CN110853027A - Three-dimensional synthetic image no-reference quality evaluation method based on local variation and global variation - Google Patents

Abstract

The invention relates to a reference-free quality evaluation method for three-dimensional composite images based on local changes and global changes. The color feature is encoded using the local binary mode to obtain the structure and color feature maps, and the structure and color feature maps are calculated to obtain the structural features and color features, so as to obtain the distortion information of the local structure and color; then, for the global change detection, Luminance features are extracted to evaluate the naturalness of 3D synthetic images; finally, based on the extracted visual features above, a random forest regression model is trained to map the extracted features to subjective quality scores. The experimental results on three public databases show that the proposed method exhibits good effectiveness and superiority compared with the existing no-reference image quality evaluation methods and full-reference 3D synthetic image quality evaluation methods.

Description

A no-reference quality assessment of 3D composite images based on local and global changes Valuation method

technical field

The invention belongs to the technical field of multimedia, in particular to the technical field of digital images and digital image processing, and particularly relates to a three-dimensional composite image quality evaluation method without reference based on local changes and global changes.

Background technique

Free-view video (FVV) and 3D film and television can bring people an immersive experience, and these technologies have gained a lot of attention from academia and industry in the past few decades. Usually a method to obtain a free-view video is to use multiple cameras to capture different images of the same scene at different viewpoints, and then stitch these images captured from different viewpoints. Feel free to change their perspective. As the demand for a better experience increases, the number of views in free-view video continues to grow, and the load on storage and transmission also increases. To solve this problem, multi-view video deepening (MVD) technology has been developed.

The multi-view video deepening technology only needs the original view map and the depth map of other cameras, and the remaining virtual views can be generated by rendering-based deepening of pictures (DIBR). However, the DIBR process can cause new visual distortions such as blurring, discontinuities, blockiness and image stretching, etc. These distortions are clearly different from traditional distortions such as Gaussian blur, Gaussian noise and white noise, which produce The distortion will seriously affect the experience of the end user. Therefore, it is necessary to design a reliable and effective 3D composite image quality assessment method to predict the visual quality of 3D composite images. In the past few years, some quality assessment methods for synthetic distortion have been proposed, but the prediction effect of these methods is not accurate enough. Therefore, the present invention proposes a three-dimensional synthetic image quality assessment method based on local changes and global changes, which can effectively Predict the visual quality of synthetic images.

SUMMARY OF THE INVENTION

The invention relates to a reference-free quality evaluation method for three-dimensional composite images based on local changes and global changes. The color feature is encoded using the local binary mode to obtain the structure and color feature maps, and the structure and color feature maps are calculated to obtain the structural features and color features, so as to obtain the distortion information of the local structure and color; then, for the global change detection, Luminance features are extracted to evaluate the naturalness of 3D synthetic images; finally, based on the extracted visual features above, a random forest regression model is trained to map the extracted features to subjective quality scores. The experimental results on three public databases show that the proposed method exhibits good effectiveness and superiority compared with the existing no-reference image quality evaluation methods and full-reference 3D synthetic image quality evaluation methods.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A method for evaluating the quality of three-dimensional composite images without reference based on local changes and global changes, which is characterized by comprising the following steps:

A. Use Gaussian derivatives to extract the structure and color features of synthetic images;

B. Use the local binary mode to encode the obtained structure and color features respectively, obtain the structure and color feature maps, and calculate the structure features and color features based on the structure feature map and color feature map respectively, so as to obtain the local structure and color features. distorted information;

C. For global changes, extract brightness features to evaluate the naturalness of 3D composite images;

D. Combined with the extracted feature information, a random forest regression model is used to learn the mapping relationship between visual features and subjective quality scores to predict the quality scores of 3D synthetic images.

Further, using Gaussian derivatives to extract the structural features and color features of the image.

