A quality evaluation method for dynamic scene multi-exposure fusion light field images

Abstract

The application discloses a quality evaluation method of a dynamic scene multi-exposure fusion light field image, which fully considers distortion characteristics and internal high-dimensional structure characteristics of a distorted multi-exposure fusion light field image from aspects of local, global and angle, etc., firstly performs tensor decomposition on an angle dimension of the distorted multi-exposure fusion light field image, which can effectively retain the high-dimensional characteristics and extract features more related to subjective perception on a tensor slice; then, a foreground part with significant artifacts is detected by designing a pre-background segmentation scheme before a first component, and edge structure and luminance / chrominance features of the foreground part are extracted; then, local features of low dark, normal and high bright areas are extracted based on an HSI model after luminance segmentation of the first component; subsequently, a perception feature vector is formed by combining global features and angle features extracted by other components, and an objective quality score is predicted by pooling, so that the predicted objective quality score has better consistency with visual perception of the human eye.

Description

Technical Field

[0001] This invention relates to an image quality assessment technique, and more particularly to a method for assessing the quality of dynamic scene multi-exposure fused light field images. Background Technology

[0002] Compared to traditional 2D images, light field images contain richer scene information, including complete parallax, light intensity, and direction information, making them suitable for rendering, object detection, and 3D reconstruction. However, the dynamic range of natural scenes is extremely large (up to nine orders of magnitude). Due to the limitations of the potential well capacity of imaging sensors, overexposure or underexposure can occur in environments with large illumination ranges. Existing commercial light field cameras struggle to record complete information about the large brightness dynamic range of a scene in a single exposure. To address this issue, multi-exposure light field imaging techniques can reconstruct light field images with large dynamic ranges. However, multi-exposure light field imaging inevitably introduces corresponding distortions, especially in dynamic scene imaging. Therefore, quality assessment of multi-exposure fused light field images presents a significant challenge in light field image processing.

[0003] To date, although some studies have considered combining light field imaging with multi-exposure fusion techniques, no researchers have conducted work on the visual quality assessment of multi-exposure fused light field images. Many researchers have proposed several relatively mature methods for assessing the quality of light field images, which can predict the quality of light field images relatively accurately. For example, Shi et al. proposed a blind quality estimator for light field images, which combines spatial and angular features of the light field image to estimate its quality. Another example is Xiang et al.'s method for assessing the quality of no-reference light field images based on pseudo-videos and refocused images. This method uses multi-scale, multi-directional shear wave transform to extract the structure, motion, and parallax information of the PV, and utilizes the spatial structure, depth, and semantic information of the RIs to evaluate the visual LFI of the visualization. Yet another example is Pan et al.'s blind light field image quality assessment method combining tensor slicing and singular values. This method extracts sharpness features from the first component of the tensor and uses the information distribution of other components and angular singular value features to measure the distortion of the light field image. Another example is the referenceless light field image quality evaluation method proposed by Liu et al., which uses the parallax relationship between sub-apertures to construct a pseudo-reference sub-aperture image and extracts features based on the pseudo-reference sub-aperture image and microlens image to measure the quality of the light field image.

[0004] While the methods described above are generally applicable to light field images, they may not yield satisfactory results for evaluating the quality of multi-exposure fused light field images. This is because the visual quality of multi-exposure fused light field images is influenced by factors such as multi-exposure fusion and tone mapping. For instance, during the fusion of dynamic scenes, dynamic objects may produce artifacts, brightness and chromaticity may be distorted during the fusion process, and the angular consistency of the light field image may be compromised. All of these factors contribute to a significant deterioration in the quality of multi-exposure fused light field images, with distortion patterns differing from those of typical light field images. Therefore, there is an urgent need to develop a novel image quality evaluation method specifically for multi-exposure fused light field images. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a quality evaluation method for multi-exposure fused light field images of dynamic scenes, which makes full use of the local, global and angular features of multi-exposure fused light field images to predict objective quality scores, and the predicted objective quality scores have better consistency with human visual perception.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for quality evaluation of dynamic scene multi-exposure fusion light field images, characterized by including the following steps:

[0007] Step 1: Denote the distorted multi-exposure fused light field image to be evaluated as _____. in, express It is five-dimensional, with U and V representing respectively. Angular dimension, where X and Y represent respectively The spatial dimension, C represents The color dimension; then Considered as a 2D sub-aperture image array, the sub-aperture image array consists of U×V sub-aperture images, each with a resolution of X×Y. The sub-aperture image with viewpoint coordinates (u,v) in the sub-aperture image array is denoted as G. u,v Where 1≤u≤U, 1≤v≤V;

[0008] Step 2: For Perform Tucker tensor decomposition on the angular dimension to obtain the decomposed components, denoted as . in, express It is five-dimensional, with U1 and V1 representing respectively. The angular dimension, U1 is less than or equal to U, V1 is less than or equal to V, and X and Y are also... The spatial dimension, C is also The number of color dimensions; then fix The first two dimensions U1 and V1 are obtained All tensor slices are represented as: in, L 1,1 This represents a tensor slice where the values ​​of U1 and V1 are fixed at 1 and 1 respectively. u,v L represents a tensor slice where the values ​​of U1 and V1 correspond to u and v, respectively. U,V This represents a tensor slice where the values ​​of U1 and V1 correspond to U and V respectively.

