CN106934806A - It is a kind of based on text structure without with reference to figure fuzzy region dividing method out of focus - Google Patents

Abstract

The present invention disclose it is a kind of based on text structure without with reference to figure fuzzy region dividing method out of focus, comprise the following steps：(1) zoomed image, by about 1/4 times that image scaling is original image area；(2) definition difference is calculated, the text structure of image correspondence position image block after artwork and scaling is calculated respectively, and calculate the difference of the two；(3) fuzzy region is extracted, the noise of error image is filtered, fuzzy region is partitioned into using image segmentation algorithm, and the result after segmentation is up-sampled.For the fuzzy region out of focus segmentation of non-reference picture, the present invention constructs zoomed image using original image, the definition of zoomed image and original image is calculated respectively, and then obtains fuzziness distributed image, be finally fast and effeciently partitioned into image focus fuzzy region.

Description

It is a kind of based on text structure without with reference to figure fuzzy region dividing method out of focus

Technical field

The present invention relates to Digital image technology, and in particular to it is a kind of based on text structure without with reference to figure confusion region out of focus Domain splitting method.

Background technology

Image blurring is exactly a kind of common image degradation process, it is usually because in exposure process the movement of camera, The movement or focusing of object are forbidden and are produced, and are artificially formed to pursue certain artistic effect in the case of also having, It is included in the image processing process in photography or later stage.It is out of focus it is fuzzy be it is a kind of common image blurring, mainly by focusing Caused by inaccurate.The image blurring loss that result in information in image, is that further treatment image causes difficulty.It is accurately high The fuzzy pixel of discovery of effect suffers from important and actual in fields such as image segmentation, object detection, scene classification, picture edittings Using.In with reference to artwork paste region segmentation application, only one input picture, current non-reference picture fuzzy region out of focus is divided Segmentation method mainly includes picture frequency domain method, Space domain and some methods for combining machine learning algorithm.But these Algorithm is primarily present two problems, and first, processing speed is slow, and practicality is not strong, and secondly, segmentation effect is not good.

One preferable image quality evaluation index can effectively distinguish fuzzy and picture rich in detail, therefore can be used to do Split image obscuring area.Without reference graph structure resolution chart as quality evaluation index (NRSS) is exactly that an effect is preferably schemed As quality evaluation index, it is realized without with reference to figure image quality evaluation on the basis of text structure (SSIM) index.But It is that the method for carrying out image obscuring area segmentation using NRSS at present is also fewer, even if there is last segmentation effect also poor.

The content of the invention

Goal of the invention：It is an object of the invention to solve the deficiencies in the prior art, there is provided one kind is clear based on structure Clear degree without with reference to figure fuzzy region dividing method out of focus.

Technical scheme：It is of the present invention it is a kind of based on text structure without with reference to figure fuzzy region segmentation side out of focus Method, comprises the following steps：

(1) original image is scaled：Original image equal proportion is scaled about 0.25 times of original image area, i.e. original graph The size of picture is M × N, and the size of image is after scalingUsing bilinear interpolation to the image after scaling Enter row interpolation, so as to the image after being scaled；

(2) image definition difference is calculated：

(2.1) piecemeal is carried out to the image after scaling in original image and step (1), so as to obtain original image set of blocks Image block set S after R and scaling；

(2.2) the clear journey for being used to weigh each image block without reference graph structure similitude of each image block is calculated respectively Degree；

(2.3) difference of each image block readability is calculated respectively, obtains matrix of differences；

(3) it is partitioned into fuzzy region：

(3.1) matrix of differences obtained in step (2) is filtered using Steerable filter and filters noise；

(3.2) image after denoising is split using Otsu threshold method, i.e., first finding out can maximize inter-class variance Gray value, and binaryzation is carried out with this gray value；

(3.3) image after segmentation is up-sampled, reverts to original image yardstick.

Further, the detailed process of the step (2.1) is：

First, image block is chosen from original image, step-length is (2,2), i.e.,：Since the upper left corner of image, first every time 2 pixels are moved along the x-axis, 2 pixels are moved then along Y, repeated the above steps, untill all image blocks of acquisition, choosing The image block set for taking is R, and each image block is designated as R_i,j；

Wherein, i represents line flag, and j represents row mark, it is assumed that the size of image block is 2m × 2n, then

Then, image block is chosen from the image after scaling, step-length is (1,1), i.e.,：Since the upper left corner of image, first 1 pixel is moved along X-axis every time, 1 pixel is moved then along Y-axis, repeated the above steps, until obtaining all image blocks Untill；The image block sequence of selection is designated as S, and each image block is designated as S_i,j；

Wherein, i represents line flag, and j represents row mark, it is assumed that the size of image block is m × n, then

The number of image block should be consistent in set R and S.

