CN102740114B - A kind of nothing ginseng appraisal procedure of Subjective video quality - Google Patents

Abstract

The invention discloses the nothing ginseng appraisal procedure of a kind of Subjective video quality, it includes three below aspect: 1) visuality of the noise comprised in compression video is carried out qualitative assessment, set up contacting between noise visuality and subjective visual quality, it is achieved the evaluation to Subjective video quality；2) video quality focusing on difference utilizes passivation fuzzy evaluation algorithm to carry out qualitative assessment, sets up and focuses on contacting between subjective visual quality, it is achieved the evaluation to Subjective video quality；3) picture quality of different contrast is carried out quantitative predication, set up contacting between contrast and subjective visual quality, it is achieved the evaluation to Subjective video quality.The present invention, in the case of not having original video sequence reference, solves the key technology to the coherence's relation Accurate Model between image pixel, it is achieved the subjective quality of video sequence is carried out printenv assessment.

Description

Non-parameter evaluation method for subjective quality of video

Technical Field

The invention relates to a parameter-free evaluation method for subjective quality of a video, in particular to a parameter-free evaluation method for subjective quality of a video of a monitored video sequence.

Background

The internationally accepted method of assessing video quality that is most reliable is the subjective assessment method, which is particularly important because in most video applications the ultimate recipient of the video is the human eye. At present, when video is subjectively evaluated, the video is manually scored by a plurality of professional evaluators by using professional experience in the industry, and the result of subjective quality of the video is the average value of scores of the plurality of evaluators on the video. Because the scoring is carried out manually, the subjective quality evaluation result of the video is inevitably influenced and interfered by human factors; meanwhile, due to the fact that the number of videos to be evaluated is too large, when an evaluator evaluates each video, the evaluator needs to completely watch the video, and the evaluation time is long.

Disclosure of Invention

The invention aims to provide an innovative and efficient video subjective quality non-reference evaluation system, which can realize evaluation of video subjective quality without human intervention, and is a technical problem to be solved in the field of measuring the subjective quality of a video by establishing a relation between noise visibility, focus, contrast and subjective visual quality according to a human eye visual model technology and priori knowledge. The method and the device can be used for evaluating the subjective quality of the video without reference, so that the interference of human factors on an evaluation result can be reduced, and the evaluation speed and the result fairness are greatly improved.

The technical problem solved by the invention can be realized by adopting the following technical scheme:

a non-parameter evaluation method for subjective quality of video comprises the following three aspects:

1) quantitatively evaluating the visibility of noise contained in the compressed video, establishing a relation between the noise visibility and subjective visual quality, and realizing the evaluation of the subjective quality of the video;

2) carrying out quantitative evaluation on the video quality of different focuses by using a passivation fuzzy evaluation algorithm, establishing a relation between the focuses and subjective visual quality, and realizing the evaluation on the subjective quality of the videos;

3) and quantitatively estimating the image quality with different contrasts, establishing a relation between the contrast and the subjective visual quality, and realizing the evaluation of the subjective quality of the video.

2. The method according to claim 1, wherein the establishing of the relationship between the noise visibility and the subjective visual quality is performed by an image structure similarity calculator (SSIM); the SSIM index includes:

image brightness comparison equation:

whereinμ_xAnd mu_yIs the mean of the images x and y, C1 isA stability constant at near zero;

contrast comparison equation:

wherein sigma_xAnd σ_yVariance of images x and y, C₂Is thatA stability constant at near zero;

structural correlation equations for images x and y:

wherein sigma_xyIs the correlation of the images x and y, C₃Is a stability constant;

the method consists of the following three equations:

SSIM(x，y)＝[l(x，y)]^α·[c(x，y)]^β·[s(x，y)]^γ。

in one embodiment of the invention, the linking between focus and subjective visual quality is achieved by a Spatial sharp Map relationship; a Spectral sharpness map is defined as follows:

$S_1\left(I\right)=1-\frac{1}{1+e^{-3(a_I-2)}};$

the total variation of Spatial domain in Spatial Measure of Sharpness is used to Measure the Sharpness and blur of an image, an image block I_bThe Total Variation of (I) v (I)_b) Can be calculated by the following equation:wherein I_biAnd I_bjIs I_bThe eight fields of (1);

v(I_b) Effectively display image block I_bThe sum of the absolute differences with its offset blocks, from which we define a spatialsharpness map as follows:

$S_2\left(I\right)=\frac{1}{4}{\max }_{\mathrm{\τ}\∈I}v\left(I_b\right);$

with S₃(I) The average of the 1% maximum measures the quality of the image:

$S_{3\_INDEX}=\frac{1}{N}\underset{k=1}{\overset{N}{\Σ}}{\tilde{S}}_3\left(K\right)$

whereinIs S₃(I) N is 1% of the magnitude of the number of image pixels.

