CN103971360A - Method and device for calculating picture texture feature - Google Patents

Abstract

An embodiment of the invention provides a method and a device for calculating picture texture feature. The method includes dividing a source picture into at least one picture chip of X*Y; performing wavelet transformation on each picture chip, and calculating texture feature of each picture chip; calculating overall texture feature of the source picture according to the texture feature of each picture chip. Accuracy and speed in extraction of the texture feature of each picture chip of X*Y size are taken into consideration; after the source picture is reasonably partitioned, wavelet transformation is performed on each picture chip to extract the texture feature, and the overall texture feature of the source picture is calculated, so that calculating speed and accuracy are high.

Description

A kind of method and apparatus that calculates picture textural characteristics

Technical field

The present invention relates to image processing technique, refer to especially a kind of method and apparatus that calculates picture textural characteristics.

Background technology

The classic method of extracting the textural characteristics of graphic image is statistics, for example gray level co-occurrence matrixes method, but its amount that need to add up and calculate is larger, has reduced the efficiency of extracting, and do not make full use of the mankind's vision perception characteristic, cause cutting apart image and be suitable for precision have very large restriction.

Based on transform methods such as wavelet transformations, extraction efficiency is high, takes full advantage of the mankind's vision perception characteristic, and part sensitivity to high frequency.

There are the following problems for prior art: the subsequent treatment that wavelet transformation extracts after detail coefficients does not have general way, if process according to concrete environment and experimental result, can in performance, bring huge difference.

Summary of the invention

The technical problem to be solved in the present invention is to provide a kind of method and apparatus that calculates picture textural characteristics, solves in prior art, and the subsequent treatment after wavelet transformation extraction detail coefficients does not have the defect of the general way of function admirable.

For solving the problems of the technologies described above, embodiments of the invention provide a kind of method of calculating picture textural characteristics, and method comprises: the picture fritter that is at least one X*Y by source picture segmentation; Each picture fritter is carried out to wavelet transformation, calculate the textural characteristics of picture fritter; Calculate the overall textural characteristics of source picture according to the textural characteristics of each picture fritter.

In described method, before the picture fritter that is at least one X*Y by source picture segmentation, also comprise: the gray-scale value that calculates each pixel in the picture of source.

In described method, the picture fritter that is at least one X*Y by source picture segmentation, specifically comprises: X=4 in each picture fritter, and Y=4, the matrix of a corresponding 4*4, the element lacking in matrix mends 0.

In described method, each picture fritter is carried out to wavelet transformation, specifically comprise: each picture fritter comprises: low frequency part LL and HFS HL/LH/HH, HFS HH/HL/LH is corresponding level detail, vertical detail and diagonal line details respectively, comprises respectively 2*2 detail coefficients; For each picture fritter, adopt HAAR wavelet function to calculate the detail coefficients of HFS HL/LH/HH.

In described method, the textural characteristics that calculates picture fritter specifically comprises: textural characteristics $f_{\mathrm{block}}=\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HH}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HL}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{LH},$ Wherein, c _ijit is the detail coefficients after wavelet transformation.

In described method, calculate the overall textural characteristics of source picture according to the textural characteristics of each picture fritter, comprising: f _picbe the overall textural characteristics of source picture, m is the number of picture fritter, f _ithe textural characteristics f of each picture fritter _block.

A kind of picture processing device, comprising: point module unit, for the picture fritter that is at least one X*Y by source picture segmentation; Wavelet transform unit, for each picture fritter is carried out to wavelet transformation, calculates the textural characteristics of picture fritter; Texture cell, for calculating the overall textural characteristics of source picture according to the textural characteristics of each picture fritter.

In described device, wavelet transform unit comprises: frequency band processing module, for for each picture fritter, adopts HAAR wavelet function to calculate the detail coefficients of HFS HL/LH/HH; Each picture fritter comprises: low frequency part LL and HFS HL/LH/HH, HFS HH/HL/LH is corresponding level detail, vertical detail and diagonal line details respectively, comprises respectively 2*2 detail coefficients.

