CN104376565A - Non-reference image quality evaluation method based on discrete cosine transform and sparse representation - Google Patents

Abstract

The invention discloses a non-reference image quality evaluation method based on discrete cosine transform and sparse representation, which mainly solves the problem of inaccurate evaluation of the non-reference image quality in the prior art. The implementation steps are: input a grayscale image, perform discrete cosine transform on it and extract the natural scene statistical features; extract a series of natural scene statistical features of images with different distortion types and different contents, and construct the original feature by combining the average subjective difference score Dictionary; cluster the original feature dictionary, and adaptively select atoms to form a sparse representation dictionary according to the similarity between the test image features and the various types in the original feature dictionary; use the sparse representation to stretch the test image features in the feature space and calculate the sparse representation coefficients, Combined with the subjective evaluation values in the sparse representation dictionary, the linear weighted sum is obtained to obtain the image quality measure. The invention has better consistency with the subjective evaluation results, and is suitable for quality evaluation of various distortion types of images.

Description

No-reference Image Quality Assessment Method Based on Discrete Cosine Transform and Sparse Representation

technical field

The invention belongs to the field of image processing, relates to objective evaluation of image quality, and can be used for image collection, coding compression, and network transmission.

Background technique

Image is an important way for human beings to obtain information. Image quality indicates the ability of image to provide information to people or equipment, and is directly related to the adequacy and accuracy of the information obtained. However, in the process of image acquisition, processing, transmission and storage, due to various factors, there will inevitably be degradation problems, which brings great difficulties to information acquisition or post-processing of images. Therefore, it is very important to establish an effective image quality evaluation mechanism. For example, it can be used for performance comparison and parameter selection of various algorithms in the process of image denoising and image fusion; in the field of image coding and communication, it can be used to guide the entire image transmission process and evaluate system performance; in addition, in image processing algorithm optimization , Biometric identification and other scientific fields are also of great significance.

Image quality evaluation can be divided into subjective evaluation method and objective evaluation method from the method. The subjective method is to evaluate the image quality based on the subjective perception of the experimenter, while the objective method is to measure the image quality by simulating the perception mechanism of the human visual system based on the quantitative indicators given by the model. Since people are the final recipients of images, subjective quality evaluation is the most reliable evaluation method. At present, the most commonly used subjective methods are the average subjective score method MOS and the difference subjective score method DMOS. However, because the evaluation results are easily affected by subjective factors, it is time-consuming and cannot be realized automatically when the number of images is large, so it is necessary to The research of objective image evaluation method is particularly important.

According to the degree of dependence on reference images during evaluation, objective image quality assessment can be divided into full reference image quality assessment FR-IQA, partial reference image quality assessment RR-IQA and no reference image quality assessment NR-IQA.

The biggest advantage of the FR-IQA method for full-reference image quality evaluation is that it can accurately predict the quality of distorted images. Currently, the commonly used full-reference methods include methods based on pixel error statistics MSE and PSNR, methods based on structural similarity SSIM and methods based on human visual system The method of HVS et al. However, since these methods require complete prior knowledge of the original image and require a large amount of data to be stored and transmitted, which limits their application in many practical fields, partial reference image quality evaluation methods have become one of the hot research topics.

Partial reference image quality assessment RR-IQA method does not need the complete original reference image, but needs to use the feature combination of some reference images to obtain the quality score of the distorted image. Although the accuracy of the quality evaluation method can be guaranteed on the basis of reducing the amount of information to be transmitted, part of the information of the original image still needs to be transmitted. In most practical applications, the information of the original image cannot be obtained at all or the acquisition cost is very high.

No-reference image quality assessment NR-IQA method is a method that directly evaluates the quality of distorted images without any prior information of the original image. Due to the limited understanding of the human visual system and the corresponding brain cognitive processes, the design and implementation of its algorithms are more difficult. At present, there are no reference image quality assessment methods: "Z.M.P.Sazzad, Y.Kawayoke, and Y.Horita, No-reference image quality assessment for jpeg2000based on spatial features, Signal Process" proposed by Sazzad et al. .Image Commun.,vol.23,no.4,pp.257–268,Apr.2008", but this method is only for JPEG2000 compression and is not suitable for evaluating the impact of blur, noise, etc. on images. Moorthy et al proposed a learning-based approach "A.K.Moorthy and A.C.Bovik, A two-step framework for constructing blind image quality indices, IEEE Signal Process. Lett., vol.17, no.5, pp.513–516, May 2010." , this method directly simulates the mapping relationship between image features and quality, and its mathematical model and model parameters need to be trained or manually obtained. Since the mapping relationship cannot be accurately simulated, the final quality evaluation result is not accurate enough. Anush Krishna Moorthy et al. proposed a method based on natural scene statistics "the Distortion Identification-based Image Verity and INtegrity Evaluation (DIIVINE) index". The feature has as many as 88 dimensions, and the feature extraction process takes too long, and the huge feature dimension is not convenient for the training of classification and regression models.

