CN105513048A - Sub-band-information-entropy-measure-based image quality evaluation method - Google Patents

Abstract

The invention discloses a sub-band-information-entropy-measure-based image quality evaluation method. With the method, a problem that evaluation of a noise image by a computer does not conform to the perception of the human eyes can be solved. The method comprises: step one, block discrete cosine transform is carried out on a to-be-tested image; step two, similarity frequency band coefficients of all sub block transform coefficients are combined; step three, the combined coefficients are reconstructed and the identical frequency band coefficients form reconstruction sets; step four, an information entropy of each reconstruction set coefficient is calculated; step five, the operations from the step to step four are carried out on a reference image corresponding to the to-be-tested image to obtain an information entropy of the reference image; step six, differences of information contents on all frequency bands of the to-be-tested image and the reference image are calculated; step seven, a weighted sum of all frequency band information content differences is calculated to obtain a quality value of the to-be-tested image; and step eight, according to the quality value, quality evaluation of the to-be-tested image is obtained. According to the invention, the obtained evaluation result conforms to the perception of the human eyes. The method can be used for identifying image transmission and image retrieving on the internet and measuring the quality of the image processing index.

Description

Based on the image quality evaluating method of sub-band information entropy tolerance

Technical field

The invention belongs to technical field of image processing, particularly a kind of image quality evaluating method, can be used for differentiating the quality of the transmission of internet epigraph, image retrieval and measurement image procossing index.

Technical background

Internet era, the exchange between information and circulation very general, digital picture and digital video, due to its expression form intuitively, become the main carriers of information gradually.Along with the lifting of digital image processing techniques, image plays an important role in increasing application scenario.In the communications, sending a width high-definition image has very high requirement to bandwidth sum code stream length, and due to restriction and the saving code stream of channel width in reality, image is sent by transmission channel after usually being compressed again.But, in compression and transmitting procedure, inevitably introduce noise.Get by the image of noise pollution at receiving end, the comfort level of Human Perception can be reduced, even can affect the correct understanding of people to picture material.Therefore, the picture quality obtained to allow receiving end can meet the specific demand of people, and guide the compressibility setting of transmitting terminal to image and the control of channel width, the quality of vision facilities energy automatic Evaluation one amplitude and noise acoustic image is highly significant, and this impels people to design the image quality evaluation algorithm consistent with human-eye visual characteristic.

In the past few decades, image quality evaluating method achieves larger progress, and a large amount of evaluation algorithms is suggested.According to the degree of dependence to reference picture, these algorithms can be roughly divided into three classes: full reference image quality appraisement algorithm, partial reference image quality appraisement algorithm and non-reference picture quality appraisement algorithm.Full reference mass evaluation algorithms needs the full detail of original image, the Y-PSNR PSNR be such as widely used.In a practical situation, we can not get all information of original image sometimes, and therefore the application of full reference mass evaluation algorithms is restricted.Because reference-free quality evaluation algorithm does not need any information of original image, so attracted the concern of researchist.But existing non-reference picture quality appraisement algorithm can not show reliable performance on large-scale database.As the compromise of full reference and reference-free quality evaluation algorithm, part reference mass evaluation algorithms is suggested owing to only needing a small part information of reference picture.

As everyone knows, the performance quality of part reference mass evaluation algorithms depends primarily on the validity of characteristics of image.A good feature, should be able to the complete information of overview diagram picture, and to different noise-sensitive.At present, the method extracting characteristics of image is in the transform domain as illustrated widely used in part reference mass evaluation algorithms, and shows good performance.These class methods have a common structure, namely first do corresponding conversion to image, and such as conventional wavelet transformation, discrete cosine transform etc., the conversion coefficient then for image processes.

The people such as Rajiv article " Rredindices:reducedreferenceentropicdifferencingforimage qualityassessment; " IEEETransactionsonImageProcessing, vol.21, no.2, pp.517-526, points out in 2012, and the wavelet coefficient of image well can obey Gaussian Mixture distribution, therefore adopt Gaussian Mixture Distribution Model to image wavelet coefficient modeling, propose a kind of quality evaluation algorithm based on wavelet transformation.The people such as Ma article " Reduce-referenceimagequalityassessmentusingreorganizeddc t-basedimagerepresentation; " IEEETransactionsonMultimedia, vol.13, no.4, pp.824-829, point out in 2011, because discrete cosine transform has very strong energy accumulating and quick computational algorithm, utilize Generalized Gaussian Distribution Model to the modeling of image discrete cosine transform coefficient, model parameter by as characteristics of image, carries out quality assessment by the feature difference calculating image to be tested and its reference picture.