In described device, wavelet transform unit also comprises: coefficient of colligation computing module, and for calculating the textural characteristics of each picture fritter $f_{\mathrm{block}}=\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HH}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HL}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{LH},$ Wherein, c _ijit is the detail coefficients after wavelet transformation.

In described device, texture cell comprises: calculate execution module, for calculating f _picbe the overall textural characteristics of source picture, m is the number of picture fritter, f _ithe textural characteristics f of each picture fritter _block.

The beneficial effect of technique scheme of the present invention is as follows: the picture fritter of X*Y size has been taken into account degree of accuracy and the speed of texture feature extraction, source picture is carried out after reasonable piecemeal, first each picture fritter is carried out to wavelet transformation with texture feature extraction, the overall textural characteristics that calculates again source picture, computing velocity is fast and accuracy is high.

Brief description of the drawings

Fig. 1 represents a kind of method flow schematic diagram that calculates picture textural characteristics;

Fig. 2 represents the principle schematic of HAAR wavelet function;

Fig. 3 represents a picture fritter to carry out the schematic diagram of wavelet transformation.

Embodiment

For making the technical problem to be solved in the present invention, technical scheme and advantage clearer, be described in detail below in conjunction with the accompanying drawings and the specific embodiments.

Adopt a width gray level image to describe the principle of two-dimensional wavelet transformation, a 8*8 image block in gray level image is expressed as by matrix A

$\begin{array}{ccccccccc}A& =[64& 2& 3& 61& 60& 6& 7& 57\\ & 9& 55& 54& 12& 13& 51& 50& 16\\ & 17& 47& 46& 20& 21& 43& 42& 24\\ & 40& 26& 27& 37& 36& 30& 31& 33\\ & 32& 34& 35& 29& 28& 38& 39& 25\\ & 41& 23& 22& 44& 45& 19& 18& 48\\ & 49& 15& 14& 52& 53& 11& 10& 56\\ & 8& 58& 59& 5& 4& 62& 63& 1]\end{array}$

An image block is a two-dimentional data array, each line translation of advancing to data array while carrying out wavelet transformation, then each row of the data array after line translation are converted, obtain the data array after conversion, finally the data array after conversion is encoded.

In matrix A, the pixel value of the first row is [64 23 61 60 67 57], adopts average (averaging) to comprise with the process of asking difference (differencing) to process the pixel value of the first row:

Step 1, gets the mean value of every a pair of pixel in the first row, result is placed on to front 4 positions of the first row, on all the other rear 4 positions, is the difference of its corresponding mean value of the first number of every a pair of pixel, and these differences are put into rear 4 positions of the first row.

Step 2, front 4 numbers to the first row are used the method identical with step 1, obtain two mean values, are placed on front 2 positions of the first row, and obtain two differences as detail coefficients (detail coefficient), be placed on third and fourth position of the first row;

Keep the position of remaining 4 detail coefficients not change.

Step 3, in the first row, adopts the method identical with step 1 and step 2, and two mean values that are positioned at front 2 positions are averaged and difference.

Use step 1～step 3 to calculate each row of image block matrix A, obtain matrix A '

$\left[\begin{array}{cccccccc}[32.5& 0& 0.5& 0.5& 31& -29& 27& -25\\ 32.5& 0& -0.5& -0.5& -23& 21& -19& 17\\ 32.5& 0& -0.5& -0.5& -15& 13& -11& 9\\ 32.5& 0& 0.5& 0.5& 7& -5& 3& -1\\ 32.5& 0& 0.5& 0.5& -1& 3& -5& 7\\ 32.5& 0& -0.5& -0.5& 9& -11& 13& -15\\ 32.5& 0& -0.5& -0.5& 17& -19& 21& -23\\ 32.5& 0& 0.5& 0.5& -25& 27& -29& 31]\end{array}\right]$

Matrix A ' in, first element of every a line is the mean value of this row pixel value, remaining is the detail coefficients of this row.