Contents of the invention

The purpose of the present invention is to address the shortcomings of the above-mentioned existing methods, and propose a non-reference image quality evaluation method based on discrete cosine transform and sparse representation, which can fully utilize the extracted data information while reducing the feature dimension of image extraction. The sparse representation method efficiently models image feature information and improves the accuracy of quality evaluation results.

Achieving the technical solution of the object of the present invention comprises the following steps:

(1) Read in a grayscale image I, perform discrete cosine transform on it, and extract a series of natural scene statistical features f associated with subjective perception;

(2) set up the original feature dictionary of the training image;

(2a) Repeat step (1), extract the natural scene statistical features of n training images, and form a feature matrix F, F=[f ₁ , f ₂ ,...,f _i ,...,f _n ], where f _i is The natural scene statistical features of the i-th training image, i=1,2,...,n;

(2b) Integrate the average subjective difference scores of n training images and form a quality vector M, M=[m ₁ ,m ₂ ,...,m _i ,...,m _n ], where m _i is the i-th training Average subjective difference scores of images, i = 1, 2, ..., n;

(2c) Correspondingly combine the quality vector M and the feature matrix F to construct the original feature dictionary D:

$\mathrm{D.}=\left[\begin{array}{c}f\\ m\end{array}\right]=[{\stackrel{\‾}{f}}_1,{\stackrel{\‾}{f}}_2,...,{\stackrel{\‾}{f}}_i,...,{\stackrel{\‾}{f}}_{\mathrm{no}}],$ in ${\stackrel{\‾}{f}}_i=\left[\begin{array}{c}{f}_i\\ m_i\end{array}\right],i=\mathrm{1,2},...,\mathrm{no};$

(3) Use the K-means algorithm to cluster the original feature dictionary D into H categories, and the cluster center of the kth category is C _k , where $C_k=\left[\begin{array}{c}{\mathrm{Fc}}_k\\ {\mathrm{Mike}}_k\end{array}\right],$ Fc _k is the feature clustering center, Mc _k is the quality clustering center, k=1,2,...,H;

(4) read in a test image I', repeat step (1), extract the natural scene statistical feature f' of the test image I';

(5) Calculate the Euclidean distance dis _k between the natural scene statistical feature f' of the test image I' and the feature cluster center Fc _k of the kth class in the original feature dictionary D, and use P _k to represent the test image and the original feature dictionary D The degree of approximation of the kth class, where

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; dis</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; dis</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; ;</span>$

(6) According to the degree of approximation P _k , select the first N _k samples from the kth class of the original feature dictionary D with the smallest Euclidean distance to the feature cluster center Fc _k as the sparse representation atoms of the test image I' Where i=1,2,...,N,k=1,2,...,H, the H-type atoms together constitute a sparse representation dictionary D' for the test image I':

${\mathrm{D.}}^{\&prime;}=[{\mathrm{D.}}_1,{\mathrm{D.}}_2,...,{\mathrm{D.}}_k,...,{\mathrm{D.}}_h]=\left[\begin{array}{c}{f}^{\&prime;}\\ m^{\&prime;}\end{array}\right],$ in ${\mathrm{D.}}_k=[{\stackrel{\&OverBar;}{f}}_1^k,{\stackrel{\&OverBar;}{f}}_2^k,...,{\stackrel{\&OverBar;}{f}}_i^k,...,{\stackrel{\&OverBar;}{f}}_{N_k}^k],{\stackrel{\&OverBar;}{f}}_i^k=\left[\begin{array}{c}{f}_i^k\\ m_i^k\end{array}\right],$ f _i ^k is the natural scene statistical feature of the i-th sample in the k-th class selected from the original feature dictionary D, is the average subjective difference score of the i-th sample in the k-th category, i=1,2,...,N _k , N _k =δ*P _k , k=1,2,...,H, δ is a normal number;