Further, using the Gaussian derivative to extract the structural features of the image, the specific steps are:

A. The local Taylor series expansion can represent the local features of the image. Through the local Gaussian derivative, the coefficient of the local Taylor series can be obtained; the Gaussian derivative of an image can be defined as follows:

Among them, m≥0 and n≥0 are the derivatives along the horizontal x and vertical y directions, and the symbol * represents the convolution operation; G ^σ (x, y, σ) is a Gaussian function, and its standard deviation σ is defined as follows:

通过计算高斯导数，所得矩阵

可以表示为：B. Use the second-order Gaussian derivative to extract structural features, first calculate 1≤m+n≤2

By computing the Gaussian derivative, the resulting matrix

It can be expressed as:

Further, using the Gaussian derivative to extract the color feature of the image, the specific steps are:

A. Two color features that are not affected by brightness are used on the first-order Gaussian derivative of the color channel, one of which is defined as follows:

R,G，B在颜色空间中分别代表红、绿、蓝通道；另一个颜色特征

定义如下：in,

R, G, B represent the red, green, and blue channels, respectively, in the color space; another color feature

Defined as follows:

where R', G' and B' represent the first-order Gaussian derivative values of the R, G and B channels in the horizontal direction, respectively, and ρ=2R'-G'-B', δ=2G'-R'-B' , τ=2B'-R'-G'.

Further, the local binary mode is used for the obtained structure and color features to obtain the structure and color feature maps, and the quality features are calculated to obtain the distortion information of the local structure and color.

Further, the structure diagram is obtained by using the local binary mode method, and the distortion information of the local structure is obtained, and the specific steps are:

中每个像素进行操作，计算得出基于

绝对值的特征图谱

其计算公式如下：A. Using the local rotation invariant binary mode method to

Each pixel in the operation is operated, and the calculation is based on

Feature map of absolute value

Its calculation formula is as follows:

和

其中

描述了中心像素点和局部区域临近像素点之间的关系；Among them, s∈{s ₁ , s ₂ , s ₃ , s ₄ , s ₅ }; LBP represents LBP operation; riu2 represents rotation invariant unified mode; D and E represent the number and calculation radius of surrounding pixels, and the surrounding pixels are The number D is set to 8, and the calculated radius E is 1; thus, 5 feature maps are obtained, which are

and

in

Describes the relationship between the central pixel and adjacent pixels in the local area;

的像素进行累加获得加权直方图，其定义公式如下：B. Use the weighted histogram to represent the local structural distortion information, through the same local binary pattern operator pair

The pixels are accumulated to obtain a weighted histogram, and the definition formula is as follows:

汇总高斯导数来融合映射高斯导数中根据LBP图强度值的像素值，并且经过归一化操作得到特征向量，增强图片区域中高对比度的变化，以此反映局部结构失真信息。Among them, N represents the number of picture pixels; K represents the index of LBP, K∈[0,D+2], is the weight and its value is the feature map

The Gaussian derivatives are aggregated to fuse the pixel values according to the LBP map intensity values in the mapped Gaussian derivatives, and the feature vector is obtained through the normalization operation to enhance the high-contrast changes in the image area to reflect the local structural distortion information.

Further, the chromaticity map is obtained by encoding, and the distortion information of the local chrominance is obtained; wherein,

然后将特征图转变成特征向量，其定义公式如下：A. Use the local binary model to obtain the feature vector: perform the LBP ^riu2 operation on the extracted color feature x ₁ to obtain the feature map

Then the feature map is transformed into a feature vector, and its definition formula is as follows:

是由局部二值模式算子得到的权值,其值为特征图

in,

is the weight obtained by the local binary pattern operator, and its value is the feature map

然后x₂的加权直方图计算定义如下：B. Use weighted histogram to represent local chromaticity information: perform LBP ^riu2 operation on the extracted color feature x ₂ to obtain a feature map

Then the weighted histogram calculation for _x2 is defined as follows:

为特征图谱

最后，表示图像彩色信息的单一特征向量由如下公式计算：Among them, the weight

feature map

Finally, a single feature vector representing the color information of the image is calculated by the following formula:

Further, for global variation, the naturalness of the 3D composite image is evaluated by extracting luminance features: where,

A. Use a Gaussian distribution to fit the luminance coefficient, the luminance coefficient (L ^′ ) is defined as follows:

where (i,j) represents the spatial position of the pixel, and i∈{1,2,…,a},j∈{1,2,…,b}, where a and b represent the height and width of the image, respectively , μ(i,j) and σ(i,j) are defined as follows:

Among them, ω is a two-dimensional centrosymmetric Gaussian weight function, ω={ω _a,b |a∈[-3,3],b∈[-3,3]};

Then a zero-mean generalized Gaussian distribution is used to fit the luminance parameter L'(i,j), which is defined by the following formula:

并且

其中参数α控制分布的形状，σ控制方差；in,

and

where the parameter α controls the shape of the distribution and σ controls the variance;

B. Subsequently, 4 parameters including the shape parameter and variance of the generalized Gaussian distribution and the kurtosis and skewness of the luminance coefficient are calculated on the 5 scales of the synthesized image, resulting in a total of 20-dimensional features; The difference between the filtered images is calculated to obtain a Laplacian pyramid image, a generalized Gaussian distribution model is used to fit the pixel values in the Laplacian pyramid image to obtain the shape coefficient and variance, and the Laplacian is calculated. The kurtosis and skewness of the pyramid plot, and extract these four parameters from five scales, resulting in a total of 20-dimensional features.