The method providing in step 1～step 3 is provided, each row of A ' are calculated, obtain A ' '

$\begin{array}{cccccccc}[32.5& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 0\\ 0& 0& 0& 0& 4& -4& 4& -4\\ 0& 0& 0& 0& 4& -4& 4& -4\\ 0& 0& 0.5& 0.5& 27& -25& 23& -21\\ 0& 0& -0.5& -0.5& -11& 9& -7& 5\\ 0& 0& 0.5& 0.5& -5& 7& -9& 11\\ 0& 0& -0.5& -0.5& 21& -23& 25& -27]\end{array}$

The mean value of the pixel value of the element representation 8*8 image block in the upper left corner, all the other elements are detail coefficients.

HARR wavelet function is defined in interval [0,1] upper, as shown in Figure 2, be the one in the simplest in conventional wavelet transformation, also be the orthogonal wavelet function with tight support of using the earliest, wherein, [0,1/2) interval value is 1, [1/2,1) interval value is-1, and other interval value is 0, is defined as:

Have in the one dimension initial pictures of 4 pixels at a width, corresponding pixel value or the coefficient of picture position are respectively [9 73 5], and the step of calculating the detail coefficients of its haar wavelet transform comprises:

Step 1, average (averaging), calculates the right mean value of neighbor, obtain the lower new images of a width resolution, corresponding pixel value is [84], and in new images, the number of pixel is 2, resolution is 1/2 of initial pictures, obviously the partial loss of information of image.

Step 2, asks difference (differencing), in order to reconstruct the original image of 4 pixels from the new images of 2 pixel compositions, need to store some detail coefficients, to give the information of loss for change in the time of reconstruct.

First right pixel pixel value is deducted to the mean value that this pixel is right, or uses difference that this pixel is right divided by 2, obtain:

First detail coefficients is (9-8)=1,8th, mean value, and storage detail coefficients 1 can be recovered the first two pixel value of original image.

Second detail coefficients is (3-4)=-1, and rear 2 pixel values can be recovered in storage detail coefficients-1.

Original image represents with [841-1] that comprise two mean values and two detail coefficients.

Step 3, to new images repeating step 1～step 2, obtains image and detail coefficients that resolution is lower.

Finally, adopt mean value 6 and three detail coefficients 2,1 and-1 presentation video of a pixel, [6 21-1] is called a haar wavelet transform coefficient.

The embodiment of the present invention provides a kind of method of calculating picture textural characteristics, as shown in Figure 1, comprising:

The picture fritter that step 101 is at least one X*Y by source picture segmentation;

Step 102, carries out wavelet transformation to each picture fritter, calculates the textural characteristics of picture fritter;

Step 103, calculates the overall textural characteristics of source picture according to the textural characteristics of each picture fritter.

The technology that application provides, the picture fritter of X*Y size has been taken into account degree of accuracy and the speed of texture feature extraction, source picture is carried out after reasonable piecemeal, first each picture fritter is carried out to wavelet transformation with texture feature extraction, the overall textural characteristics that calculates again source picture, computing velocity is fast and accuracy is high.

In a preferred embodiment, before the picture fritter that is at least one X*Y by source picture segmentation, also comprise:

Calculate the gray-scale value of each pixel in the picture of source.

Source picture generally adopts rgb space to represent, therefore the gray-scale value of wavelet transformation based on source images need first to calculate the gray-scale value G of each pixel in the picture of source:

G=R*0.30+G*0.59+B*0.11, R, G, B is respectively red, green, blue triple channel numerical value.

In a preferred embodiment, the picture fritter that is at least one X*Y by source picture segmentation, specifically comprises:

X=4 in each picture fritter, Y=4, the matrix of a corresponding 4*4, the element lacking in matrix mends 0.

Source picture is divided into after picture fritter, to each picture fritter texture feature extraction respectively, picture fritter is too large, the textural characteristics extracting is inaccurate and coarse, the too little calculated amount of picture fritter is waited a moment greatly, can know that the picture fritter that source picture is divided into 4*4 is to take into account processing speed and precision according to practical experience balance after both.