(7) Solve the sparse representation coefficient of the natural scene statistical feature f' of the test image I' under the feature matrix F' according to the method of sparse representation: Where argmin means the assignment of α=[α ¹ ,α ² ,…,α ^k ,…,α ^H ] ^T that will make the objective function λ·||α|| ₁ +||f'-D′α|| ₂ the smallest give α ^* , k=1,2,...,H, i=1,2,...,N _k , R ^L represents L-dimensional real number space, Indicates the 1-norm of vector β=[β ₁ ,β ₂ ,…,β _l ,…,β _L ] ^T , Represents the 2-norm of the vector β, l=1,2,...,L, λ is a normal number used to balance the fidelity term and the regular term, and T represents the transpose operation;

(8) Calculate the quality of the test image I' according to the constructed sparse representation: Q∈[0,100], where is the quality score of the i-th atom f _i ^k of the k-th class in the sparse representation dictionary D', α _k,i represents the natural scene statistical feature f' of the test image I' in the k-th class i-th in the feature matrix F' The representation coefficient of atom f _i ^k , Q is the final quality measure of the test image I';

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

1. The present invention can make full use of the information of the database, and carry out dictionary training and learning on the statistical features of natural scenes of images of different distortion types in the database, thereby being able to evaluate the quality of images of different distortion types. Compared with most of the existing no-reference image quality assessment algorithms that only target specific distortion types, it has a wider scope of application.

2. The present invention utilizes the feature dictionary that has been trained, does not need any original reference image information, and directly evaluates the image quality effectively, which is more convenient than the full-reference image quality evaluation method and the partial reference image quality evaluation method. Wider range.

3. The present invention can adaptively select the corresponding sparse representation dictionary from the trained image feature dictionary for a specific test image. Compared with other existing algorithms, the evaluation result has better subjective consistency.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is the natural scene statistical feature extracted after discrete cosine transform is carried out to grayscale image I among the present invention;

Fig. 3 is the result of K-means clustering to original feature dictionary D in the present invention;

Fig. 4 is a comparison chart of objective evaluation scores and average subjective difference scores of test images using the present invention.

detailed description

Below in conjunction with accompanying drawing, specific implementation steps and effects of the present invention are described in further detail:

With reference to Fig. 1, the realization steps of the present invention are as follows:

Step 1, extracting the natural scene statistical features f _s of the grayscale image I on the first scale;

(1a) Read in a grayscale image I, decompose the grayscale image I into 5*5 overlapping image blocks, perform discrete cosine transform on each image block, and remove the direct current of discrete cosine transform coefficients weight;

(1b) Use the generalized Gaussian distribution model to fit the discrete cosine transform coefficients of each image block after removing the DC component to obtain the shape factor γ of each image block, and take the mean value of the shape factor γ of all image blocks as the first The first element f _s,1,1 of the natural scene statistical feature f _s,1 on a scale, arrange the shape factors of all image blocks from small to large, and take the mean value of the first 10% as the first scale The second element f _s,1,2 of the natural scene statistical feature f _s, 1, where the generalized Gaussian distribution model function for calculating the shape factor γ is: $f\left(x\right|\mathrm{\&alpha;},\mathrm{\&beta;},\mathrm{\&gamma;})=\mathrm{\&alpha;}*e^{-{\left(\mathrm{\&beta;}{(x-\mathrm{\&mu;})}^T{\mathrm{\&Sigma;}}^{-1}(x-\mathrm{\&mu;})\right)}^{\mathrm{\&gamma;}}},\mathrm{\&beta;}=\frac{1}{\mathrm{\&sigma;}}\sqrt{\frac{\mathrm{\&Gamma;}(3/\mathrm{\&gamma;})}{\mathrm{\&Gamma;}(1/\mathrm{\&gamma;})}},$ μ is the mean, σ is the variance, γ is the shape factor, and β is the scale factor;

(1c) Calculate the frequency variation coefficient of each image block Take the frequency variation coefficients of all image blocks The mean value of is used as the third element f _s,1,3 of the natural scene statistical feature f _s,1 on the first scale, and the frequency variation coefficient for all image blocks Arranged in descending order, take the mean value of the top 10% as the fourth element f _s,1,4 of the natural scene statistical features f _s , ₁ on the first scale, where μ _|x| and σ _|x| are the mean and standard deviation of the absolute value of the DC component-removed discrete cosine transform coefficients of each image block, respectively, and || indicates the absolute value;