Although the method for carrying out modeling for conversion coefficient is in the transform domain as illustrated widely used, the performance of these algorithms sharply declines along with the minimizing of characteristic number.In addition, when noise intensity is larger, the conversion coefficient of contaminated image may not above-mentioned the carried Gaussian Mixture Distribution Model of matching preferably or Generalized Gaussian Distribution Model.Therefore, between the model and real coefficient distribution of matching, there is error of fitting, this will affect the accuracy of quality assessment.

Summary of the invention

The object of the invention is the deficiency for existing in above-mentioned prior art, proposes a kind of image quality evaluating method based on sub-band information entropy tolerance, to utilize a small amount of information obtained from reference picture, estimates the mass value of noise image accurately.

Technical scheme of the present invention is achieved in that

The present invention calculates image to be tested and the quantity of information of reference picture on each subband thereof respectively in discrete cosine transform domain, by the decay asking poor mode to obtain quantity of information on each subband of image to be tested, again summation is weighted to the decay of quantity of information on each subband, finally obtain the mass value of image to be tested, implementation step comprises as follows:

(1) input the image I to be tested that size is M × N, undertaken, without aliasing piecemeal, then doing discrete cosine transform to each sub-block respectively by B × B to this image I, obtain the transform coefficient matrix C of each sub-block _l, wherein B=8, l represent from top left to bottom right to the index of each sub-block in image I to be tested, expression is not more than maximum integer;

(2) to each matrix of coefficients C _lin coefficient that often group has a similar frequency band ask two norms respectively, make C _lin original B ²individual coefficient becomes 2B-1 new coefficient after merging, and forms new coefficient set C with these new coefficients _l';

(3) from new coefficient set C _l' in take out frequency band successively and be coefficient, obtain recombination coefficient S set _i, i=0,1 ..., 2B-2;

(4) each recombination coefficient S set is calculated respectively _ithe information entropy e of middle coefficient _i, the value of each information entropy corresponds to the quantity of information of image I to be tested in frequency band;

(5) the reference picture U corresponding to image I to be tested is carried out the operation of above-mentioned steps (1)-(4), obtain the quantity of information r of reference picture U on each frequency band _i;

(6) the poor d of image I to be tested and reference picture U quantity of information on each frequency band is calculated _i;

(7) by the poor d of quantity of information on each frequency band _ibe weighted summation, be the mass value Q of image I to be tested;

(8) quality treating test pattern I according to mass value Q judges:

If Q=0, then represent that this test pattern I is not by noise pollution;

If 0 < Q≤5, then represent that this test pattern I is by noise slight pollution;

If 5 < Q≤8, then represent that this test pattern I is by noise intermediate pollution;

If Q > 8, then represent that this test pattern I is by noise serious pollution.

Compared with prior art, tool has the following advantages in the present invention:

1) the partial reference image quality appraisement algorithm of the present invention's proposition, only utilize the feature of 15 reference pictures just can treat test pattern and carry out quality assessment exactly, significantly reduce the degree of dependence to reference picture, decrease the cost to reference picture characteristic transmission.

2) the present invention extracts characteristics of image by the information entropy of computed image conversion coefficient on each frequency band, avoids and carries out modeling to conversion coefficient and the error of fitting introduced, improve evaluation precision.

3) the present invention has taken into full account the visual characteristic of human eye system to image different frequency bands information, by arranging to the distortion of quantity of information on each frequency band the mass value that different weights evaluates image to be tested, makes evaluation result consistent with Human Perception.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention.

Embodiment

With reference to Fig. 1, specific implementation step of the present invention is as follows:

Step 1, input size is the image I to be tested of M × N, to go forward side by side the conversion of line frequency territory, obtain the transform coefficient matrix C of each sub-block to its piecemeal _l.

Conventional image frequency domain transform method has: Fourier transform, discrete cosine transform, wavelet transformation, Walsh transform etc., because discrete cosine transform has very strong energy accumulating and the fast algorithm of computing, this example adopts discrete cosine transform, and its step is as follows:

(1a) input the image I to be tested that a width size is M × N, being undertaken without aliasing piecemeal by B × B, is P to each sub block number _l, wherein B=8, l represent from top left to bottom right to the call number of each sub-block in image I to be tested, expression is not more than maximum integer;

(1b) each sub-block P in test pattern I is treated respectively _ldo discrete cosine transform, obtain each sub-block P _ltransform coefficient matrix C _l, the coefficient C in this transform coefficient matrix _l(u, v) is:

$C_l(u,v)=\sqrt{\frac{2}{M}}\sqrt{\frac{2}{N}}\underset{x=0}{\overset{M-1}{\&Sigma;}}\underset{y=0}{\overset{N-1}{\&Sigma;}}cos\&lsqb;\frac{u\mathrm{\&pi;}}{2M}\left(2x+1\right)\&rsqb;cos\&lsqb;\frac{v\mathrm{\&pi;}}{2N}\left(2y+1\right)\&rsqb;P_l(x,y)$

Wherein, (x, y) is sub-block P _lthe coordinate of middle pixel, (u, v) is transform coefficient matrix C _lthe coordinate of middle coefficient.