In one embodiment of the invention, said establishing a link between contrast and subjective visual quality is performed by an overall contrast function;

$D_{n-1}=\frac{1}{N_{pix}(N_{pix}-1)}\underset{i=0}{\overset{M-2}{\Σ}}\underset{j=i+1}{\overset{M-1}{\Σ}}{H}_{n-1}\left(i\right)H_{n-1}\left(j\right)(j-i),i,j\∈\[0,M-1\];$

the [0, M-1 ]]Is the gray scale range of the image, N_pixIs the number of pixel points of the image, and h (i) is the amplitude of the image gray i in the histogram.

Drawings

Fig. 1 is a schematic diagram of a parameter-free video subjective quality assessment method according to the present invention.

Fig. 2 is a schematic diagram of the 8 × 8 block moving range according to the present invention.

FIG. 3 is a schematic diagram of a passivation ambiguity assessment algorithm according to the present invention.

Detailed Description

In order to make the technical means, the creation characteristics, the achievement purposes and the effects of the invention easy to understand, the invention is further described with the specific embodiments.

As shown in fig. 1, the non-parameter evaluation method for subjective quality of video according to the present invention includes the following steps:

step 1: the method comprises the steps of carrying out quantitative estimation on the visibility of noise contained in a compressed video, fitting an estimated noise visibility dimension with subjective quality assessment of an image, establishing a relation between the noise visibility and subjective visual quality, namely an image structure similarity calculator SSIM, and realizing the evaluation on the subjective quality of the video by utilizing the internal relation.

Step 2: the quality of the videos with different focuses is quantitatively estimated by a passivation fuzzy estimation algorithm, the relation between the focuses and the subjective visual quality is established, namely a Spatial Sharpness Map relation is established, and the evaluation on the subjective quality of the videos is realized by utilizing the intrinsic relation.

And step 3: based on the Center-surround model, the image quality with different contrasts is quantitatively estimated, the relation between the contrasts and the subjective visual quality is established, namely a total contrast function, and the evaluation on the subjective quality of the video is realized by utilizing the internal relation.

It is known that some redundant information of the compressed image can be applied to predict the information of the original image, and the structural similarity between the compressed image and the original image can reflect the evaluation of the image quality in the visual perception of human eyes. This process is also similar to the process of subjective evaluation of images by humans. In the 1860 s, Hermann von Helmholtz has found that human evaluation of visual objects involves complex psychological and physiological processes. In his conclusion, we assume that the human brain has an internal generative model that can use non-local information to infer information about the reasoning of the original image. Meanwhile, a human takes the difference between the external input and the internal inference as an evaluation of image quality. In step 1, The shifted windows filtered algorithm is used to describe a human brain internal generation model, and The difference between external input and original image information deduced from The generation model is measured by an operator SSIM for measuring image structure similarity.

As shown in fig. 1, each image block is moved within a predetermined interval to find a similar block. The similar block is determined by determining whether the Mean Absolute Error (MAE) between the two image blocks is less than a predetermined value. Typically, we move each 8 x 8 block to be in the range of-8, as shown in FIG. 1.

Since the discontinuity of blocks is caused by quantization tables, we have chosen an optimized mae (MAXMAE) to evaluate block similarity, as follows:

$MAXMAE=\frac{1}{8}qt(0,0)\exp (1-\underset{u=0}{\overset{7}{\Σ}}\underset{v=0}{\overset{7}{\Σ}}qt(u,v)/L),$

wherein qt (u, v) is a quantization table. L255 is the maximum gray level quantization step.

The set of similar blocks can be represented by:

$\mathrm{\Θ}(a,b)=\left\{(i,j)\right|\frac{1}{64}\underset{i=0}{\overset{7}{\Σ}}\underset{j=0}{\overset{7}{\Σ}}|f_{(a,b)}^{(i,j)}(n,m)-f_{(a,b)}(n,m)|\≤MAXMAE\}$

where (i, j) is the motion parameter for the block, -8 ≦ i, j ≦ 8.(n, m) is the index value for the pixel in block (a, b).