[0009] Step 3: Set the viewpoint coordinates in the sub-aperture image array to... The sub-aperture image is taken as the central sub-aperture image, denoted as G. centre Then calculate G. centre The depth map, denoted as D centre Then, the maximum inter-class algorithm is used in... and D is obtained from between centre Optimal threshold To D centre Perform foreground and background segmentation to obtain a binary image, denoted as D. centre,val Then D centre,val As weight and L 1,1 Combined, soon D centre,val With L 1,1 Multiply, and we get L 1,1 The foreground portion exhibiting obvious artifacts is denoted as L. 1,1,fore ; then L 1,1,fore Transforming from RGB space to HSI model, we obtain L 1,1,fore The hue component, saturation component, and intensity component are denoted as H. 1,1,fore S 1,1,fore and I 1,1,fore Finally, the I-scale model was analyzed using an eight-scale Gaussian difference model, histograms, generalized Gaussian functions, and a moment-matching method. 1,1,fore Edge structure features are extracted to obtain I. 1,1,fore The GGD fitted feature vector is denoted as Use histograms to analyze H 1,1,fore Color component feature extraction was performed to obtain H. 1,1,fore The order moment eigenvector is denoted as . Use histograms for I 1,1,fore Brightness features were extracted to obtain I. 1,1,fore The order moment eigenvector is denoted as . joint and The local feature vector extracted based on foreground and background segmentation is denoted as F1. in, The floor function is called the floor function. `min()` returns the minimum value, and `max()` returns the maximum value. D represents centre The minimum pixel value in D represents centre The maximum pixel value in Its dimension is 1×16. Its dimension is 1×4. The dimension of F1 is 1×4, and "[]" is a vector or matrix representation symbol. The dimension of F1 is 1×24.

[0010] Step 4: Place L 1,1 The minimum, maximum, and median values ​​of grayscale images are denoted as p. graymin p graymax p graymed Then use the maximum inter-class algorithm on p graymin and p graymed L is obtained from the interval 1,1 The first optimal threshold In p graymed and p graymax L is obtained from the interval 1,1 The second optimal threshold In this way, L 1,1 The area is divided into three regions with different brightness intensities: a dark region, a normal exposure region, and a bright region, denoted as L. 1,1,low L 1,1,nor and L 1,1,high Next, L 1,1,low Transforming from RGB space to HSI model, we obtain L 1,1,low The hue component, saturation component, and intensity component are denoted as H. 1,1,low S 1,1,low and I 1,1,low Similarly, L 1,1,nor Transforming from RGB space to HSI model, we obtain L 1,1,nor The hue component, saturation component, and intensity component are denoted as H. 1,1,nor S 1,1,nor and I 1,1,nor ; will L 1,1,high Transforming from RGB space to HSI model, we obtain L 1,1,high The hue component, saturation component, and intensity component are denoted as H. 1,1,high S 1,1,high and I 1,1,high Finally, extract I. 1,1,low The feature vectors are used to describe brightness and texture information, denoted as . Extract I 1,1,high The feature vectors are used to describe texture information, denoted as . Extract H 1,1,high The influence of the halo is characterized by its feature vector, denoted as . Extract S 1,1,high The influence of the halo is characterized by its feature vector, denoted as . Extract I 1,1,nor The feature vectors are used to describe texture information, denoted as . joint and The resulting local feature vector, denoted as F2, is constructed based on brightness segmentation. Among them, L 1,1,low The grayscale values ​​in the grayscale image at p graymin and Between, L 1,1,nor The gray values ​​in the grayscale image and Between, L 1,1,high The gray values ​​in the grayscale image and p gray max between, Indicate I 1,1,low The mean of the pixel values ​​of all pixels. Indicate I 1,1,low The standard deviation of the pixel values ​​of all pixels. Indicate I 1,1,low The information entropy of the pixel values ​​of all pixels, Indicate I 1,1,high The information entropy of the pixel values ​​of all pixels, H represents 1,1,high The mean of the pixel values ​​of all pixels. H represents 1,1,high The standard deviation of the pixel values ​​of all pixels. S represents 1,1,high The mean of the pixel values ​​of all pixels. Indicate I 1,1,nor The information entropy of the pixel values ​​of all pixels, and the dimension of F2 is 1×8;

[0011] Step 5: Place L 1,1 Transforming from RGB space to HSI model, we obtain L 1,1 The hue component, saturation component, and intensity component are denoted as H. 1,1 S 1,1 and I 1,1 ; then extract I 1,1 The contrast characteristics are denoted as Then, using an 8-scale Gaussian difference model, histogram, generalized Gaussian function, and moment matching-based method, I was extracted. 1,1 The GGD fitted feature vector is denoted as Finally, the joint and The feature vector extracted globally is denoted as F3. Where R represents the number of pixel gray levels, R = 256, γ i,k P represents the absolute difference between grayscale value i and grayscale value k. i,k This represents the probability that a pixel with grayscale value i and a pixel with grayscale value k appear as adjacent pixels. The dimension of F1 is 1×16, and the dimension of F3 is 1×17;

[0012] Step 6: [The sentence is incomplete and requires more context to be translated accurately.] Except for L 1,1 All remaining tensor slices are transformed from RGB space to HSI model to obtain the hue, saturation, and intensity components of each tensor slice. Then, a GLBP algorithm based on 5 scales is used to extract gradient texture feature vectors at different scales for the intensity component of each remaining tensor slice. The dimension of the gradient texture feature vector at each scale is 1×10. Then, the average vector of the gradient texture feature vectors of the intensity component of each remaining tensor slice at the same scale is calculated. Finally, the average vectors of the 5 scales are used to construct the feature vector based on angle extraction, denoted as F4. The dimension of F4 is 1×50.

[0013] Step 7: Combine F1, F2, F3, and F4 to form The perceptual feature vector, denoted as F LGA ,F LGA = [F1, F2, F3, F4]; where F LGA The dimension is 1×99;

[0014] Step 8: F LGA As input, the LIBSVM package is used to train the SVR with a radial basis function kernel for regression training, and the predictions are obtained. The objective quality score, denoted as Q. predict ; where Q predict The larger the value, the better the input F. LGA The better the quality of the corresponding multi-exposure fused light field image, the better; conversely, the worse the quality, the lower the quality of the input F-index. LGA The lower the quality of the corresponding multi-exposure fused light field image.