Further, a random forest regression method is used to train a quality prediction model, where,

A. Selection of feature information, through local features and global features, a total of 310-dimensional feature vectors are obtained, including 270-dimensional local variable features and 40-dimensional global natural features;

B. Use the random forest regression method to train the visual quality prediction model to map the quality features to the subjective evaluation. The 3D synthetic view quality database was randomly divided into training and test sets for 1000 training runs, and finally the mean Pierce Linear Correlation Coefficient (PLCC), Spearman Rank Correlation Coefficient (SRCC), Kendall Rank Correlation Coefficient (KRCC) and Root Mean Square Error (RMSE) as the final result.

Description of drawings

FIG. 1 is an algorithm framework diagram of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, but not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

Wherein, the technical features, abbreviations/abbreviations, symbols, etc. involved in this document are explained, defined/illustrated on the basis of the common knowledge/common understanding of those skilled in the art.

Inspired by the proposed local and global distribution of 3D composite image degradation and the fact that the human visual system is very sensitive to structure, color information and global natural changes, the present invention designs a novel and effective 3D composite view based on local and global changes without reference. The quality evaluation method (LVGC) can effectively improve the effect of quality evaluation of three-dimensional composite images. The visual features used include structural features, chromaticity features, image naturalness, and regression prediction.

The concrete operations of each part of the present invention are as follows:

(1) Structural feature extraction:

The research shows that the local Taylor series expansion can represent the local features of the image and the coefficients of the local Taylor series can be obtained by the local Gaussian distribution. The Gaussian distribution of an image can be defined as:

where m≥0 and n≥0 are the derivatives along the horizontal (defined as x) and vertical (defined as y) directions, in particular, the symbol * represents the convolution operation and ^Gσ (x,y,σ) is the Gaussian function , and its positive deviation σ is defined as follows:

通过计算高斯导数，矩阵

可以表示为：Inspired by other researches, the present invention applies the second-order Gaussian derivative to extract structural features; first, it is calculated under 1≤m+n≤2 to obtain

By computing the Gaussian derivative, the matrix

It can be expressed as:

中每个像素进行操作实现旋转不变性，可以计算得出基于

绝对值的特征图谱

其计算公式如下：Then, using the Local Uniform Rotation Invariant Binary Pattern Operator (ULBP) pair

Rotation invariance is achieved by operating on each pixel in , which can be calculated based on

Feature map of absolute value

Its calculation formula is as follows:

和

)，其中

描述了中心像素点和局部区域临近像素点之间的关系，并且这个局部细节可以有效地捕捉到由于不同失真类型而造成的复杂降解。where s∈{s ₁ , s ₂ , s ₃ , s ₄ , s ₅ }; LBP stands for LBP operation; riu2 stands for rotation-invariant uniform mode; D and E stand for the number of surrounding pixels and the adjacent radius. In particular, set the number of surrounding pixels D to 8 and the adjacent radius E to 1 to obtain 5 feature maps (respectively

and

),in

The relationship between the central pixel and adjacent pixels in the local area is described, and this local detail can effectively capture the complex degradation caused by different distortion types.

Although the local binary mode can detect the difference between the center pixel and its neighboring pixels, it cannot accurately capture the gradient information, it encodes the difference between adjacent pixels, which will weaken the ability of the local binary mode to distinguish the local variation. This is critical, as changes in local contrast can have a large impact on assessing the visual quality of an image. It is well known that contrast transformation is highly related to the visual quality of pictures. Therefore, the present invention accumulates through the same local binary pattern operator to obtain a weighted histogram, which is defined as follows:

为权值，其值为特征图谱

本发明运用高斯导数汇总融合映射高斯导数中图强度的像素值，并且经过归一化操作得到特征向量；经过这些操作，可以增强图片区域中高对比度的变化。where N represents the number of picture pixels; K represents the index of the LBP, K∈[0,D+2], and

is the weight, and its value is the feature map

The present invention uses the Gaussian derivative to summarize and fuse the pixel values of the image intensity in the Gaussian derivative, and obtains a feature vector through normalization operations; through these operations, the changes of high contrast in the image area can be enhanced.

