different super-resolution methods compares.

Embodiment

Referring to Fig. 1, Fig. 1 is a process flow diagram of the present invention.Be input as the gray scale or the color text image of low resolution, be output as high-resolution gray scale or color text image.Practical implementation step of the present invention is following:

1. adopt stingy picture technology that the low resolution text image is divided into low resolution foreground F based on guiding filtering ^L, low resolution background colour B ^LAnd low resolution foreground to transparent degree Q ^LThree parts.Its concrete steps are:

1) to low resolution text image I ^LCarry out binaryzation and obtain two-value text image A.The threshold value TH of binaryzation adopts big Tianjin method to choose.

$A(x,y)=\left\{\begin{array}{cc}1& I^L(x,y)>;\mathrm{TH}\\ 0& \mathrm{otherwise}\end{array}\right.---\left(1\right)$

2) shown in formula (2), with former text image I ^LFor the guiding text image, the two-value text image A that obtains is carried out the stingy picture based on guiding filtering, obtain the low foreground to transparent degree part Q that differentiates text image ^L

Q ^L=GD(A,I ^L) （2）

I wherein ^LBe to be referenced text image (low resolution gray document image; If text image is the color text image then is converted into gray document image through the mean value that calculates three Color Channels); A is a low resolution binaryzation text image of treating that the step 1) of filtering obtains, and guiding filtering output result is the foreground to transparent degree part Q of low resolution text image ^L.

The practical implementation details of guiding filtering is following: at first adopt linear regression method calculating sealing of cost function shown in formula (3) to separate a _kWith b _k(shown in formula (4), (5)).

$E(a_k,b_k)=\underset{i\∈{\mathrm{\ω}}_k}{\mathrm{\Σ}}({(a_k{i}_i^L+b_k-A_i)}^2+\mathrm{\ϵ}{a}_k^2)---\left(3\right)$

$a_k=\frac{\frac{1}{\left|\mathrm{\ω}\right|}{\mathrm{\Σ}}_{i\∈{\mathrm{\ω}}_k}{I}_i^L{S}_i-{\mathrm{\μ}}_k{\stackrel{\‾}{S}}_k}{{\mathrm{\δ}}_k^2+\mathrm{\ϵ}}---\left(4\right)$

$b_k={\stackrel{\‾}{S}}_k-a_k{\mathrm{\μ}}_k---\left(5\right)$

ω wherein _kBe meant with k pixel to be that radius centered is the square window of r, | ω | represent the number of this window interior pixel, μ _kWith

Represent and guide text image the window ω that is the center with k pixel respectively _kThe average of interior pixel and variance,

The window ω that the filtering text image is the center with k pixel is treated in expression _kThe average of interior pixel, ε be constant (in we are bright, windows radius r=2, ε=0.1 ³).Then according to the output of formula (6) calculating filter:

$Q_i^L={\stackrel{\‾}{a}}_i{I}_i^L+{\stackrel{\‾}{b}}_i---\left(6\right)$

Wherein ${\stackrel{\‾}{a}}_i=\frac{1}{\left|\mathrm{\ω}\right|}{\mathrm{\Σ}}_{k\∈{\mathrm{\ω}}_i}{a}_k$ , ${\stackrel{\‾}{b}}_i=\frac{1}{\left|\mathrm{\ω}\right|}{\mathrm{\Σ}}_{k\∈{\mathrm{\ω}}_i}{b}_k$

3) calculate the foreground to transparent degree Q of text image ^LAfter, we propose two kinds of diverse ways and estimate text image foreground and background colour:

Method one: according to foreground to transparent degree Q ^LWith former text image I ^LCalculate foreground image I _fWith background image I _b, then the two is corroded expansion and obtains corresponding foreground F ^LWith background colour B ^L

$B^L=I_b\⊕H---\left(8\right)$

Wherein Expression corrosion operation,

The expression expansive working, I _f=I ^LQ ^L, I _b=I ^L(1-Q ^L), H is one a square structure element:

$H=\left[\begin{array}{ccc}1& 1& 1\\ 1& 1& 1\\ 1& 1& 1\end{array}\right]---\left(9\right)$

Method two: for Q ^LValue is the pixel of big (greater than 0.9) very, and its foreground directly substitutes with former text image pixel value, for Q ^LThe pixel that is worth very little (less than 0.1), its background colour directly substitute with former text image pixel value, and for Q ^LValue is positioned at [0.1-0.9] interval pixel, and its foreground and background colour are separated (like formula (11), shown in (12)) according to the sealing of the cost function shown in the formula (10) and asked for.