(1d) According to the order from high frequency to low frequency, the discrete cosine transform coefficients of each image block are divided into 3 frequency bands, and the frequency band energy change rate R _n of each image block is calculated. The specific calculation is as follows:

in Represents the variance of the nth frequency band, n represents each frequency band, E _n represents the energy of the nth frequency band, n=1,2,3, j is a positive integer;

(1e) Take the average value of the frequency band energy change rate R _n of all image blocks as the fifth element f _s,1,5 of the natural scene statistical feature f _s,1 on the first scale, for the frequency bands of all image blocks The energy change rate R _n is arranged in descending order, and the average value of the top 10% is taken as the sixth element f _s,1,6 of the natural scene statistical feature f _s,1 on the first scale;

(1f) Divide the discrete cosine transform coefficients of each image block into subbands along the three directions of 45°, 90°, and 135°, and use the generalized Gaussian distribution function to fit the subbands in each direction respectively to obtain 3 The frequency variation coefficients of the three direction subbands, and calculate the variance of the frequency variation coefficients of the three direction subbands

(1g) Take the variance of the frequency variation coefficients of the three direction subbands of all image blocks The mean value of is taken as the seventh element f _s,1,7 of the natural scene statistical feature f _s,1 on the first scale, and the variance of the frequency variation coefficient of the three direction sub-bands of all image blocks Arranged from largest to smallest, take the mean value of the top 10% as the eighth element f _s,1,8 of the natural scene statistical feature f _s,1 on the first scale;

(1h) Obtain the 8-dimensional natural scene statistical features f _s,1 of the grayscale image I on the first scale:

f _s,1 =[f _s,1,1 ,f _s,1,2 ,...,f _s,1,i ...,f _s,1,8 ], i=1,2,...,8.

Step 2, extracting the natural scene statistical features f of the grayscale image I on multiple scales;

(2a) carry out the downsampling operation to the grayscale image I to obtain the downsampled image I2, repeat step 1 operation to the downsampled image I2, obtain the natural scene statistical feature f _s,2 of the grayscale image I on the second scale;

(2b) Perform the downsampling operation on the downsampled image I2 again to obtain the second downsampled image I3, repeat the step 1 operation on the second downsampled image I3, and obtain the natural scene statistical feature f of the grayscale image I on the third scale _s,3 ;

(2c) Obtain the statistical characteristics of the natural scene of the grayscale image I: f=[f _s,1 ,f _s,2 ,f _s,3 ] ^T , where T represents the transpose operation.

Step 3, extracting the natural scene statistical feature f' of the test image I' according to step 1 and step 2;

Step 4, construct the original feature dictionary D;

(4a) Perform step 1 and step 2 operations on each image in the training set in turn, extract the natural scene statistical features of n training images, and form a feature matrix F, F=[f ₁ ,f ₂ ,…,f _i , ..., f _n ], where f _i is the natural scene statistical feature of the ith secondary training image, i=1,2,...,n;

(4b) Integrate the average subjective difference scores of n training images to form a quality vector M, M=[m ₁ ,m ₂ ,...,m _i ,...,m _n ], where m _i is the i-th training Average subjective difference scores of images, i = 1, 2, ..., n;

(4c) Correspondingly combine the quality vector M and the feature matrix F to construct the original feature dictionary D: $\mathrm{D.}=\left[\begin{array}{c}f\\ m\end{array}\right]=[{\stackrel{\&OverBar;}{f}}_1,{\stackrel{\&OverBar;}{f}}_2,...,{\stackrel{\&OverBar;}{f}}_i,...,{\stackrel{\&OverBar;}{f}}_{\mathrm{no}}],$ in ${\stackrel{\&OverBar;}{f}}_i=\left[\begin{array}{c}{f}_i\\ m_i\end{array}\right],i=\mathrm{1,2},...,\mathrm{no};$

Step 5, select a corresponding sparse representation dictionary in the image feature dictionary for a specific test image;

(5a) Use the K-means algorithm to cluster the original feature dictionary D into H classes, and the cluster center of the kth class in H class is C _k , where $C_k=\left[\begin{array}{c}{\mathrm{Fc}}_k\\ {\mathrm{Mike}}_k\end{array}\right],$ Fc _k is the feature clustering center, Mc _k is the quality clustering center, k=1,2...H;

(5b) read in a pair of test image I', perform step 1 and step 2 operations in sequence, and extract the natural scene statistical feature f' of the test image I';