[0015] In step 2, the following will be done: The Tucker tensor decomposition is represented as: In the Tucker tensor decomposition process It is regarded as a fifth-order tensor. Represents the core tensor. D (1) This represents the factor matrix of the first angular dimension. D (2) This represents the factor matrix of the second angular dimension. D (3) The factor matrix representing the third dimension of space. D (4) The factor matrix representing the fourth dimension of space. D (5) The factor matrix representing the fifth dimension, color dimension. U1 and V1 are also The angular dimension, X1 and Y1 respectively represent The spatial dimension, X1 is less than or equal to X, Y1 is less than or equal to Y, and C1 represents The color dimension, C1 is less than or equal to C; then the core tensor Represented as: Here, the superscript "T" indicates the transpose of a vector or matrix; then the... Understanding Tucker decomposition from the perspective dimension as The factor matrix D with the first angular dimension (1) The factor matrix D in the second angular dimension (2) By multiplying along the pattern, the decomposed components are obtained. Then according to and get and make In The spatial dimension X1 equals X, Y1 equals Y, and the color dimension C1 equals C, thus achieving dimensionality reduction of the angular dimension.

[0016] In step 3, The acquisition process is as follows: I is obtained using an 8-scale Gaussian difference model. 1,1,fore Processing yields I 1,1,fore Eight edge maps at different scales were generated; then histograms were used to obtain I. 1,1,fore The coefficient distribution of the edge map at each scale is obtained and fitted using the shape and variance parameters of a generalized Gaussian function. Then, a moment-matching method is used to estimate the shape and variance parameters of the generalized Gaussian function, targeting I... 1,1,fore The shape parameter and variance parameter of the generalized Gaussian function corresponding to the Gaussian difference model at the j-th scale are denoted as β. j and σ j ; and then regarding I 1,1,fore By combining the shape and variance parameters of the generalized Gaussian function corresponding to the Gaussian difference model at eight scales, we obtain I.1,1,fore GGD fitted feature vector Where 1≤j≤J, J=8, β I,GGD Indicates that it is aimed at I 1,1,fore The shape parameter vector of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model, β I,GGD =[β1,β2,…,β J ],β1,β2,…,β J Corresponding to I 1,1,fore The shape parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the shape parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the shape parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model, σ I,GGD Indicates that it is aimed at I 1,1,fore The variance parameter vector σ of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model. I,GGD =[σ1,σ2,…,σ J ],σ1,σ2,…,σ J Corresponding to I 1,1,fore The variance parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the variance parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the variance parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model;

[0017] The acquisition process is as follows: use histograms to analyze H 1,1,fore Processing yields H 1,1,fore The first, second, third, and fourth moments; then H 1,1,fore The first, second, third, and fourth moments are arranged in sequence to form H. 1,1,fore eigenvectors of the order moments

[0018] The acquisition process is as follows: use histograms to analyze I 1,1,fore Processing yields I 1,1,fore The first, second, third, and fourth moments; then I 1,1,fore The first, second, third, and fourth moments are arranged in sequence to form I. 1,1,fore eigenvectors of the order moments

[0019] In step 5, an 8-scale Gaussian difference model is used to analyze I. 1,1 Processing yields I 1,1Eight edge maps at different scales were generated; then, histograms were used to obtain I. 1,1 The coefficient distribution of the edge map at each scale is obtained and fitted using the shape and variance parameters of a generalized Gaussian function. Then, a moment-matching method is used to estimate the shape and variance parameters of the generalized Gaussian function, targeting I... 1,1 The shape parameter and variance parameter of the generalized Gaussian function corresponding to the Gaussian difference model at the j-th scale are denoted as β. j 'and σ' j ; and then regarding I 1,1 By combining the shape and variance parameters of the generalized Gaussian function corresponding to the Gaussian difference model at eight scales, we obtain I. 1,1 GGD fitted feature vector Where 1≤j≤J, J=8, Indicates that it is aimed at I 1,1 The shape parameter vectors of the generalized Gaussian function corresponding to the eight scales of the Gaussian difference model. β′1,β′2,…,β' J Corresponding to I 1,1 The shape parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the shape parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the shape parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model. Indicates that it is aimed at I 1,1 The variance parameter vector of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model, σ I,GGD =[σ′1,σ'2,…,σ' J ],σ′1,σ′2,…,σ′ J Corresponding to I 1,1 The variance parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the variance parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the variance parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model.

[0020] Compared with the prior art, the advantages of the present invention are as follows:

[0021] (1) The present invention starts from the distortion generated in dynamic scenes by combining light field imaging technology and existing multi-exposure fusion algorithms, and proposes a quality evaluation method for multi-exposure fusion light field images in dynamic scenes, filling the gap in this field.

[0022] (2) Compared with existing light field image quality assessment methods, the method of this invention fully considers the distortion characteristics and inherent high-dimensional structural features of distorted multi-exposure fused light field images from multiple aspects, including local, global, and angular perspectives. Specifically, firstly, tensor decomposition is performed on the angular dimension of the distorted multi-exposure fused light field image, which effectively preserves its high-dimensional characteristics and extracts features more relevant to subjective perception from the tensor slices; then, a foreground / background segmentation scheme is designed for the first component, i.e., the first tensor slice, to detect the foreground part with significant artifacts, and edge structure and brightness / chromaticity features are extracted from the foreground part; next, after brightness segmentation of the first component, local features of dark, normal, and bright areas are extracted based on the HSI model; then, the perceptual feature vector is formed by combining global features and angular features extracted from other components, and pooling is used to predict the objective quality score. Therefore, the objective quality score predicted by the method of this invention has better consistency with human visual perception. Attached Figure Description

[0023] Figure 1 This is a block diagram illustrating the overall implementation of the method of the present invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0025] This invention proposes a quality assessment method for dynamic scene multi-exposure fusion light field images, the overall implementation block diagram of which is shown below. Figure 1 As shown, it includes the following steps:

[0026] Step 1: Denote the distorted multi-exposure fused light field image to be evaluated as _____. in, express It is five-dimensional, with U and V representing respectively. Angular dimension, where X and Y represent respectively The spatial dimension, C represents The color dimension; then Considered as a 2D sub-aperture image array, the sub-aperture image array consists of U×V sub-aperture images, each with a resolution of X×Y. The sub-aperture image with viewpoint coordinates (u,v) in the sub-aperture image array is denoted as G. u,v , where 1≤u≤U, 1≤v≤V.