Finally, reconstructing an image f'_(a，b)(x, y) is defined as follows:

${f^{\′}}_{(a,b)}(n,m)=\frac{1}{N\left(\mathrm{\Θ}\right(a,b))}\underset{i,j\∈\mathrm{\Θ}(a,b)}{\Σ}{f}_{(a,b)}^{(i,j)}(n,m).$

the algorithm for the SSIM evaluation criteria is shown below.

SSIM is an algorithm mainly used to measure image structure similarity, and it consists of three parts:

image brightness comparison equation:wherein mu_xAnd mu_yThe score is the mean of the images x and y. C₁Is thatA stability constant at near zero.

Contrast comparison equation:wherein sigma_xAnd σ_yAre the variances of the images x and y, respectively. C₂Is thatA stability constant at near zero.

Structural correlation equations for images x and y:wherein sigma_xyIs the correlation of the images x and y, C₃Is a stable constant.

The SSIM index consists of the above three equations:

SSIM(x，y)＝[l(x，y)]^α·[c(x，y)]^β·[s(x，y)]^γ。

usually we replace the SSIM index with MSSIM to increase the computation speed.

$MSSIM(x,y)=\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}SSIM(x_i,y_i),$

Where M is the number of local windows x_iAnd y_iIs the content of the ith window of images x and y.

From the two sections above, we can derive a new approach to blockiness evaluation:

${\mathrm{MSSIM}}_{NF}=\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}SSIM(f_i^{\′},f_i),$

wherein f'_iIs the predicted original image.

As shown in fig. 2, the image blurring algorithm in step 2 is composed of two parts, i.e., measuring the slope of the local amplitude spectrum, and measuring the local maximum variation of the spatial domain. It is well known that the loss of high frequency components of an image causes blurring of the image. A good way to measure this effect is to measure m (f) of the image magnitude spectrum, since it decreases inversely with frequency. For example, M (f). varies.. f^-aWhere f is the frequency. Although the evaluation of the slope of the amplitude spectrum can be used for measuring the sharpness or the blurring degree of the image, the method does not consider the contrast of the image, and experiments show that the contrast of the image has a direct influence on the sharpness and the blurring degree of the image. To take into account the effect of contrast on image sharpness and blur, we measure the effect of contrast on image sharpness and blur by measuring the local maximum variation of the spatial domain. The two parts are effectively combined to well measure the blurring degree of the image.

In the Spectral sharpness map, the image I (x, y) is subjected to DFT change to obtain I_DFT(u, v), then I_DFT(u, v) into polar equation I_DFT(f, θ), wherein m is the size of the local block.

Next we calculate the sum of the amplitudes in all directions at the same frequency as follows:

$Z_I\left(f\right)=\underset{\mathrm{\θ}}{\Σ}\left|I_{DFT}(f,\mathrm{\θ})\right|.$

the slope of the amplitude spectrum of image I (x, y) is defined as-a_IIt can be calculated from the slope of-alogf + log β, and its mathematical expression is as follows:

$a_I=\arg \underset{a}{\min }\left|\right|{\mathrm{\βf}}^{-a}-Z_I\left(f\right)|{|}^2.$

from the above formula, the structure of a_IThe larger the higher frequency content of the image, the more blurred the image. Here, the Spectralsharpness map is defined as follows:

$S_1\left(I\right)=1-\frac{1}{1+e^{-3(a_I-2)}}.$

the total amount of Spatial domain variation in Spatial Measure of Sharpness is used by us to Measure Sharpness and blur of images. An image block I_bThe Total Variation of (I) v (I)_b) Can be calculated by the following equation:

$v\left(I_b\right)=\frac{1}{255}\underset{i,j}{\Σ}|I_{bi}-I_{bj}|.$

wherein I_biAnd I_bjIs I_bEight fields of the invention.

v(I_b) Effectively display image block I_bThe sum of absolute differences from its offset blocks. We thus define a spatialsharpness map as follows.

$S_2\left(I\right)=\frac{1}{4}{\max }_{\mathrm{\τ}\∈I}v\left(I_b\right),$

Where τ is a 2 × 2 block in I.

We synthesize the matching graph obtained in the spatial and frequency domains into a new match using the following formula:

S₃(I)＝S₁(I)^γ×S₂(I)^1-γ

we use S₃(I) The average of the 1% maximum measures the quality of the image:

$S_{3\_INDEX}=\frac{1}{N}\underset{k=1}{\overset{N}{\Σ}}{\tilde{S}}_3\left(K\right),$

whereinIs S₃(I) N is 1% of the magnitude of the number of image pixels.