(5c) Calculate the Euclidean distance dis _k between the natural scene statistical feature f' of the test image I' and the feature cluster center Fc _k of the kth class in the original feature dictionary D, and obtain the test image I' and the kth class in the original feature dictionary D Class approximation P _k , where

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; dis</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; dis</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; ;</span>$

(5d) According to the degree of approximation P _k , select the first N _k samples with the smallest Euclidean distance from the feature cluster center Fc _k from the kth class of the original feature dictionary D as the sparse representation atoms of the test image I' Where i=1,2,...,N,k=1,2,...,H, the H-type atoms together constitute a sparse representation dictionary D' for the test image I':

${\mathrm{D.}}^{\&prime;}=[{\mathrm{D.}}_1,{\mathrm{D.}}_2,...,{\mathrm{D.}}_k,...,{\mathrm{D.}}_h]=\left[\begin{array}{c}{f}^{\&prime;}\\ m^{\&prime;}\end{array}\right],$ in ${\mathrm{D.}}_k=[{\stackrel{\&OverBar;}{f}}_1^k,{\stackrel{\&OverBar;}{f}}_2^k,...,{\stackrel{\&OverBar;}{f}}_i^k,...,{\stackrel{\&OverBar;}{f}}_{N_k}^k],{\stackrel{\&OverBar;}{f}}_i^k=\left[\begin{array}{c}{f}_i^k\\ m_i^k\end{array}\right],$ f _i ^k is the natural scene statistical feature of the _i -th sample in the k-th class selected from the original feature dictionary D, is the average subjective difference score of the i-th sample in the k-th class, i=1,2,...,N _k , N _k =δ*P _k , k=1,2,...,H, δ is a normal number.

Step 6, sparsely represent the features of the test image;

Solve the sparse representation coefficient of the natural scene statistical feature f' of the test image I' under the feature matrix F' according to the sparse representation method: Where arg min means α=[α ¹ ,α ² ,…,α ^k ,…,α ^H ] ^T that will make the objective function λ·||α|| ₁ +||f'-D′α|| ₂ the smallest assigned to α ^* , k=1,2,...,H, i=1,2,...,N _k , R ^L represents L-dimensional real number space, Indicates the 1-norm of vector β=[β ₁ ,β ₂ ,…,β _l ,…,β _L ] ^T , Indicates the 2-norm of the vector β, λ is a normal number used to balance the fidelity term and the regular term, and T represents the transpose operation.

Step 7, use the following formula to calculate the quality evaluation measure of the test image I':

Q∈[0,100], where is the quality score of the i-th atom f _i ^k of the k-th class in the sparse representation dictionary D', α _k,i represents the natural scene statistical feature f' of the test image I' in the k-th class i-th in the feature matrix F' The representation coefficient of atom f _i ^k , Q is the final quality measure of the test image I', the smaller the Q value, the higher the quality of the test image.

Effect of the present invention can be further illustrated by following experiments:

1. Experimental conditions and scoring criteria:

This experiment was carried out on the second-generation LIVE image quality assessment database of the University of Texas in the United States, which contains 29 high-resolution undistorted RGB color images and corresponding five types of distorted images, including 175 JPEG images, 169 JPEG2000 images, 145 white noise WN images, 145 Gaussian blurred Gblur images and 145 distorted images after fast fading FF channel. The database gives the distorted image an average subjective difference score describing the quality of the distorted image.

In order to test the consistency of the objective evaluation results of the image quality proposed by the present invention and the subjective perception, this experiment has selected the following three kinds of metrics: the first is the linear correlation coefficient LCC, which reflects the accuracy of the objective evaluation method prediction; the second is the root mean square The error RMSE reflects the error of the objective evaluation method; the third is the Spearman rank order correlation coefficient SROCC, which reflects the monotonicity of the objective evaluation result prediction.

In the experiment, the image database is divided into a training image set and a test image set. The training image set is used to train the original feature dictionary, and the test image is used to predict the evaluation results. According to the different number of original images, there are two grouping methods. The first group: the training set has 15 original images; the second group: the training set has 23 original images.

Experimental parameter settings: 1) H is 4 when the original feature dictionary is clustered; 2) δ is 4 when forming a sparse representation dictionary; 3) λ is 0.0001 based on experience when using sparse representation to solve the sparse representation coefficient.