Further, the detailed process of the step (2.2) is：

Assuming that given image block P, as follows without the calculation procedure with reference to graph structure similitude：Blurred picture block P, uses Gauss Obscure that image block obscure and obtain image P_b；Gradient is extracted, the horizontal and vertical side of Sobel operator extraction image blocks is used To gradient, image block P and P_bGradient image be designated as G and G respectively_b；Find out the most abundant N number of image of information in gradient image Block, the abundant degree of gradient information is weighed using the variance of image, that is, find out the wherein maximum N number of image block of variance；Calculate figure It is without the computing formula with reference to graph structure definition NRSS, NRSS as block P：

Wherein, image block is designated as G in Gⁱ,G_bMiddle image block is designated as Gⁱ _b；SSIM functions are used for calculating two knots of picture block Structure similitude, the function considers the correlation of the brightness, contrast and structural information between image block simultaneously.In given two figures In the case of as block a, b, SSIM functions can be expressed as：

SSIM (a, b)=[l (a, b)]^α[c(a,b)]^β[s(a,b)]^γ

Wherein,

S (a, b)=(σ_ab+C₃)/(σ_aσ_b+C₃),

u_a,u_bThe gradation of image average of image block a, b, σ are represented respectively_a,σ_bThe gradation of image of image block a, b is represented respectively Standard deviation, σ_abThe gradation of image covariance of image block a, b is represented, α, beta, gamma is parameter item, C₁,C₂,C₃Constant term, with case Only denominator close to zero when there is the unstable situation of computing.

And then respectively obtain NRSS (S_i,j) and NRSS (R_i,j)。

Further, the method for the step (2.3) is as follows：

Calculate the difference M ' of correspondence image block definition in image block set R and S_i,j, so as to obtain matrix of differences M '= {M′_i,j}：

M′_i,j=NRSS (S_i,j)-NRSS(R_i,j)；

Matrix M ' is normalized and obtains matrix M={ M_i,j, normalized using max min：

M_i,j=(M '_i,j- min (M '))/(max (M ')-min (M ')), the maximum of element wherein in max (M ') representing matrix Value, the minimum value of element in min (M ') representing matrix.

Beneficial effect：Without loss of generality, after image down, originally fuzzy image can be apparent from the present invention, therefore can Using use reduce after image and original image poor definition as new discriminant criterion.Image can be obtained based on this method Fuzziness is distributed, and with reference to image segmentation algorithm, is finally effectively partitioned into image obscuring area.By image scale transform, institute The data volume that need to be processed is substantially reduced, so algorithm process speed of the invention is fast, meanwhile, segmentation effect is better than existing algorithm.

In sum, the present invention have the advantages that splitting speed is fast, segmentation effect is good, suitable for non-reference picture.

Brief description of the drawings

Fig. 1 is original-gray image in embodiment；

Fig. 2 is the image being manually partitioned into using other method in embodiment；

Fig. 3 is matrix of differences after the normalization drawn using the present invention in embodiment；

Fig. 4 is the segmentation result obtained using the present invention in embodiment；

Fig. 5 is effect contrast figure in embodiment.

Specific embodiment

Technical solution of the present invention is described in detail below, but protection scope of the present invention is not limited to the implementation Example.

Embodiment：

Step 1：Original color image is read, coloured image matrix is obtained；

Step 2：It is gray level image matrix by coloured image matrix conversion, so as to obtain gray level image as described in Figure 1, this When determine image size be 640x621；

Step 3：Zoomed image, it is determined that the size of image is 320x310 after scaling, using linear interpolation shown in Fig. 1 Image of the image scaling into above-mentioned size；

Step 4：Image block simultaneously calculates the definition of each image block.Image after original image and scaling is divided Block, in the process, the moving step length of original image is (2,2), and the moving step length of image is (1,1), original image after scaling Tile size be 32x32, after scaling the tile size of image be 16x16, obtain original picture block after the completion of this process Image block set S after set R and scaling.To setIn each image block calculate without refer to graph structure definition (NRSS), it is used to weigh the readability of each image block, original image definition matrix R is obtained after the completion of this process^cAnd contracting Put rear image definition matrix S^c；

Step 5：The difference of definition is calculated, according to the definition matrix R being calculated in step 4^cAnd S^c, can obtain difference Matrix M '=S^c-R^c；

Step 6：Matrix of differences M ' is normalized and obtains M, specific method for normalizing can be minimum using maximum Value method for normalizing, i.e. M_i,j=(M '_i,j- min (M '))/(max (M ')-min (M ')), the normalized image for obtaining such as Fig. 3 institutes Show.

Step 7：Steerable filter is carried out to matrix M, using M itself as reference picture；

Step 8：Matrix of differences M is split using Da-Jin algorithm；

Step 9：Enlarged drawing, is up-sampled to the result after segmentation and is entered row interpolation using linear interpolation algorithm, is obtained To final segmentation result, as shown in Figure 4.

In addition, obtaining image as shown in Figure 2 using artificial segmentation to the original-gray image in Fig. 1 again, then will adopt The structure split with two methods carries out Contrast on effect, as shown in figure 5, white is overlap clear area in figure, black is to overlap Fuzzy region, grey is erroneous judgement region.