Step 3 is mainly based on the application of local band-pass filter and center-peripheral receiving area model (center-surround receiving field model). Some views of the physiological studies on the influence of chromatic aberration on the human eye's image contrast sensitivity, and the view of sub-threshold contrast performance, are also considered in our algorithm. Considering the presence of multi-spectral channels in the human eye, our algorithm will work on the multi-spectral channels and then synthesize the resulting contrast-related values on the different channels into a scalar quantity with respect to the image contrast by means of the Lp norm.

We consider the contrast in the grey scale map to be seen as the ratio of the local variation to the local average. While we used the local band-limited contrast to measure the contrast responsible for the image by analyzing the weber & Michelson definition of contrast.

The local band-limited contrast is defined as follows:

$c(x,y)=\frac{\mathrm{\β}(x,y)}{\mathrm{\λ}(x,y)}$

wherein:

β(x，y)＝f(x，y)*b(x，y)

λ(x，y)＝f(x，y)*l(x，y)

f (x, y) is the gray value of the image at the coordinates (x, y), b is a band-pass filter, l is a low-pass filter, and x is the convolution operation.

We simply consider the perception phenomena of the human retina, the two main image-receiving neurons being rods and cones, with the signals generated by light conduction in the retina producing excitatory and inhibitory fields that in turn result in the center-surround visual acceptance field. DOG this in turn can be used to model the center-surround visual acceptance field.

O(x，y)＝C(x，y)-S(x，y)

Wherein,

C(x，y)＝f(x，y)*g₁

S(x，y)＝f(x，y)*g₂

this person O (x, y) is the output based on the Center-surround model DoG, C stands for Center and S stands for surround. g₁And g₂Is two gaussian equations, we now use the following DoG to set the theory above with the following equation:

$C(x,y)=\frac{\mathrm{\β}(x,y)}{\mathrm{\λ}(x,y)}=\frac{O(x,y)}{S(x,y)}$

considering that the human eye is multi-channel, in order to simulate the characteristics of the human eye, we set up

${\mathrm{\σ}}_{g_1}=\frac{2log\left(\frac{1}{M^2}\right)}{v^2(1-M^2)},v=\frac{72\mathrm{\π}}{80},\frac{69\mathrm{\π}}{80},\mathrm{...},\frac{6\mathrm{\π}}{80},\frac{3\mathrm{\π}}{80}$

Is g₁Of standard deviation of (A), while. Typically we set M to 3. In this way we constructed 24 channels. Thus we get a multi-valued criterion for the contrast of the image, and we use a multi-dimensional norm to integrate the values of the multiple channels together in order to get a single valued calibration value.

$C_L(x,y)=\underset{{\mathrm{\σ}}_{g_1}}{\Σ}{(P^{-1}\×|c_{{\mathrm{\σ}}_{g_1}}\left(x,y\right){|}^{\left(p\right)})}^{\left(\frac{1}{p}\right)}$

Since 1-norm mainly reflects a threshold contract period, when p is 1, max-norm mainly reflects a threshold contract period, when p is infinity, we mainly consider these two norms,

the local band-limited contrast (LBPC) of the whole image is

LBPC＝std(C₁)+std(C_∞。)

For a better measure of the contrast of an image, we also consider the global contrast by the variance of the conventional image histogram in addition to the local contrast, we apply the following approach in our invention to evaluate the overall contrast:

$D_{n-1}=\frac{1}{N_{pix}(N_{pix}-1)}\underset{i=0}{\overset{M-2}{\Σ}}\underset{j=i+1}{\overset{M-1}{\Σ}}{H}_{n-1}\left(i\right)H_{n-1}\left(j\right)(j-i),i,j\∈\[0,M-1\]$

here, [0, M-1 ]]Is the gray scale range of the image, N_pixIs the number of pixel points of the image, and h (i) is the amplitude of the image gray i in the histogram. The reason for choosing D as the evaluation criterion is: a larger D indicates a larger distance between each gray level element inside the histogram and a larger contrast of the image from one side.