2. Experimental content and results:

Experiment 1: Evaluation of image distortion on the LIVE_database2 database

Utilize the non-reference image quality evaluation method based on discrete cosine transform and sparse representation proposed by the present invention, use two grouping methods to conduct quality evaluation on the test images in the LIVE_database2 database, and calculate the 3 of the image quality objective evaluation results and subjective perception consistency respectively There are two metrics: Linear Correlation Coefficient LCC, Spearman Rank-Order Correlation Coefficient SROCC and Root Mean Square Error RMSE.

Table 1 shows the distortion evaluation on the LIVE_database2 database. Wherein DCTSR1 is the test result that the present invention uses the first grouping mode, DCTSR2 is the test result that the present invention uses the second grouping mode, and wavelet transform and sparse representation combination method 1 is that the wavelet transform and sparse representation combination method uses the first grouping The test results of the method, combined with wavelet transform and sparse representation Method 2 is the test result of the combined method of wavelet transform and sparse representation using the second grouping method.

Table 1 Comparison test results between this method and other image quality evaluation methods

As can be seen from Table 1, when using the same grouping method, the present invention has good advantages compared with most of the existing no-reference methods: 1) higher prediction accuracy is arranged, that is, the linear correlation coefficient LCC is higher than the existing Most of the methods are large; 2) the monotonicity of the prediction is stricter, that is, the rank order correlation coefficient SROCC is larger than most of the existing methods; 3) the present invention can predict images of various distortion types well.

Experiment 2: Consistency experiment between objective image quality evaluation results and subjective perception

According to the quality evaluation measurement of the test image obtained in Experiment 1, draw the scatter relationship diagram of the present invention to the quality evaluation measurement Q of the test image and the average subjective difference score DMOS of the test image, and use logarithmic function fitting to obtain the best Matching curve, as shown in Figure 4. In Fig. 4, the abscissa represents the quality evaluation measure of the image, the ordinate represents the average subjective difference score DMOS of the image, '·' represents the image quality evaluation measure obtained by the present invention, the distribution of '·' and the best matching curve fitted The closer it is, the better the effect. in:

4 (a) is the effect of the present invention in JPEG2000 compression distortion image sub-library,

4 (b) is the effect of the present invention in the JPEG compression distortion image sub-library,

4(c) is the effect of the present invention in the white noise WN distorted image sub-library,

4(d) is the effect of the present invention in the Gaussian blur Gblur distortion image sub-library,

4(e) is the effect of the present invention in the distorted image sub-library after passing through the fast fading FF channel,

4(f) is the effect of the present invention in the entire LIVE_database2 database.

It can be seen from Fig. 4 that the distribution of '·' in the present invention is relatively close to the fitted best matching curve, with less deviation from the curve, and the performance is very stable in different distorted image sub-libraries of the whole database, which is consistent with subjective visual perception The consistency is high.

Claims (2)

1., based on a non-reference picture quality appraisement method for discrete cosine transform and rarefaction representation, comprise the steps:

(1) read in a width gray level image I, discrete cosine transform is carried out to it, extract a series of natural scene statistical nature f be associated with subjective perception;

(2) the primitive character dictionary of training image is set up;

(2a) repeat step (1), extract the natural scene statistical nature of the secondary training image of n, and constitutive characteristic matrix F, F=[f ₁, f ₂..., f _i..., f _n], wherein f _ibe the natural scene statistical nature of the n-th secondary training image, i=1,2 ..., n;

(2b) integrate the mean subjective discrepancy score of the secondary training image of n, and form quality vector M, M=[m ₁, m ₂..., m _i..., m _n], wherein m _ibe the mean subjective discrepancy score of the n-th secondary training image, i=1,2 ..., n;

(2c) quality vector M and eigenmatrix F is carried out correspondence to combine, builds primitive character dictionary D: $D=\left[\begin{array}{c}F\\ M\end{array}\right]=[\stackrel{\&OverBar;}{f_1},\stackrel{\&OverBar;}{f_2},...,\stackrel{\&OverBar;}{f_i},...,\stackrel{\&OverBar;}{f_n}],$ Wherein $\stackrel{\&OverBar;}{f_i}=\left[\begin{array}{c}{f}_i\\ m_i\end{array}\right],i=\mathrm{1,2},...,n;$

(2d) gather for H class with K-means algorithm by primitive character dictionary D, in H class, kth class cluster centre is C _k, wherein $C_k=\left[\begin{array}{c}{\mathrm{Fc}}_k\\ {\mathrm{Mc}}_k\end{array}\right],$ Fc _kfor feature clustering center, Mc _kfor clustering quality center, k=1,2...H.