[0027] Step 2: For Performing Tucker tensor decomposition (also known as higher-order singular value decomposition) on the angular dimension yields the decomposed components, denoted as . in, express It is five-dimensional, with U1 and V1 representing respectively. The angular dimension, U1 is less than or equal to U, V1 is less than or equal to V, and X and Y are also... The spatial dimension, C is also The color dimension is difficult to handle due to the high-dimensionality of multi-exposure fused light field images. Therefore, performing high-order singular value decomposition on distorted multi-exposure fused light field images can effectively reduce dimensionality while preserving the original color information of the light field image; then fix... The first two dimensions U1 and V1 are obtained All tensor slices are represented as: in, L 1,1 This represents a tensor slice where the values ​​of U1 and V1 are fixed at 1 and 1 respectively. u,v L represents a tensor slice where the values ​​of U1 and V1 correspond to u and v, respectively. U,V This represents a tensor slice where the values ​​of U1 and V1 correspond to U and V, respectively.

[0028] In this embodiment, in step 2, the The Tucker tensor decomposition is represented as: In the Tucker tensor decomposition process It is regarded as a fifth-order tensor. Represents the core tensor. D (1) This represents the factor matrix of the first angular dimension. D (2) This represents the factor matrix of the second angular dimension. D (3) The factor matrix representing the third dimension of space. D (4) The factor matrix representing the fourth dimension of space. D (5) The factor matrix representing the fifth dimension, color dimension. U1 and V1 are also The angular dimension, X1 and Y1 respectively represent The spatial dimension, X1 is less than or equal to X, Y1 is less than or equal to Y, and C1 represents The color dimension, C1 is less than or equal to C; since the factor matrices are usually orthogonal, when the core tensor When the dimensions U1 and V1, X1 and Y1, C1 are less than those of U and V, X and Y, C respectively, the core tensor can be considered as such. It is a compressed version of the light field image, thus achieving dimensionality reduction, and therefore the core tensor is then... Represented as: Here, the superscript "T" indicates the transpose of a vector or matrix; then the... Understanding Tucker decomposition from the perspective dimension as The factor matrix D with the first angular dimension (1) The factor matrix D in the second angular dimension (2) By multiplying along the pattern, the decomposed components are obtained. Then according to and get and make In The spatial dimension X1 equals X, Y1 equals Y, and the color dimension C1 equals C, thus achieving dimensionality reduction of the angular dimension.

[0029] Step 3: Set the viewpoint coordinates in the sub-aperture image array to... The sub-aperture image is taken as the central sub-aperture image, denoted as G. centre Then, G is calculated using existing technology. centre The depth map, denoted as D centre Then, the maximum inter-class algorithm is used in... and D is obtained from between centre Optimal threshold To D centre Perform foreground and background segmentation to obtain a binary image, denoted as D. centre,val Then D centre,val As weight and L 1,1 Combined, soon D centre,val With L 1,1 Multiply, and we get L 1,1 The foreground portion exhibiting obvious artifacts is denoted as L. 1,1,fore ; then L 1,1,fore The conversion from RGB space to the HSI model (Hue, Saturation, Intensity) yields L. 1,1,fore The hue component, saturation component, and intensity component are denoted as H. 1,1,fore S 1,1,fore and I 1,1,fore Finally, an 8-scale Difference of Gaussian (DOG) model, histograms, generalized Gaussian function (GGD), and moment matching-based methods were used to analyze I. 1,1,fore Edge structure features are extracted to obtain I. 1,1,fore The GGD fitted feature vector is denoted as Use histograms to analyze H 1,1,fore Color component feature extraction was performed to obtain H. 1,1,fore The order moment eigenvector is denoted as . Use histograms for I 1,1,fore Brightness features were extracted to obtain I. 1,1,fore The order moment eigenvector is denoted as . joint and The local feature vector extracted based on foreground and background segmentation is denoted as F1. in, The floor function is called the floor function. `min()` returns the minimum value, and `max()` returns the maximum value. D represents centre The minimum pixel value in D represents centre The maximum pixel value in Its dimension is 1×16. Its dimension is 1×4. The dimension of F1 is 1×4, and "[]" is a vector or matrix representation symbol. The dimension of F1 is 1×24. The eight scales of the Gaussian difference model refer to the standard deviation of the Gaussian filter. The Gaussian kernels at each scale are set to (0.4,0.5), (0.5,0.6), (0.6,0.7), (0.7,0.8), (0.8,0.9), (1.0,1.6), (1.6,2.56), and (2.56,4.096).

[0030] In this embodiment, in step 3, The acquisition process is as follows: I is obtained using an 8-scale Gaussian difference model. 1,1,fore Processing yields I 1,1,fore Eight edge maps at different scales were generated; then histograms were used to obtain I. 1,1,fore The coefficient distribution of the edge map at each scale is obtained and fitted using the shape and variance parameters of a generalized Gaussian function. Then, a moment-matching method is used to estimate the shape and variance parameters of the generalized Gaussian function, targeting I... 1,1,fore The shape parameter and variance parameter of the generalized Gaussian function corresponding to the Gaussian difference model at the j-th scale are denoted as β. j and σ j ; and then regarding I 1,1,fore By combining the shape and variance parameters of the generalized Gaussian function corresponding to the Gaussian difference model at eight scales, we obtain I. 1,1,fore GGD fitted feature vector Where 1≤j≤J, J=8, β I,GGD Indicates that it is aimed at I 1,1,fore The shape parameter vector of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model, β I,GGD =[β1,β2,…,β J ],β1,β2,…,β J Corresponding to I 1,1,foreThe shape parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the shape parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the shape parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model, σ I,GGD Indicates that it is aimed at I 1,1,fore The variance parameter vector σ of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model. I,GGD =[σ1,σ2,…,σ J ],σ1,σ2,…,σ J Corresponding to I 1,1,fore The variance parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the variance parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the variance parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model. The acquisition process is as follows: use histograms to analyze H 1,1,fore Processing yields H 1,1,fore The first, second, third, and fourth moments; then H 1,1,fore The first, second, third, and fourth moments are arranged in sequence to form H. 1,1,fore eigenvectors of the order moments The acquisition process is as follows: use histograms to analyze I 1,1,fore Processing yields I 1,1,fore The first, second, third, and fourth moments; then I 1,1,fore The first, second, third, and fourth moments are arranged in sequence to form I. 1,1,fore eigenvectors of the order moments