(2) Color feature extraction:

In order to extract color features, the present invention adopts two color luminances that are not affected by luminance on the chrominance channel of the first-order Gaussian derivative; experiments have confirmed that the information of the first-order Gaussian derivative of color can be used to perceive the degradation of local structures, one of which is Color features can be defined as:

然后将特征图转变成特征向量，计算公式定义如下：in, R, G, and B represent the red, green, and blue channels, respectively, in the color space. Then, perform LBP ^riu2 operation on x ₁ to extract feature maps

Then the feature map is transformed into a feature vector, and the calculation formula is defined as follows:

是权值，其值为特征图谱

另外一个特征为

其定义如下：in,

is the weight, and its value is the feature map

Another feature is

It is defined as follows:

操作应用于x₂计算得到特征图谱

权重直方图计算如下：Among them, R', G' and B' represent the first-order Gaussian derivative values of the R, G, and B channels in the horizontal direction, respectively, and ρ=2R'-G'-B', δ=2G'-R'-B' , τ=2B'-R'-G'. followed by

The operation is applied to x ₂ to calculate the feature map

The weight histogram is calculated as follows:

是被定义为

的权重，计算公式如下：in,

is defined as

The weight of , the calculation formula is as follows:

Color features are invariant to luminance and luminance-related scene information effects such as shadows, therefore, because they are not affected by lighting, they can represent strong structural information; not only that, images caused by a single factor (such as blurring) Distortions can damage the structure of an image, but they are not necessarily related to the effects of brightness.

(3) Image naturalness representation:

The brightness distortion of the 3D composite image may affect the naturalness, and a high-definition 3D composite image should have the natural characteristics of a natural picture. Therefore, the present invention uses the brightness-based quality feature to evaluate the naturalness of the three-dimensional composite image. Considering that the brightness parameter of the natural picture will follow a Gaussian distribution, the brightness coefficient is used to calculate the naturalness of the composite image; the brightness parameter (L') Defined as follows:

where (i,j) represents the spatial index, and i∈{1,2,…,a},j∈{1,2,…,b}, where a and b represent the height and width of the image, respectively.

In particular, μ(i,j) and σ(i,j) are defined as follows:

where ω is a 2D centrosymmetric Gaussian weight function sampled by three standard deviations and rescaled to unity, ω={ω _a,b |a∈[-3,3],b∈[-3,3]} .

The luminance parameter L'(i,j) is fitted with a zero-mean generalized Gaussian distribution, and the formula is defined as follows:

并且

参数α控制分布的一般形状，σ控制方差，两个参数(α，σ²)通过这个模型得到评估；并且亮度系数的峰度和偏度由来自5个以上的经典分布计算得出，产生了总共20个特征。in,

and

The parameter α controls the general shape of the distribution, σ controls the variance, and two parameters (α, σ ² ) are evaluated by this model; and the kurtosis and skewness of the luminance coefficients are calculated from more than 5 classical distributions, yielding 20 features in total.

Then, a Laplacian pyramid is computed by the difference between the composite image and its low-pass filtered version, and a generalized Gaussian distribution model is used to fit the distribution of pixel values in the Laplacian pyramid, Taking the kurtosis and skewness of the Laplacian pyramid graph as features, the present invention extracts quality-conscious features from five levels, resulting in a total of 20 features.

(4) Regression model and quality prediction:

Studies have shown that multi-scale properties exist in the human visual system when perceiving visual information, so extracting visual features that extract images at multiple scales can be better characterized. By representing the model by local features and global features, the present invention can obtain a total of 310-dimensional feature vectors, including 270-dimensional local variable features and 40-dimensional global natural features; then, the random forest method is used to train the visual quality prediction model , so that quality features can be mapped to subjective evaluations; by randomly partitioning the database: 80% of the image samples and corresponding subjective evaluation scores are used in the database for training, and the remaining 20% are used for testing; finally, the Pierce linear The mean correlation coefficient (PLCC), Spearman rank correlation coefficient (SRCC), Kendall rank correlation coefficient (KRCC) and root mean square error (RMSE) were used as final results.

The process of the present invention is shown in Figure 1, and the concrete process is as follows:

Step 1: Extract structure and color features using Gaussian derivatives;

Step 2: Use the local binary pattern to encode the structure and color feature maps, which are used to calculate the quality-aware features to obtain local structure and color distortion;

Step 3: Extract luminance features through global changes to evaluate the naturalness of the 3D composite image;

Step 4: Based on the above extracted feature information, use random forest regression to map from visual features to subjective scores to train a quality prediction model.