$E((F^L-B^L),B^L)=\underset{w_k}{\mathrm{\Σ}}[{((F^L-B^L)q^L+B^L-I^L)}^2+\mathrm{\ϵ}{(F^L-B^L)}^2]---\left(10\right)$

$F_k^L-B_k^L=\frac{1}{\left|w_i\right|}{\mathrm{\&Sigma;}}_{k\&Element;{\mathrm{\&omega;}}_{\mathrm{<figure-callout\; id="1"\; label="i"\; filenames="HDA00002029248000011.png"\; state="\{\{state\}\}">i</figure-callout>}}}\frac{\frac{1}{\left|w_k\right|}{\mathrm{\&Sigma;}}_{j\&Element;w_k}{q}_j^L{I}_j^L-{\mathrm{\&eta;}}_k{\stackrel{\&OverBar;}{I^L}}_k}{{\mathrm{\&sigma;}}_k^2+\mathrm{\&epsiv;}}---\left(11\right)$

$B_k^L=\frac{1}{\left|w_i\right|}{\mathrm{\&Sigma;}}_{k{\&Element;w}_i}{\stackrel{\&OverBar;}{I^L}}_k-X_k{\mathrm{\&eta;}}_k---\left(12\right)$

Wherein

, w _kBe to be 3 * 3 the square window at center with k pixel, | w| representes the number of this window interior pixel, η _kWith

Represent foreground to transparent degree text image Q respectively ^LWith k the pixel window w that is the center _kThe average of interior pixel and variance.

2. the foreground F that adopts quick interpolation method that step 1 is obtained ^LWith background colour B ^LCarry out super-resolution respectively and obtain high-resolution foreground F ^HWith background colour B ^HHere said quick interpolation method mainly is meant arest neighbors interpolation, Fast numerical algorithms such as bilinear interpolation and bicubic interpolation.The arest neighbors interpolation thinks that the gray-scale value of interpolation point equals the gray-scale value of nearest known point; Bilinear interpolation utilizes around the interpolation position gray-scale value of four known points to do linear interpolation on the both direction; The bicubic interpolation method utilizes around the interpolation position gray-scale value of 16 points to do cubic interpolation, considers that not only the gray scale of 4 direct neighbor points influences, and considers the influence of gray-value variation rate between each adjoint point.

3. adopt the edge Q of the interpolation method of edge enhancing to text image ^LCarry out the edge text image Q after super-resolution obtains oversubscription ^H, the practical implementation step is following:

$Q_1^H=\mathrm{Enhance}\left(\mathrm{Bicubic}\left(Q^L\right)\right)---\left(13\right)$

$Q_2^H=\mathrm{Bicubic}\left(\mathrm{Enhance}\left(Q^L\right)\right)---\left(14\right)$

$Q^H=(1-\mathrm{\&lambda;})Q_1^H+\mathrm{\&lambda;}{Q}_2^H---\left(15\right)$

Wherein the span of λ is [0,1], is made as 0.5 at the default value of λ of the present invention, and Bicubic () representes bicubic interpolation, and Enhance () is that text image strengthens operation, and it can be represented with following formula:

Enhance(I)=I-I*○ （16）

Wherein I representes text image to be strengthened, and * representes convolution algorithm, zero expression convolution kernel as follows:

4. with high-resolution foreground F ^H, background colour B ^HAnd foreground to transparent degree part Q ^HMake up as follows:

I ^H(x,y)=Q ^H(x,y)F ^H(x,y)+(1-Q ^H(x,y))B ^H(x,y) （18）

Wherein (x y) is pixel coordinate, I ^HBe final text image super-resolution result.