Step 2, calculates each matrix of coefficients C respectively _lin often group there are two norms of similar frequency band coefficient, form new coefficient set C by these two norms _l'.

(2a) according to the basis function characteristic of two-dimension discrete cosine transform, each transform coefficient matrix C _lin coefficient on each counter-diagonal direction be considered to have similar frequency band, by transform coefficient matrix C _lin the set of coefficient on i-th counter-diagonal be expressed as:

${\mathrm{\&Psi;}}_l^i=\left\{\left(u,v\right)\right|u+v=i,0\&le;u<B,0\&le;v<B\}$

(2b) in order to reduce the number of characteristics of image, to transform coefficient matrix C _lin there is similar frequency band coefficient merged by the mode of calculating two norm:

$c_l^i=\sqrt{\frac{1}{\left|{\mathrm{\&Psi;}}_l^i\right|}\underset{u,v\&Element;{\mathrm{\&Psi;}}_l^i}{\&Sigma;}{C}_l{(u,v)}^2},i=0,1,...2B-2$

Wherein, represent transform coefficient matrix C _lin coefficient on i-th counter-diagonal merge by asking two norms the new coefficient obtained, its frequency band is represent set the number of middle coefficient;

(2c) each matrix of coefficients C is utilized respectively _lnew coefficient after middle merging form new coefficient set C _l':

$C_l^{\&prime;}=\left\{c_l^i\right\},i=0,1,...2B-2.$

Step 3, from all new coefficient set C _l' in take out frequency band successively and be coefficient, obtain each recombination coefficient S set _i:

Step 4, calculates each recombination coefficient S set respectively _ithe information entropy e of middle coefficient _i.

(4a) respectively by each recombination coefficient S set _iin coefficient be divided into K group uniformly according to its numerical range calculate recombination coefficient S set _imiddle jth group the number of middle coefficient accounts for this recombination coefficient S set _ithe ratio of middle overall coefficient number

$p_i^j=\frac{\left|{\mathrm{\&Omega;}}_i^j\right|}{\left|S_i\right|}$

Wherein, 1≤j≤K, K=10, represent recombination coefficient S set _ijth group the number of middle coefficient, | S _i| represent restructuring S set _ithe number of middle coefficient;

(4b) according to the formula of Shannon entropy, pass through calculate each recombination coefficient S set _iinformation entropy e _i:

$e_i=-\underset{j=1}{\overset{K}{\&Sigma;}}{p}_i^j{\log }_2{p}_i^j$

This information entropy e _irepresent the corresponding frequency band of image I to be tested the quantity of information had.

Step 5, carries out the operation of above-mentioned steps (1)-(4) by the reference picture U corresponding to image I to be tested, obtain the quantity of information r of corresponding i-th frequency band of reference picture U _i.

Step 6, calculates the poor d of image I to be tested and reference picture U quantity of information on each frequency band _i:

d _i＝e _i-r _i。

Step 7, by the poor d of quantity of information on each frequency band _ibe weighted summation, obtain the mass value Q of image I to be tested:

$Q={\log }_2(1+\frac{1}{D_0}\underset{i=0}{\overset{2B-2}{\&Sigma;}}{w}_i{d}_i),$

Wherein D ₀=0.1, w _ifor weight, this weight size sets the visual characteristic of image different frequency bands information according to human eye system, because human eye is more responsive than high frequency band to the information change on lower band, therefore the weight of the higher setting of frequency band is less, and little on quality evaluation result impact higher than information distortion on the frequency band of intermediate frequency, so $w_i=\left\{\begin{array}{cc}{\left(8-i\right)}^2,& 0\&le;i<8\\ 1,& 8\&le;i<15\end{array}\right..$

Step 8, judges the quality of picture quality according to mass value Q.

The quality evaluation result standard weighing image I to be tested is: Q is less, and quality is better; Q is larger, and quality is poorer:

If Q=0, then represent that this test pattern I is not by noise pollution;

If 0 < Q≤5, then represent that this test pattern I is by noise slight pollution;

If 5 < Q≤8, then represent that this test pattern I is by noise intermediate pollution;

If Q > 8, then represent that this test pattern I is by noise serious pollution.