Be divided in the process of 4*4 picture fritter, insufficient section mends 0, specifically adopts HAAR small echo to convert.

In a preferred embodiment, as shown in Figure 3, each picture fritter is carried out to wavelet transformation, specifically comprises:

Each picture fritter comprises: low frequency part LL and HFS HL/LH/HH, and HFS HH/HL/LH is corresponding level detail, vertical detail and diagonal line details respectively, comprises respectively 2*2 detail coefficients;

For each picture fritter, adopt HAAR wavelet function to calculate the detail coefficients of HFS HL/LH/HH.

As shown in Figure 3, in picture fritter, LL obtains by being expert at and listing to average, and belongs to and changes part slowly, is called low frequency part, 4 elements in the corresponding upper left corner; HL/LH/HH is that the difference by constantly calculating element and mean value obtains, and relatively belongs to the part that rate of change is large, is called HFS, respectively the element in the corresponding lower left corner, the upper right corner and the lower right corner.

For the picture fritter of 4*4, obtain in the process of LL/HL/LH/HH, first, get the mean value of every a pair of pixel in the first row, result is placed on to front 2 positions of the first row, on all the other rear 2 positions, be the difference of its corresponding mean value of the first number of every a pair of pixel, these differences are put into rear 2 positions of the first row.Afterwards, each row in 4 row are carried out to same calculating, obtain 4*4 matrix after treatment.

In 4*4 matrix, the corresponding low frequency part of LL, the corresponding HFS of HL/LH/HH.

In a preferred embodiment, the textural characteristics that calculates picture fritter specifically comprises:

Textural characteristics $f_{\mathrm{block}}=\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HH}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HL}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{LH}$

Wherein, c _ijit is the detail coefficients after wavelet transformation.

In HFS, HL/LH/HH correspondence three frequency bands, and each frequency band comprises 2*2=4 detail coefficients.What LH calculated is the detail coefficients of row, is level detail, and what HL calculated is the difference of row (vertical direction), is vertical detail, and HH has characterized diagonal, is diagonal line details.

Adopt the detail coefficients of HAAR wavelet function calculating HFS HL/LH/HH, after HAAR wavelet transformation, the comprehensively textural characteristics f of each picture fritter _blockobtain the overall textural characteristics of source picture.

In a preferred embodiment, calculate the overall textural characteristics of source picture according to the textural characteristics of each picture fritter, comprising:

F _picbe the overall textural characteristics of source picture, m is the number of picture fritter, f _ithe textural characteristics f of each picture fritter _block.

Source picture is carried out to reasonable piecemeal, rule of thumb can know that the picture fritter of 4*4 size taken into account precision and speed, use HAAR small echo to extract texture to each picture fritter, then obtain the overall textural characteristics of source picture according to the textural characteristics of each picture fritter, this entirety textural characteristics has characterized the spectral change of source picture, the i.e. spatial color distribution of source picture and light distribution.

The embodiment of the present invention provides a kind of picture processing device, comprising:

Divide module unit, for the picture fritter that is at least one X*Y by source picture segmentation;

Wavelet transform unit, for each picture fritter is carried out to wavelet transformation, calculates the textural characteristics of picture fritter;

Texture cell, for calculating the overall textural characteristics of source picture according to the textural characteristics of each picture fritter.

In a preferred embodiment, wavelet transform unit comprises:

Frequency band processing module, for for each picture fritter, adopts HAAR wavelet function to calculate the detail coefficients of HFS HL/LH/HH;

Each picture fritter comprises: low frequency part LL and HFS HL/LH/HH, HFS HH/HL/LH is corresponding level detail, vertical detail and diagonal line details respectively, comprises respectively 2*2 detail coefficients.

In a preferred embodiment, wavelet transform unit also comprises:

Coefficient of colligation computing module, for calculating the textural characteristics of each picture fritter

$f_{\mathrm{block}}=\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HH}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HL}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{LH},$ Wherein, cij is the detail coefficients after wavelet transformation.