The mean Pierce Linear Correlation Coefficient (PLCC), Spearman's Rank Correlation Coefficient (SRCC), Kendall's Correlation Coefficient (KRCC) and Root Mean Square Error (RMSE) were used as final comparison results. Generally speaking, higher PLCC and SRCC and lower RMSE value represent better performance of the algorithm, that is, the algorithm has better prediction accuracy of the algorithm. In order to verify the performance of the algorithm proposed by the present invention, the algorithm is compared with the existing quality evaluation methods with reference and without reference in three public databases MCL-3D, IRCCyN/IVC, and IETR-DIBR. The compared evaluation methods include PSNR, SSIM, BRISQUE, NIQE, BIQI, NRSL, CM-LOG, MP-PSNRr, MW-PSNR, MW-PSNRr, LOGS, Ref, APT and NIQSV+; the first seven methods are quality evaluation methods for natural pictures, the latter seven It is a quality evaluation method specially designed for synthetic viewing angle.

The MCL-3D database contains 693 stereoscopic image sets selected from nine depth image sources. The IRCCyN/IVC DIBR database consists of 12 reference pictures selected from three MVD sequences and 84 composite pictures generated by 7 different DIBR techniques. The IETR DIBR database consists of 10 MVD sequences and 7 150 composite images produced by the latest DIBR technology. Similar to the RCCyN/IVC DIBR database, IETR DIBR also mainly focuses on rendering distortion.

Table 1: Comparison of the present invention with the existing full reference method

Table 1 shows the comparison between the proposed method and the existing full reference method. From these comparisons, the no-reference method proposed by the present invention performs better.

Table 2: Comparison of the present invention with existing methods without reference

Table 2 shows the comparison between the proposed method and the existing no-reference method. From these comparisons, the no-reference method proposed by the present invention performs better on the tested database.

The above-mentioned embodiment is an illustration of the present invention, not a limitation of the present invention. It can be understood that various changes, modifications, replacements and modifications can be made to these embodiments without departing from the principle and spirit of the present invention. The protection scope of the present invention It is defined by the appended claims and their equivalents.

Claims (9)

1. A three-dimensional synthetic image no-reference quality evaluation method based on local variation and global variation is characterized in that: the method comprises the following steps:

A. extracting the structure and color characteristics of the synthesized image by using a Gaussian derivative;

B. coding the obtained structure and color characteristics by using a local binary pattern respectively to obtain a structure characteristic diagram and a color characteristic diagram, and calculating the structure characteristic diagram and the color characteristic diagram respectively based on the structure characteristic diagram and the color characteristic diagram to obtain distortion information of the local structure and the color;

C. extracting brightness characteristics to evaluate the naturalness of the three-dimensional synthetic image for global change;

D. and (4) combining the extracted characteristic information, and using a random forest regression model to learn the mapping relation between the visual characteristics and the subjective quality scores to predict the quality scores of the three-dimensional synthetic images.

2. The method of claim 1, wherein: structural features and color features of the image are extracted using gaussian derivatives.

3. The method of claim 2, wherein: the structural features of the image are extracted by using Gaussian derivatives, and the method comprises the following specific steps:

A. the local Taylor series expansion can represent the local characteristics of the image, and the coefficients of the local Taylor series can be obtained through local Gaussian derivatives; the gaussian derivative of an image can be defined as follows:

where m ≧ 0 and n ≧ 0 are derivatives along the horizontal x and vertical y directions, the symbol denotes the convolution operation; g^σ(x, y, σ) is a Gaussian function whose standard deviation σ is defined as follows:

B. using the second order Gaussian derivative to extract structural features, first calculating the sum of m + n < 1 > and 2

By calculating the Gaussian derivatives, the resulting matrix

Can be expressed as:

4. the method of claim 2, wherein: extracting color features of the image by using Gaussian derivatives, which comprises the following specific steps:

A. two color features that are not affected by luminance are employed on the first order gaussian derivative of the color channel, one of which is defined as follows:

wherein,

r, G and B respectively represent red, green and blue channels in a color space; another color characteristic

The definition is as follows:

wherein R ', G', and B 'respectively represent gaussian first derivative values of R, G, and B channels in the horizontal direction, and ρ ═ 2R' -G '-B', δ ═ 2G '-R' -B ', τ ═ 2B' -R '-G'.