It can thus be seen that segmentation effect of the present invention is good, more precisely.

Claims (4)

1. it is a kind of based on text structure without with reference to figure fuzzy region dividing method out of focus, comprise the following steps：

(1) original image is scaled：Original image equal proportion is scaled about 0.25 times of original image area, i.e. original image Size is M × N, and the size of image is after scalingThe image after scaling is carried out using bilinear interpolation Interpolation, so as to the image after being scaled；

(2) image definition difference is calculated：

(2.1) piecemeal is carried out to the image after scaling in original image and step (1), so as to obtain original image set of blocks R and Image block set S after scaling；

(2.2) readability for being used to weigh each image block without reference graph structure similitude of each image block is calculated respectively；

(2.3) difference of each image block readability is calculated respectively, obtains matrix of differences；

(3) it is partitioned into fuzzy region：

(3.1) matrix of differences obtained in step (2) is filtered using Steerable filter and filters noise；

(3.2) image after denoising is split using Otsu threshold method, i.e., first finding out can maximize the ash of inter-class variance Angle value, and binaryzation is carried out with this gray value；

(3.3) image after segmentation is up-sampled, reverts to original image yardstick.

2. it is according to claim 1 based on text structure without with reference to figure fuzzy region dividing method out of focus, its feature It is：The detailed process of the step (2.1) is：

First, image block is chosen from original image, step-length is (2,2), i.e.,：Since the upper left corner of image, first every time along X-axis Mobile 2 pixels, 2 pixels are moved then along Y, are repeated the above steps, untill all image blocks of acquisition, the figure of selection As set of blocks is R, each image block is designated as R_i,j；

Wherein, i represents line flag, and j represents row mark, it is assumed that the size of image block is 2m × 2n, then

Then, image block is chosen from the image after scaling, step-length is (1,1), i.e.,：Since the upper left corner of image, first every time 1 pixel is moved along X-axis, 1 pixel is moved then along Y-axis, repeated the above steps, be until obtaining all image blocks Only；The image block sequence of selection is designated as S, and each image block is designated as S_i,j；

Wherein, i represents line flag, and j represents row mark, it is assumed that the size of image block is m × n, then

The number of image block should be consistent in set R and S.

3. it is according to claim 1 based on text structure without with reference to figure fuzzy region dividing method out of focus, its feature It is：The detailed process of the step (2.2) is：

Assuming that given image block P, as follows without the calculation procedure with reference to graph structure similitude：Blurred picture block P, uses Gaussian Blur Image block obscure and obtains image P_b；Extract gradient, using Sobel operator extraction image blocks both horizontally and vertically Gradient, image block P and P_bGradient image be designated as G and G respectively_b；Find out the most abundant N number of image block of information in gradient image, ladder The abundant degree of degree information is weighed using the variance of image, that is, find out the wherein maximum N number of image block of variance；Calculate image block P Be without the computing formula with reference to graph structure definition NRSS, NRSS：

$NRSS\left(P\right)=1/N\underset{1}{\overset{N}{\Σ}}SSIM(G^i,G_b^i);$

Wherein, image block is designated as G in Gⁱ,G_bMiddle image block is designated asThe structure that SSIM functions are used for calculating two picture blocks is similar Property, the function considers the correlation of the brightness, contrast and structural information between image block simultaneously.Given two image block a, In the case of b, SSIM functions can be expressed as：

SSIM (a, b)=[l (a, b)]^α[c(a,b)]^β[s(a,b)]^γ

Wherein,

$l(a,b)=(2u_a{u}_b+C_1)/(u_a^2+u_b^2+C_1),$

$c(a,b)=(2{\mathrm{\σ}}_a{\mathrm{\σ}}_b+C_2)/({\mathrm{\σ}}_a^2+{\mathrm{\σ}}_b^2+C_2),$

S (a, b)=(σ_ab+C₃)/(σ_aσ_b+C₃),

u_a,u_bThe gradation of image average of image block a, b, σ are represented respectively_a,σ_bThe gradation of image standard of image block a, b is represented respectively Difference, σ_abThe gradation of image covariance of image block a, b is represented, α, beta, gamma is parameter item, C₁,C₂,C₃It is constant term, is used to prevent point There is the unstable situation of computing when mother is close to zero；

And then respectively obtain NRSS (S_i,j) and NRSS (R_i,j)。

4. it is according to claim 1 based on text structure without with reference to figure fuzzy region dividing method out of focus, its feature It is：The method of the step (2.3) is as follows：

Calculate the difference M ' of correspondence image block definition in image block set R and S_i,j, so as to obtain matrix of differences M '= {M′_i,j}：

M′_i,j=NRSS (S_i,j)-NRSS(R_i,j)；

Matrix M ' is normalized and obtains matrix M={ M_i,j, normalized using max min：

M_i,j=(M'_i,j- min (M'))/(max (M')-min (M')), the maximum of element in wherein max (M') representing matrix, The minimum value of element in min (M') representing matrix.