[0031] Step 4: Place L 1,1 The minimum, maximum, and median values ​​of grayscale images are denoted as p. graymin p graymax p graymed Then use the maximum inter-class algorithm on p graymin and p graymed L is obtained from the interval 1,1 The first optimal threshold In p graymed and p graymax L is obtained from the interval 1,1 The second optimal threshold In this way, L 1,1 The area is divided into three regions with different brightness intensities: a dark region, a normal exposure region, and a bright region, denoted as L. 1,1,low L 1,1,nor and L 1,1,highNext, L 1,1,low Transforming from RGB space to HSI model, we obtain L 1,1,low The hue component, saturation component, and intensity component are denoted as H. 1,1,low S 1,1,low and I 1,1,low Similarly, L 1,1,nor Transforming from RGB space to HSI model, we obtain L 1,1,nor The hue component, saturation component, and intensity component are denoted as H. 1,1,nor S 1,1,nor and I 1,1,nor ; will L 1,1,high Transforming from RGB space to HSI model, we obtain L 1,1,high The hue component, saturation component, and intensity component are denoted as H. 1,1,high S 1,1,high and I 1,1,high Finally, extract I. 1,1,low The feature vectors are used to describe brightness and texture information, denoted as . Extract I 1,1,high The feature vectors are used to describe texture information, denoted as . Extract H 1,1,high The influence of the halo is characterized by its feature vector, denoted as . Extract S 1,1,high The influence of the halo is characterized by its feature vector, denoted as . Extract I 1,1,nor The feature vectors are used to describe texture information, denoted as . joint and The resulting local feature vector, denoted as F2, is constructed based on brightness segmentation. in, L 1,1,low The grayscale values ​​in the grayscale image at p gray min and th1 L1,1 Between, L 1,1,nor The gray values ​​in the grayscale image and Between, L 1,1,high The gray values ​​in the grayscale image and p gray max between, Indicate I 1,1,low The mean of the pixel values ​​of all pixels. Indicate I 1,1,low The standard deviation of the pixel values ​​of all pixels. Indicate I 1,1,lowThe information entropy of the pixel values ​​of all pixels, Indicate I 1,1,high The information entropy of the pixel values ​​of all pixels, H represents 1,1,high The mean of the pixel values ​​of all pixels. H represents 1,1,high The standard deviation of the pixel values ​​of all pixels. S represents 1,1,high The mean of the pixel values ​​of all pixels. Indicate I 1,1,nor The information entropy of the pixel values ​​of all pixels is given by F2, which has a dimension of 1×8.

[0032] Step 5: Place L 1,1 Transforming from RGB space to HSI model, we obtain L 1,1 The hue component, saturation component, and intensity component are denoted as H. 1,1 S 1,1 and I 1,1 ; then extract I 1,1 The contrast characteristics are denoted as Then, using an 8-scale Difference of Gaussian (DOG) model, histograms, generalized Gaussian function (GGD), and a moment-matching-based method, I was extracted. 1,1 The GGD fitted feature vector is denoted as Finally, the joint and The feature vector extracted globally is denoted as F3. Where R represents the number of pixel gray levels, R = 256, γ i,k P represents the absolute difference between grayscale value i and grayscale value k. i,k This represents the probability that a pixel with grayscale value i and a pixel with grayscale value k appear as adjacent pixels. The dimension of F1 is 1×16, and the dimension of F3 is 1×17. The eight scales of the Gaussian difference model refer to the standard deviation of the Gaussian filter, and the Gaussian kernel at each scale is set to (0.4,0.5), (0.5,0.6), (0.6,0.7), (0.7,0.8), (0.8,0.9), (1.0,1.6), (1.6,2.56), (2.56,4.096).

[0033] In this embodiment, in step 5, an 8-scale Gaussian difference model is used to analyze I. 1,1 Processing yields I 1,1 Eight edge maps at different scales were generated; then, histograms were used to obtain I. 1,1The coefficient distribution of the edge map at each scale is obtained and fitted using the shape and variance parameters of a generalized Gaussian function. Then, a moment-matching method is used to estimate the shape and variance parameters of the generalized Gaussian function, targeting I... 1,1 The shape parameter and variance parameter of the generalized Gaussian function corresponding to the Gaussian difference model at the j-th scale are denoted as β. j 'and σ' j ; and then regarding I 1,1 By combining the shape and variance parameters of the generalized Gaussian function corresponding to the Gaussian difference model at eight scales, we obtain I. 1,1 GGD fitted feature vector Where 1≤j≤J, J=8, Indicates that it is aimed at I 1,1 The shape parameter vectors of the generalized Gaussian function corresponding to the eight scales of the Gaussian difference model. β′1,β′2,…,β' J Corresponding to I 1,1 The shape parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the shape parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the shape parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model. Indicates that it is aimed at I 1,1 The variance parameter vector of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model, σ I,GGD =[σ′1,σ'2,…,σ' J ],σ′1,σ′2,…,σ′ J Corresponding to I 1,1 The variance parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the variance parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the variance parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model.

[0034] Step 6: [The sentence is incomplete and requires more context to be translated accurately.] Except for L 1,1All remaining tensor slices (a total of U×V-1 tensor slices) are transformed from RGB space to the HSI model to obtain the hue, saturation, and intensity components of each remaining tensor slice. Then, a GLBP (Gradient Local Binary Pattern) algorithm based on 5 scales is used to extract gradient texture feature vectors at different scales for the intensity component of each remaining tensor slice. The dimension of the gradient texture feature vector at each scale is 1×10. Then, the average vector of the gradient texture feature vectors of the intensity component of each remaining tensor slice at the same scale is calculated. Finally, the average vectors of the 5 scales are used to construct the angle-based feature vector, denoted as F4, where the dimension of F4 is 1×50.