(3) read in a secondary test pattern I', repeat step (1), extract the natural scene statistical nature f' of test pattern I';

(4) the feature clustering center Fc of kth class in the natural scene statistical nature f' of test pattern I' and primitive character dictionary D is calculated _keuclidean distance dis _k, obtain the degree of approximation P of kth class in test pattern I' and primitive character dictionary D _k, wherein $P_k=\frac{1/{\mathrm{dis}}_k}{{\mathrm{\&Sigma;}}_{k=1}^{H}1/{\mathrm{dis}}_k},$ k＝1,2,...,H；

(5) according to degree of approximation P _k, select and feature clustering center Fc from the kth class of primitive character dictionary D _kthe minimum front N of Euclidean distance _kindividual sample is as the rarefaction representation atom of test pattern I' wherein i=1,2 ..., N, k=1,2 ..., H, H class atom forms the rarefaction representation dictionary D' for test pattern I' jointly: $D^{\&prime;}=[D_1,D_2,...,D_k,...,D_H]=\left[\begin{array}{c}{F}^{\&prime;}\\ M^{\&prime;}\end{array}\right],$ Wherein $D_k=[{\stackrel{\&OverBar;}{f}}_1^k,{\stackrel{\&OverBar;}{f}}_2^k,...,{\stackrel{\&OverBar;}{f}}_i^k,...,{\stackrel{\&OverBar;}{f}}_{N_k}^k],{\stackrel{\&OverBar;}{f}}_i^k=\left[\begin{array}{c}{f}_i^k\\ m_i^k\end{array}\right],$ for the natural scene statistical nature of i-th sample in the kth class selected from primitive character dictionary D, for the mean subjective discrepancy score of the sample of i-th in kth class, i=1,2 ..., N _k, N _k=δ * P _k, k=1,2 ..., H, δ are normal number;

(6) the rarefaction representation coefficient of natural scene statistical nature f' under eigenmatrix F' of test pattern I' is solved according to the method for rarefaction representation: wherein argmin represents and will make objective function λ || α || ₁+ || f'-D ' α || ₂minimum α=[α ¹, α ²..., α ^k..., α ^h] ^tassignment is to α ^*, k=1,2 ..., H, i=1,2 ..., N _k, R ^lrepresent that L ties up real number space, $L=\underset{k=1}{\overset{H}{\mathrm{\&Sigma;}}}{N}_k,{\left|\right|\mathrm{\&beta;}\left|\right|}_1=\underset{l=1}{\overset{L}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&beta;}}_l\right|$ Represent vectorial β=[β ₁, β ₂..., β _l..., β _l] ^t1 norm, ${\left|\right|\mathrm{\&beta;}\left|\right|}_2=\sqrt{\underset{l=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_l^2}$ Represent 2 norms of vectorial β, λ is the normal number for balancing fidelity item and regular terms, and T represents matrix transpose operation;

(7) quality of test pattern I' is calculated according to constructed rarefaction representation: q ∈ [0,100], wherein for kth class i-th atom in rarefaction representation dictionary D' massfraction, α _k,irepresent natural scene statistical nature f' kth class i-th atom in eigenmatrix F' of test pattern I' expression coefficient, Q is that the final mass of test pattern I' is estimated.

2. the non-reference picture quality appraisement method based on discrete cosine transform and rarefaction representation according to claim 1, wherein described in step (1), discrete cosine transform is carried out to gray level image I, extract a series of natural scene statistical nature f be associated with subjective perception, concrete steps are as follows:

(1.1) the natural scene statistical nature f of gray level image I on first yardstick is extracted _{s, 1};

(1.1a) what gray level image I is decomposed into 5*5 has overlapping image block each other, carries out discrete cosine transform, and remove the DC component of discrete cosine transform coefficient to each image block;