More than describing is only example of the present invention, does not form any limitation of the invention.Obviously for those skilled in the art; after having understood content of the present invention and principle; all may when not deviating from the principle of the invention, structure; carry out the various amendment in form and details and change, but these corrections based on inventive concept and change are still within claims of the present invention.

Claims (4)

1., based on an image quality evaluating method for sub-band information entropy tolerance, comprise the steps:

(1) input the image I to be tested that size is M × N, undertaken, without aliasing piecemeal, then doing discrete cosine transform to each sub-block respectively by B × B to this image I, obtain the transform coefficient matrix C of each sub-block _l, wherein B=8, l represent from top left to bottom right to the call number of each sub-block in image I to be tested, expression is not more than maximum integer;

(2) to each matrix of coefficients C _lin coefficient that often group has a similar frequency band ask two norms respectively, make C _lin original B ²individual coefficient becomes 2B-1 new coefficient after merging, and forms new coefficient set C with these new coefficients _l';

(3) from new coefficient set C _l' in take out frequency band successively and be coefficient, obtain recombination coefficient S set _i, i=0,1 ..., 2B-2;

(4) each recombination coefficient S set is calculated respectively _ithe information entropy e of middle coefficient _i, the value of each information entropy corresponds to the quantity of information of image I to be tested in frequency band;

(5) the reference picture U corresponding to image I to be tested is carried out the operation of above-mentioned steps (1)-(4), obtain the quantity of information r of reference picture U on each frequency band _i;

(6) the poor d of image I to be tested and reference picture U quantity of information on each frequency band is calculated _i;

(7) by the poor d of quantity of information on each frequency band _ibe weighted summation, obtain the mass value Q of image I to be tested;

(8) quality treating test pattern I according to mass value Q judges:

If Q=0, then represent that this test pattern I is not by noise pollution;

If 0 < Q≤5, then represent that this test pattern I is by noise slight pollution;

If 5 < Q≤8, then represent that this test pattern I is by noise intermediate pollution;

If Q > 8, then represent that this test pattern I is by noise serious pollution.

2. the method for claim 1, to each matrix of coefficients C in its step (2) _lin coefficient that often group has a similar frequency band ask two norms respectively, obtained by following formulae discovery:

$c_l^i=\sqrt{\frac{1}{\left|{\mathrm{\&Psi;}}_l^i\right|}\underset{u,v\&Element;{\mathrm{\&Psi;}}_l^i}{\&Sigma;}{C}_l{(u,v)}^2},i=0,1,...2B-2$

${\mathrm{\&Psi;}}_l^i=\left\{(u,v)\right|u+v=i,0\&le;u<B,0\&le;v<B\}$

Wherein, represent matrix of coefficients C _lin the new coefficient of coefficient by asking two norms to obtain on i-th counter-diagonal, (u, v) represents matrix of coefficients C _lin the coordinate of each coefficient, represent matrix of coefficients C _lin the set of coefficient on i-th counter-diagonal, represent set the number of middle coefficient.

3. the method for claim 1, calculates each recombination coefficient S set respectively in its step (4) _ithe information entropy e of middle coefficient _i, carry out as follows:

(4a) respectively by each recombination coefficient S set _iin coefficient be divided into K group uniformly according to its numerical range calculate recombination coefficient S set _imiddle jth group the number of middle coefficient accounts for this recombination coefficient S set _ithe ratio of middle overall coefficient number

$p_i^j=\frac{\left|{\mathrm{\&Omega;}}_i^j\right|}{\left|S_i\right|}$

Wherein, 1≤j≤K, K=10, represent recombination coefficient S set _ijth group the number of middle coefficient, | S _i| represent recombination coefficient S set _ithe number of middle coefficient;

(4b) according to ratio calculate each recombination coefficient S set _iinformation entropy e _i:

$e_i=-\underset{j=1}{\overset{K}{\&Sigma;}}{p}_i^j{\log }_2{p}_i^j.$

4. the method for claim 1, described in its step (7) by the poor d of quantity of information on each frequency band _ibe weighted summation, its formula is as follows:

$Q={\log }_2(1+\frac{1}{D_0}\underset{i=0}{\overset{2B-2}{\&Sigma;}}{w}_i{d}_i)$

Wherein, Q is the mass value of image I to be tested, D ₀=0.1, w _ifor weight, $w_i=\left\{\begin{array}{cc}{\left(8-i\right)}^2,& 0\&le;i<8\\ 1,& 8\&le;i<15\end{array}\right..$