5. The method of claim 2, wherein: and obtaining a structure characteristic graph and a color characteristic graph by respectively using the obtained structure and color characteristics through a local binary pattern, and calculating quality characteristics so as to obtain distortion information of the local structure and color.

6. The method of claim 5, wherein: obtaining a structure chart by using a local binary pattern method to obtain distortion information of a local structure, wherein the method comprises the following specific steps of:

A. binary pattern method pair using local rotation invariance

Each pixel of the image is operated on, the calculation is based onCharacteristic map of absolute value

The calculation formula is as follows:

wherein s ∈ { s ∈ [)₁,s₂,s₃,s₄,s₅}; LBP stands for LBP operation; riu2 represents a rotation invariant unified mode; d and E represent the number and the calculation radius of surrounding pixels, the number D of the surrounding pixels is set to be 8, and the calculation radius E is 1; thereby obtaining 5 characteristic maps which are respectively

And

wherein

Describing the relationship between the central pixel point and the adjacent pixel points in the local area;

B. representing local structure distortion information by using weighted histogram, and using same local binary pattern operator pair

The pixels of (a) are accumulated to obtain a weighted histogram, which is defined by the following formula:

wherein N represents the number of picture pixels; k denotes the index of LBP, K ∈ [0, D +2 ]],

Is a weight value which is a characteristic map

And summarizing the Gaussian derivatives to fuse and map pixel values in the Gaussian derivatives according to the LBP image intensity value, and obtaining a characteristic vector through normalization operation to enhance the change of high contrast in the image area so as to reflect local structure distortion information.

7. The method of claim 5, wherein: obtaining a chromaticity diagram by using coding to obtain distortion information of local chromaticity; wherein,

A. obtaining a feature vector by using a local binary model: at the extracted color feature x₁Carry out LBP^riu2Operation acquisition feature mapThe feature map is then converted into feature vectors, which define the formula:

wherein,

is a weight value obtained by a local binary pattern operator, and the value is a characteristic diagram

B. The local chrominance information is represented by a weighted histogram: at the extracted color feature x₂Carry out LBP^riu2Operation acquisition feature map

Then x₂The weighted histogram calculation of (a) is defined as follows:

wherein, the weight value

Is a characteristic map

Finally, a single feature vector representing the color information of the image is calculated by the following formula:

8. the method of claim 1, wherein: for global variation, the naturalness of the three-dimensional synthetic image is evaluated by extracting the brightness features: wherein,

A. the luminance coefficient is fitted using a gaussian distribution, and the luminance coefficient (L') is defined as follows:

where (i, j) represents the spatial position of the pixel, and i e {1,2, …, a }, j e {1,2, …, b }, where a and b represent the height and width of the image, respectively, and μ (i, j) and σ (i, j) are defined as follows:

where ω is a two-dimensional, centrosymmetric gaussian weighting function, ω ═ ω_a,b|a∈[-3,3],b∈[-3,3]}；

The luminance parameter L' (i, j) is then fitted using a zero-mean generalized gaussian distribution, which defines the formula:

wherein,

and is

Where parameter α controls the shape of the distribution, σ controls the variance;

B. subsequently, 4 parameters including shape parameters and variance of the generalized gaussian distribution and kurtosis and skewness of the luminance coefficient are calculated on 5 scales of the composite image, resulting in a total of 20-dimensional features; in addition, a laplacian pyramid image is computed from the difference between the composite image and its low-pass filtered image, the shape coefficients and variances are obtained using a generalized gaussian distribution model to fit the pixel values in the laplacian pyramid, the kurtosis and skewness of the laplacian pyramid are computed, and the four parameters are extracted from the five scales, yielding a total of 20-dimensional features.

9. The method of claim 1, wherein: a quality prediction model is trained using a random forest regression method, wherein,

A. selecting feature information, namely obtaining a total 310-dimensional feature vector through local features and global features, wherein the total 310-dimensional feature vector comprises 270-dimensional local variation features and 40-dimensional global naturalness features;

B. and training a visual quality prediction model by using a random forest regression method, and mapping the quality characteristics to subjective evaluation. The three-dimensional synthetic view quality database is randomly divided into a training set and a testing set to be trained for 1000 times, and finally, a pierce linear correlation coefficient mean (PLCC), a pierce Mandarin level correlation coefficient (SRCC), a Kendall level correlation coefficient (KRCC) and a Root Mean Square Error (RMSE) are used as final results.

Images