Method proposed by the invention with based on bicubic interpolation with compare based on the text image super-resolution method of rarefaction representation.A left side is played first to classify as is the low resolution text image in the accompanying drawing 2; Second classifies the high resolving power text image that uses the bicubic interpolation method to obtain as; The 3rd row are based on the high resolving power text image that the super-resolution method of rarefaction representation obtains, and last classifies the high resolving power text image that adopts the inventive method to obtain as.Can find out the present invention and the sharpness that can both when amplifying text image, effectively promote text image based on the method for rarefaction representation, but bicubic interpolation and method based on rarefaction representation all can cause in various degree fuzzy of text image edge.

The computing time that table 1 has been listed algorithms of different relatively, in the table institute free all be to use MATLAB to be the working time on the computer of 2.8GHz at CPU frequency.Can find out; Although the algorithm based on rarefaction representation can obtain the super-resolution efect suitable with this algorithm, its computing velocity is very slow, and the color text image of handling 256 * 256 needs 815 seconds; It is fast a lot of that yet the method that the present invention proposes is wanted; The color text image of handling 256 * 256 only needs 0.186 second, and the speed of one 256 * 256 color text image of C++ routine processes of this algorithm can reach 0.047 second in addition, so the present invention has important practical value.

The working time of the different super-resolution algorithms of table 1 relatively

Claims (9)

1. one kind based on scratching the fast text image super-resolution method of picture with the edge enhancing, comprises following concrete steps:

1) the low resolution text image is carried out binaryzation, obtain the bianry image of low resolution text image.

2) with low resolution text image and bianry image thereof as input, adopt the foreground to transparent degree image that obtains text image as algorithm of scratching based on guiding filtering;

3) according to former text image and step 2) the foreground to transparent degree image that obtains estimates foreground and background colour;

4) foreground and the background colour to the low resolution text image partly adopts interpolation algorithm to carry out SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

5) the foreground to transparent degree to the low resolution text image partly adopts edge enhancing interpolation algorithm to carry out SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

6) foreground, background colour and the combination of foreground to transparent degree that will pass through after the SUPERRESOLUTION PROCESSING FOR ACOUSTIC obtain final high resolving power text image;

2. according to claim 1 based on scratching the fast text image super-resolution method of picture with the edge enhancing, described step 1) adopts big Tianjin method to choose the threshold value of text image binaryzation.

3. according to claim 1 based on scratching the fast text image super-resolution method of picture with the edge enhancing; Described step 2) with the low resolution text image as the guiding text image; Thereby the binary image channeling conduct filtering to the low resolution text image obtains scratching picture result, i.e. the foreground to transparent degree of low resolution text image.

4. according to claim 1 based on scratching the fast text image super-resolution method of picture with the edge enhancing; Utilize guiding filtering algorithm separating in the described step 3) to local linear model; The foreground to transparent degree for input, obtains foreground and background colour to original image as guiding.

5. according to claim 4 based on scratching the fast text image super-resolution method that picture and edge strengthen, adopt the expansion erosion algorithm in the morphological image processing to obtain foreground and background colour in the described step 3).

6. according to claim 1 based on scratching the fast text image super-resolution method that picture and edge strengthen, adopt first interpolation to carry out the edge again in the described step 4) and strengthen and carry out the edge earlier and strengthen again the text image interpolation method that interpolation combines.

7. according to claim 1 based on scratching the fast text image super-resolution method of picture with the edge enhancing, the employing bicubic interpolation method in the described step 5).

8. according to claim 1 based on scratching picture and the fast text image super-resolution method that the edge strengthens, adopt bilinear interpolation method in the described step 5).

9. according to claim 1 based on scratching picture and the fast text image super-resolution method that the edge strengthens, adopt the arest neighbors interpolation method in the described step 5).

Images