[0035] Step 7: Combine F1, F2, F3, and F4 to form The perceptual feature vector, denoted as F LGA ,F LGA = [F1, F2, F3, F4]; where F LGA The dimension is 1×99.

[0036] Step 8: F LGA As input, the LIBSVM package is used to train an SVR (Support Vector Regression) machine with a radial basis function (RBF) kernel, and the predictions are obtained. The objective quality score, denoted as Q. predict ; where Q predict The larger the value, the better the input F. LGA The better the quality of the corresponding multi-exposure fused light field image, the better; conversely, the worse the quality, the lower the quality of the input F-index. LGA The lower the quality of the corresponding multi-exposure fused light field image.

[0037] The following experiments demonstrate the effectiveness and feasibility of the method of the present invention.

[0038] This invention compares the method with eight classic IQA (Image Quality Assessment) methods, including two MEF-FRIQA methods and six LF-NRIQA methods. The MEF-FRIQA methods include MEF-SSIM (K. Ma, K. Zeng and Z. Wang, “Perceptual quality assessment for multi-exposure image fusion,” IEEE Trans. Image Process., vol. 24, no. 11, pp. 3345-3356, Nov. 2015).d(Y. Fang, H. Zhu, K. Ma, Z. Wang and S. Li, “Perceptual evaluation for multi-exposure image fusion of dynamic scenes,” IEEE Trans. Image Process., vol. 29, pp. 1127-1138, Sept. 2020. (Perceptual evaluation for multi-exposure image fusion of dynamic scenes)). The LF-NRIQA method includes NR-LFQA (L. Shi, W. Zhou, Z. Chen and J. Zhang, “No-reference light field imagequality assessment based on spatial-angular measurement,” IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 11, pp. 4114-4128, Nov. 2020. (No-reference light field image quality assessment based on spatial-angular measurement)) and BELIF (L. Shi, S. Zhao and Z. Chen, “BELIF: Blind quality evaluator of light field image with tensor structure variation”). index,” in Proc. Int. Conf. Image Process. ICIP, pp. 3781-3785, Aug. 2019 (Blind quality assessment of light field images based on tensor structure change index)); Tensor-NLFQ (W. Zhou, L. Shi, Z. Chen and J. Zhang, “Tensor-oriented no-reference light field image quality assessment,” IEEE Trans. Image Process., vol. 29, pp. 4070-4084, Feb. 2020. (Tensor-oriented no-reference light field image quality assessment)); TSSV-LFIQA (Z. Pan, M. Yu, G. Jiang, H. Xu and Y.-S. Ho, “Combining tensor slice and singular value for blind light field image quality assessment,” IEEE J. Sel. Top. Signal Process., vol. 15, no. 3, pp. 672-687, Apr.2021 (Blind light field image quality assessment based on tensor slices and singular values), PVRIs-LFIQA (J. Xiang, M. Yu, G. Jiang, H. Xu, Y. Song and Y.-S. Ho, “Pseudo video and refocused images-based blind light field imagequality assessment,” IEEE Trans. Circuits Syst. Video Technol., vol. 31, no. 7, pp. 2575-2590, July 2021 (Blind light field image quality assessment based on pseudo-video and refocused images)), and PM-BLFIQM (Y. Liu, G. Jiang, Z. Jiang, Z. Pan, M. Yu and Y.-S. Ho, “Pseudoreference subaperture images and microlens image-based blind light field image quality measurement,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1-15, July 2021). 2021 (Blind Light Field Image Quality Measurement Based on Pseudo-Reference Sub-Aperture Images and Microlens Images).

[0039] The multi-exposure fusion light field image database used for training and testing is the NBU-MLFD database. The predictive performance of the objective quality assessment method is evaluated by comparing the consistency between the objective quality score predicted by the objective quality assessment method and the MOS value. Specifically, the perceptual feature vector of the distorted multi-exposure fusion light field image extracted by the method of this invention obtains the objective quality score through a support vector regression machine; then, the objective quality score of the distorted multi-exposure fusion light field image is nonlinearly fitted with the MOS value; finally, three standard indicators provided by the video quality assessment expert group are used to quantify the predictive performance of different objective quality assessment methods. The three standard indicators are Spearman Rank Order Correlation Coefficient (SROCC), Pearson Linear Correlation Coefficient (PLCC), and Root Mean Square Error (RMSE). SROCC measures the predictive monotonicity of the objective quality assessment method, while PLCC and RMSE measure the predictive accuracy. The value of SROCC ranges from -1 to 1, and the value of PLCC ranges from 0 to 1. The closer the absolute values ​​of SROCC and PLCC are to 1, and the closer the RMSE is to 0, the better the predictive performance of the objective quality assessment method.

[0040] Table 1 lists the method of the present invention and its comparison with MEF-SSIM. d The performance metrics SROCC, PLCC, and RMSE of eight existing objective quality assessment methods—NR-LFQA, BELIF, Tensor-NLFQ, TSSV-LFIQA, PM-BLFIQM, and PVRIS-LFIQA—were calculated on the NBU-MLFD database.

[0041] Table 1. Comparison of the method of the present invention with MEF-SSIM and MEF-SSIM. d , NR-LFQA, BELIF, Tensor-NLFQ,

[0042] Eight existing objective quality assessment methods, including TSSV-LFIQA, PM-BLFIQM, and PVRIs-LFIQA, are used in...

[0043] Results of three performance metrics, SROCC, PLCC, and RMSE, on the NBU-MLFD database.

[0044]

[0045] As shown in Table 1, the method of this invention exhibits good prediction performance on the NBU-MLFD database. It is worth noting that the method of this invention also maintains advantages compared to the full-reference method.