In a preferred embodiment, texture cell comprises:

Calculate execution module, for calculating

F _picbe the overall textural characteristics of source picture, m is the number of picture fritter, f _ithe textural characteristics f of each picture fritter _block.

Adopt this programme advantage to be afterwards: the picture fritter that rule of thumb can know 4*4 has been taken into account degree of accuracy and speed that texture extracts, source picture is carried out reasonable piecemeal and is obtained the picture fritter of 4*4, use HAAR small echo to each picture fritter texture feature extraction, then obtain the overall textural characteristics of source picture according to the textural characteristics of each picture fritter, extract that the precision of texture is high and computing velocity is fast.

The above is the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, do not departing under the prerequisite of principle of the present invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (10)

1. a method of calculating picture textural characteristics, is characterized in that, method comprises:

The picture fritter that is at least one X*Y by source picture segmentation;

Each picture fritter is carried out to wavelet transformation, calculate the textural characteristics of picture fritter;

Calculate the overall textural characteristics of source picture according to the textural characteristics of each picture fritter.

2. method according to claim 1, is characterized in that, before the picture fritter that is at least one X*Y, also comprises source picture segmentation:

Calculate the gray-scale value of each pixel in the picture of source.

3. method according to claim 1, is characterized in that, the picture fritter that is at least one X*Y by source picture segmentation, specifically comprises:

X=4 in each picture fritter, Y=4, the matrix of a corresponding 4*4, the element lacking in matrix mends 0.

4. method according to claim 1, is characterized in that, each picture fritter is carried out to wavelet transformation, specifically comprises:

Each picture fritter comprises: low frequency part LL and HFS HL/LH/HH, and HFS HH/HL/LH is corresponding level detail, vertical detail and diagonal line details respectively, comprises respectively 2*2 detail coefficients;

For each picture fritter, adopt HAAR wavelet function to calculate the detail coefficients of HFS HL/LH/HH.

5. method according to claim 4, is characterized in that, the textural characteristics that calculates picture fritter specifically comprises:

Textural characteristics $f_{\mathrm{block}}=\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HH}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HL}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{LH}$

Wherein, c _ijit is the detail coefficients after wavelet transformation.

6. method according to claim 5, is characterized in that, calculates the overall textural characteristics of source picture according to the textural characteristics of each picture fritter, comprising:

$f_{\mathrm{pic}}=\left(\underset{i=0}{\overset{m-1}{\mathrm{\Σ}}}{f}_i\right)/m$

F _picbe the overall textural characteristics of source picture, m is the number of picture fritter, f _ithe textural characteristics f of each picture fritter _block.

7. a picture processing device, is characterized in that, comprising:

Divide module unit, for the picture fritter that is at least one X*Y by source picture segmentation;

Wavelet transform unit, for each picture fritter is carried out to wavelet transformation, calculates the textural characteristics of picture fritter;

Texture cell, for calculating the overall textural characteristics of source picture according to the textural characteristics of each picture fritter.

8. device according to claim 7, is characterized in that, wavelet transform unit comprises:

Frequency band processing module, for for each picture fritter, adopts HAAR wavelet function to calculate the detail coefficients of HFS HL/LH/HH;

Each picture fritter comprises: low frequency part LL and HFS HL/LH/HH, HFS HH/HL/LH is corresponding level detail, vertical detail and diagonal line details respectively, comprises respectively 2*2 detail coefficients.

9. device according to claim 7, is characterized in that, wavelet transform unit also comprises:

Coefficient of colligation computing module, for calculating the textural characteristics of each picture fritter

$f_{\mathrm{block}}=\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HH}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{HL}+\left(\sqrt{\underset{i=0}{\overset{1}{\mathrm{\Σ}}}\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}^2}\right)\mathrm{LH},$ Wherein, c _ijit is the detail coefficients after wavelet transformation.

10. device according to claim 7, is characterized in that, texture cell comprises:

Calculate execution module, for calculating

F _picbe the overall textural characteristics of source picture, m is the number of picture fritter, f _ithe textural characteristics f of each picture fritter _block.