Compared with the traditional video subjective quality evaluation method, the non-parameter evaluation method of the video subjective quality provided by the invention has the following advantages:

1. a parameter-free evaluation technique is employed. When the video quality is evaluated, the original video sequence is not required to be referred to, and the method is suitable for most application scenes in which the original video sequence cannot be obtained, such as a highway video monitoring system and an internet bar video monitoring system used in the public security industry;

2. because the subjective algorithm is adopted to carry out the non-parameter evaluation of the video quality, the manual intervention is not needed, so that the evaluation result is not influenced by the human subjective factors and has objectivity;

3. by adopting the computer to replace the evaluator, the video sequence with large data volume and long time can be comprehensively and statistically evaluated more quickly.

The foregoing shows and describes the general principles and broad features of the present invention and advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (4)

1. A non-parameter evaluation method for subjective quality of video is characterized by comprising the following three aspects:

1) quantitatively evaluating the visibility of noise contained in a compressed video, fitting the evaluated noise visibility dimension with subjective quality evaluation of an image, establishing a relation between the noise visibility and the subjective visual quality, namely an image structure similarity operator SSIM, and evaluating the difference between external input and original image information deduced by a generated model through the operator SSIM for evaluating the image structure similarity so as to evaluate the subjective quality of the video;

2) quantitatively evaluating the video quality of different focuses by using a passivation fuzzy evaluation algorithm, establishing a relation between the focuses and subjective visual quality, and realizing the evaluation of the subjective quality of the videos, wherein the passivation fuzzy evaluation algorithm consists of two parts, namely measuring the slope of a local amplitude spectrum and measuring the local maximum variation of a spatial domain;

3) based on the application of a local band-pass filter and a central-peripheral accepting area model, the method works on multi-spectral channels, and integrates the obtained values on the contrast on different channels into a scalar quantity on the contrast of an image through an Lp norm, so as to quantitatively estimate the image quality of different contrasts, establish the relation between the contrast and subjective visual quality and realize the evaluation of the subjective quality of a video.

2. The method according to claim 1, wherein the establishing of the relationship between the noise visibility and the subjective visual quality is performed by an image structure similarity calculator (SSIM); the SSIM index includes:

image brightness comparison equation:

wherein mu_xAnd mu_yMean values of the images x and y, C, respectively₁Is thatA stability constant at near zero;

contrast comparison equation:

wherein sigma_xAnd σ_yVariance of images x and y, C₂Is thatA stability constant at near zero;

structural correlation equations for images x and y:

wherein sigma_xyIs the correlation of the images x and y, C₃Is a stability constant;

the method consists of the following three equations:

SSIM(x，y)＝[l(x，y)]^α·[c(x，y)]^β·[s(x，y)]^γ。

3. the method of claim 1, wherein the establishing of the relationship between the focus and the subjective visual quality is performed by Spatial sharp Map relation; the Spectralsharpness map is defined as follows:

$S_1\left(I\right)=1-\frac{1}{1+e^{-3(a_I-2)}};$

wherein-a_IIs the amplitude spectral slope of image I (x, y);

the total variation of Spatial domain in Spatial Measure of Sharpness is used to Measure the Sharpness and blur of an image, an image block I_bTotal amount of change v (I)_b) Can be calculated by the following equation:wherein I_biAnd I_bjIs I_bThe eight fields of (1);

v(I_b) Effectively display image block I_bThe sum of the absolute differences with its offset blocks, from which we define a spatialsharpness map as follows:

$S_2\left(I\right)=\frac{1}{4}{\max }_{\mathrm{\τ}\∈I}v\left(I_b\right);$

with S₃(I) The average of the 1% maximum measures the quality of the image:

$S_{3\_INDEX}=\frac{1}{N}\underset{k=1}{\overset{N}{\Σ}}{\tilde{S}}_3\left(K\right)$

wherein S₃(I)＝S₁(I)^γ×S₂(I)^1-γ；Is S₃(I) N is 1% of the magnitude of the number of image pixels.

4. The method according to claim 1, wherein the establishing of the relationship between the contrast and the subjective visual quality is performed by an overall contrast function;

$D_{n-1}=\frac{1}{N_{pix}(N_{pix}-1)}\underset{i=0}{\overset{M-2}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{M-1}{\mathrm{\Σ}}}{H}_{n-1}\left(i\right)H_{n-1}\left(j\right)(j-i),i,j\∈\[0,M-1\];$

the [0, M-1 ]]Is the gray scale range of the image, N_pixIs the number of pixel points of the image, and h (i) is the amplitude of the image gray i in the histogram.