Claims

1. A method for quality evaluation of dynamic scene multi-exposure fusion light field image, characterized in that comprising the steps of: Step 1: Let the distorted multi-exposure fusion light field image to be quality evaluated be denoted as wherein, represents is five-dimensional, U and V represent the angle dimensions of respectively, X and Y represent the spatial dimensions of respectively, and C represents the color dimension of ; then let be regarded as a 2-dimensional sub-aperture image array, the sub-aperture image array being composed of UxV sub-aperture images, each sub-aperture image having a resolution of XxY, and let the sub-aperture image with a view point coordinate position of (u, v) in the sub-aperture image array be denoted as G u,v , wherein 1≤u≤U and 1≤v≤V. Step 2: For Perform Tucker tensor decomposition on the angular dimension to obtain the decomposed components, denoted as . in, express It is five-dimensional, with U1 and V1 representing respectively. The angular dimension, U1 is less than or equal to U, V1 is less than or equal to V, and X and Y are also... The spatial dimension, C is also The number of color dimensions; then fix The first two dimensions U1 and V1 are obtained All tensor slices are represented as: in, L 1,1 This represents a tensor slice where the values ​​of U1 and V1 are fixed at 1 and 1 respectively. u,v L represents a tensor slice where the values ​​of U1 and V1 correspond to u and v, respectively. U,V This represents a tensor slice where the values ​​of U1 and V1 correspond to U and V respectively. Step 3: Take the sub-aperture image with the view point coordinate position as in the sub-aperture image array as the center sub-aperture image, denoted as G centre ; then calculate the depth map of G centre , denoted as D centre ; then obtain the optimal threshold of D centre between and using the maximum inter-class algorithm to perform foreground-background segmentation on D centre to obtain a binary image, denoted as D centre,val ; then combine D centre,val as a weight with L 1,1 , i.e. multiply D centre,val and L 1,1 to obtain the foreground part of L 1,1 with obvious artifacts, denoted as L 1,1,fore ; then convert L 1,1,fore from the RGB space to the HSI model to obtain the hue component, saturation component and intensity component of L 1,1,fore , denoted as H 1,1,fore , S 1,1,fore and I 1,1,fore ; finally, use an 8-scale Gaussian difference model, a histogram, a generalized Gaussian function and a method based on matrix matching to extract the edge structure features of I 1,1,fore to obtain the GGD fitting feature vector of I 1,1,fore , denoted as use the histogram to extract the color component features of H 1,1,fore to obtain the moment feature vector of H 1,1,fore , denoted as use the histogram to extract the brightness features of I 1,1,fore to obtain the moment feature vector of I 1,1,fore , denoted as combine and to form the local feature vector extracted based on foreground-background segmentation, denoted as F1, wherein is a down rounding operator, min() is a minimum value function, max() is a maximum value function, represents the minimum pixel value in D centre , represents the maximum pixel value in D centre , has a dimension of 1x16, has a dimension of 1x4, has a dimension of 1x4, and "[]" is a vector or matrix representation symbol, and F1 has a dimension of 1x24. Step 4: Place L 1,1 The minimum, maximum, and median values ​​of grayscale images are denoted as p. graymin p graymax p graymed Then use the maximum inter-class algorithm on p graymin and p graymed L is obtained from the interval 1,1 The first optimal threshold In p graymed and p graymax L is obtained from the interval 1,1 The second optimal threshold In this way, L 1,1 The area is divided into three regions with different brightness intensities: a dark region, a normal exposure region, and a bright region, denoted as L. 1,1,low L 1,1,nor and L 1,1,high Next, L 1,1,low Transforming from RGB space to HSI model, we obtain L 1,1,low The hue component, saturation component, and intensity component are denoted as H. 1,1,low S 1,1,low and I 1,1,low Similarly, L 1,1,nor Transforming from RGB space to HSI model, we obtain L 1,1,nor The hue component, saturation component, and intensity component are denoted as H. 1,1,nor S 1,1,nor and I 1,1,nor ; will L 1,1,high Transforming from RGB space to HSI model, we obtain L 1,1,high The hue component, saturation component, and intensity component are denoted as H. 1,1,high S 1,1,high and I 1,1,high Finally, extract I. 1,1,low The feature vectors are used to describe brightness and texture information, denoted as . Extract I 1,1,high The feature vectors are used to describe texture information, denoted as . Extract H 1,1,high The influence of the halo is characterized by its feature vector, denoted as . Extract S 1,1,high The influence of the halo is characterized by its feature vector, denoted as . Extract I 1,1,nor The feature vectors are used to describe texture information, denoted as . joint and The resulting local feature vector, denoted as F2, is constructed based on brightness segmentation. Among them, L 1,1,low The grayscale values ​​in the grayscale image at p graymin and Between, L 1,1,nor The gray values ​​in the grayscale image and Between, L 1,1,high The gray values ​​in the grayscale image and p graymax between, Indicate I 1,1,low The mean of the pixel values ​​of all pixels. Indicate I 1,1,low The standard deviation of the pixel values ​​of all pixels. Indicate I 1,1,low The information entropy of the pixel values ​​of all pixels, Indicate I 1,1,high The information entropy of the pixel values ​​of all pixels, H represents 1,1,high The mean of the pixel values ​​of all pixels. H represents 1,1,high The standard deviation of the pixel values ​​of all pixels. S represents 1,1,high The mean of the pixel values ​​of all pixels. Indicate I 1,1,nor The information entropy of the pixel values ​​of all pixels, and the dimension of F2 is 1×8; Step 5: L 1,1 Transforming from RGB space to HSI model, get the hue component, saturation component and intensity component of L 1,1 , denoted as H 1,1 , S 1,1 and I 1,1 respectively; then extract the contrast feature of I 1,1 , denoted as Then using 8 scales of Gaussian difference model, histogram, generalized Gaussian function and the method based on matrix matching, get the GGD fitting feature vector of I 1,1 , denoted as Finally, combine F and F to form the feature vector based on global extraction, denoted as F3, wherein R represents the number of pixel gray levels, R = 256, γ i,k represents the absolute difference value of gray value i and gray value k, P i,k represents the probability of adjacent occurrence of the pixel with gray value i and the pixel with gray value k, the dimension of F is 1 × 16, and the dimension of F3 is 1 × 17. Step 6: [The sentence is incomplete and requires more context to be translated accurately.] Except for L 1,1 All remaining tensor slices are transformed from RGB space to HSI model to obtain the hue, saturation, and intensity components of each tensor slice. Then, a GLBP algorithm based on 5 scales is used to extract gradient texture feature vectors at different scales for the intensity component of each remaining tensor slice. The dimension of the gradient texture feature vector at each scale is 1×10. Then, the average vector of the gradient texture feature vectors of the intensity component of each remaining tensor slice at the same scale is calculated. Finally, the average vectors of the 5 scales are used to construct the feature vector based on angle extraction, denoted as F4. The dimension of F4 is 1×50. Step 7: combine F1, F2, F3, F4 to form the perceptual feature vector, denoted as F LGA ,F LGA = [F1, F2, F3, F4]; where F LGA has dimension 1x99;​ Step 8: F LGA As input, the LIBSVM package is used to train the SVR with a radial basis function kernel for regression training, and the predictions are obtained. The objective quality score, denoted as Q. predict ; where Q predict The larger the value, the better the input F. LGA The better the quality of the corresponding multi-exposure fused light field image, the better; conversely, the worse the quality, the lower the quality of the input F-index. LGA The lower the quality of the corresponding multi-exposure fused light field image.