(1.1b) respectively Generalized Gaussian Distribution Model matching is used to the discrete cosine transform coefficient after the removal DC component of each image block, obtain the form factor γ of each image block, get the average of the form factor γ of all image blocks as the natural scene statistical nature f on first yardstick _{s, 1}first element f _{s, 1,1}, to the form factor of all image blocks according to arranging from small to large, get the average of front 10% as the natural scene statistical nature f on first yardstick _{s, 1}second element f _{s, 1,2}, the Generalized Gaussian Distribution Model function wherein calculating form factor γ is: $f\left(x\right|\mathrm{\&alpha;},\mathrm{\&beta;},\mathrm{\&gamma;})=\mathrm{\&alpha;}*e^{{-\left(\mathrm{\&beta;}{(x-\mathrm{\&mu;})}^T{\mathrm{\&Sigma;}}^{-1}(x-\mathrm{\&mu;})\right)}^{\mathrm{\&gamma;}}},\mathrm{\&beta;}=\frac{1}{\mathrm{\&sigma;}}\sqrt{\frac{\mathrm{\&Gamma;}(3/\mathrm{\&gamma;})}{\mathrm{\&Gamma;}(1/\mathrm{\&gamma;})}},\mathrm{\&Gamma;}\left(z\right)={\&Integral;}_0^{\&infin;}{t}^{z-1}{e}^{-t}\mathrm{dt},\mathrm{\&alpha;}=\frac{\mathrm{\&beta;\&gamma;}}{2\mathrm{\&Gamma;}(1/\mathrm{\&gamma;})},$ μ is average, and σ is variance, and γ is form factor, and β is scale factor;

(1.1c) the frequency change coefficient of each image block is calculated get the frequency change coefficient of all image blocks average as the natural scene statistical nature f on first yardstick _{s, 1}the 3rd element f _{s, 1,3}, to the frequency change coefficient of all image blocks according to arranging from big to small, get the average of front 10% as the natural scene statistical nature f on first yardstick _{s, 1}the 4th element f _{s, Isosorbide-5-Nitrae}, wherein μ _{| x|}and σ _{| x|}be respectively average and the standard deviation of the discrete cosine transform coefficient of the removal DC component of each image block, || represent and ask absolute value;

(1.1d) according to the order from high frequency to low frequency, the discrete cosine transform coefficient of each image block is divided into 3 frequency bands, calculates the frequency band energy rate of change R of each image block _n, be specifically calculated as follows formula:

$R_n=\frac{|E_n-\frac{1}{n-1}{\mathrm{\&Sigma;}}_{j<n}{E}_j|}{E_n+\frac{1}{n-1}{\mathrm{\&Sigma;}}_{j<n}{E}_j},$ Wherein $E_n={\mathrm{\&sigma;}}_n^2,$ N=1,2,3, represent the variance of the n-th frequency band, n represents each frequency band, E _nrepresent the energy of the n-th frequency band, j is positive integer;

(1.1e) get all image blocks frequency band energy rate of change R _naverage as the natural scene statistical nature f on first yardstick _{s, 1}the 5th element f _{s, 1,5}, to the frequency band energy rate of change R of all image blocks _naccording to arranging from big to small, get the average of front 10% as the natural scene statistical nature f on first yardstick _{s, 1}the 6th element f _{s, 1,6};

(1.1f) discrete cosine transform coefficient of each image block is divided into along 45 °, 90 °, the subband in 135 ° of these 3 directions, and with generalized Gaussian distribution function respectively to the subband matching in each direction, obtain the frequency change coefficient of 3 directional subbands, and calculate the variance of the frequency change coefficient of these 3 directional subbands

(1.1g) variance of the frequency change coefficient of 3 directional subbands of all image blocks is got average as the natural scene statistical nature f on first yardstick _{s, 1}the 7th element f _{s, 1,7}, to the variance of frequency change coefficient of 3 directional subbands of getting all image blocks according to order arrangement from big to small, get the average of front 10% as the natural scene statistical nature f on first yardstick _{s, 1}the 8th element f _{s, 1,8};

(1.1h) the 8 dimension natural scene statistical nature fs of gray level image I on first yardstick are obtained _{s, 1}:

f _s,1＝[f _s,1,1,f _s,1,2,...,f _s,1,i...,f _s,1,8]，i＝1,2,...,8。

(1.2) down-sampling operation is carried out to gray level image I and obtain down-sampled images I2, step (1.1) operation is repeated to down-sampled images I2, obtains the natural scene statistical nature f of gray level image I on second yardstick _{s, 2};

(1.3) down-sampling operation is carried out again to down-sampled images I2 and obtain secondary down-sampled images I3, step (1.1) operation is repeated to secondary down-sampled images I3, obtains the natural scene statistical nature f of gray level image I on the 3rd yardstick _{s, 3};

(1.4) the natural scene statistical nature of gray level image I is tried to achieve: f=[f _{s, 1}, f _{s, 2}, f _{s, 3}] ^t, T represents matrix transpose operation.