2. The method of claim 1, wherein In step 2, the following will be done: The Tucker tensor decomposition is represented as: In the Tucker tensor decomposition process It is regarded as a fifth-order tensor. Represents the core tensor. D (1) This represents the factor matrix of the first angular dimension. D (2) This represents the factor matrix of the second angular dimension. D (3) The factor matrix representing the third dimension of space. D (4) The factor matrix representing the fourth dimension of space. D (5) The factor matrix representing the fifth dimension, color dimension. U1 and V1 are also The angular dimension, X1 and Y1 respectively represent The spatial dimension, X1 is less than or equal to X, Y1 is less than or equal to Y, and C1 represents The color dimension, C1 is less than or equal to C; then the core tensor Represented as: Here, the superscript "T" indicates the transpose of a vector or matrix; then the... Understanding Tucker decomposition from the perspective dimension as The factor matrix D with the first angular dimension (1) The factor matrix D in the second angular dimension (2) By multiplying along the pattern, the decomposed components are obtained. Then according to and get and make In The spatial dimension X1 equals X, Y1 equals Y, and the color dimension C1 equals C, thus achieving dimensionality reduction of the angular dimension.

3. The method of claim 1 or 2, wherein The step 3 is as follows: The acquisition process is as follows: using 8-scale Gaussian difference model to process I 1,1,fore to obtain 8 different scale edge maps of I 1,1,fore ; then using histogram to obtain the coefficient distribution of each scale edge map of I 1,1,fore and using the shape parameter and variance parameter of generalized Gaussian function to fit, and using the method based on matrix matching to estimate the shape parameter and variance parameter of generalized Gaussian function, the shape parameter and variance parameter of generalized Gaussian function corresponding to the jth scale of Gaussian difference model of I 1,1,fore are denoted as β j and σ j ; then the shape parameter and variance parameter of generalized Gaussian function corresponding to the 8 scales of Gaussian difference model of I 1,1,fore are combined to obtain the GGD fitting feature vector of I 1,1,fore Wherein, 1≤j≤J, J=8, β I,GGD denotes the shape parameter vector of generalized Gaussian function corresponding to the 8 scales of Gaussian difference model of I 1,1,fore , β I,GGD =[β1, β2, …, β J ], β1, β2, …, β J correspond to the shape parameters of generalized Gaussian function corresponding to the 1st scale of Gaussian difference model of I 1,1,fore , the 2nd scale of Gaussian difference model, …, the Jth scale of Gaussian difference model, σ I,GGD denotes the variance parameter vector of generalized Gaussian function corresponding to the 8 scales of Gaussian difference model of I 1,1,fore , σ I,GGD =[σ1, σ2, …, σ J ], σ1, σ2, …, σ J correspond to the variance parameters of generalized Gaussian function corresponding to the 1st scale of Gaussian difference model of I 1,1,fore , the 2nd scale of Gaussian difference model, …, the Jth scale of Gaussian difference model.​ The acquisition process is as follows: use histograms to analyze H 1,1,fore Processing yields H 1,1,fore The first, second, third, and fourth moments; then H 1,1,fore The first, second, third, and fourth moments are arranged in sequence to form H. 1,1,fore eigenvectors of the order moments The acquisition process is as follows: use histograms to analyze I 1,1,fore Processing yields I 1,1,fore The first, second, third, and fourth moments; then I 1,1,fore The first, second, third, and fourth moments are arranged in sequence to form I. 1,1,fore eigenvectors of the order moments 4. The method of claim 3, wherein In step 5, an 8-scale Gaussian difference model is used to analyze I. 1,1 Processing yields I 1,1 Eight edge maps at different scales were generated; then, histograms were used to obtain I. 1,1 The coefficient distribution of the edge map at each scale is obtained and fitted using the shape and variance parameters of a generalized Gaussian function. Then, a moment-matching method is used to estimate the shape and variance parameters of the generalized Gaussian function, targeting I. 1,1 The shape parameter and variance parameter of the generalized Gaussian function corresponding to the Gaussian difference model at the j-th scale are denoted as β'. j and σ' j ; and then regarding I 1,1 By combining the shape and variance parameters of the generalized Gaussian function corresponding to the Gaussian difference model at eight scales, we obtain I. 1,1 GGD fitted feature vector Where 1≤j≤J, J=8, Indicates that it is aimed at I 1,1 The shape parameter vectors of the generalized Gaussian function corresponding to the eight scales of the Gaussian difference model. β'1,β'2,…,β' J Corresponding to I 1,1 The shape parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the shape parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the shape parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model. Indicates that it is aimed at I 1,1 The variance parameter vector of the generalized Gaussian function corresponding to the eight-scale Gaussian difference model, σ I,GGD =[σ'1,σ'2,…,σ' J ],σ'1,σ'2,…,σ' J Corresponding to I 1,1 The variance parameters of the generalized Gaussian function corresponding to the first-scale Gaussian difference model, the variance parameters of the generalized Gaussian function corresponding to the second-scale Gaussian difference model, ..., the variance parameters of the generalized Gaussian function corresponding to the J-th-scale Gaussian difference model.

Images