A method for evaluating the quality of light field images without a reference light field

Abstract

The present invention provides a method for evaluating the quality of a light field image without a reference light field. The method includes: Step 1: Obtain a color light field image to be evaluated, and perform tensor decomposition on the color light field image to obtain three 4D components, where the three 4D components include one luminance component and two chrominance components; Step 2: Rearrange each component into a 3D pseudo-video sequence, and calculate the consecutive frame differences of each pseudo-video; Step 3: Use a feature extraction method based on 3D shearlet transform to extract features from the consecutive frame differences of the pseudo-video; Step 4: Pool and regress the features extracted in Step 3 to obtain the quality score of the color light field image. The present invention takes into account the problems brought about by the high-dimensionality of the light field image and can more accurately predict the quality of the light field image.

Description

Technical Field

[0001] The present invention relates to the technical field of image quality evaluation, and particularly to a method for evaluating the quality of a reference-free light field image. Background Art

[0002] Light field imaging is one of the most promising technologies in the field of computational imaging. Light field imaging can capture the intensity and direction information of light in the free space of the real world, and thus has received increasing attention. With the wide application of light fields, the quality of light field imaging has also become an important factor affecting people's viewing of light field images. During the acquisition, transmission, decoding, and display processes of light field images, distortions are introduced due to differences in processing algorithms. These distortions will lead to the degradation of the quality of light field images, which on the one hand affects the applications of post-processing of light field images, such as refocusing, de-occlusion, and depth estimation, etc.; on the other hand, low-quality light field images will provide users with an uncomfortable visual experience. In order to guide and optimize light field image processing algorithms and provide better services to users, it is of great research significance to objectively and quantitatively evaluate the quality of light field images.

[0003] Traditional planar image quality evaluation methods have been relatively mature, and these methods can already accurately evaluate the quality of planar images. Light field images record not only the spatial information of the scene but also the angular information of the scene. The quality of light field images is mainly restricted by spatial quality and angular consistency. Therefore, planar image quality evaluation methods cannot accurately estimate the quality of light field images.

[0004] In order to accurately evaluate the quality of light field images, Tian et al. designed a multi-derivative feature model (MDFM) to measure the similarity of the first-order and second-order structures between each viewpoint of the original and distorted light field images. Min et al. measured the quality of distorted light field images from three aspects: global space, local space, and angular quality. Shi et al. obtained the spatial quality and angular consistency of light field images based on catadioptric image arrays and epipolar plane images. Zhou et al. extracted the spatial quality and angular consistency of light field images from the viewpoint stack of light field images. Xiang et al. quantified the quality of light field images based on pseudo-videos and refocused images. However, the above-mentioned light field image quality evaluation methods usually assume that the human eye is more sensitive to the brightness of images, while ignoring the exploration of the chromaticity information of light field images. Light field images are high-dimensional data, and color light field images can be defined as 5D functions. Existing color space transformations are difficult to process 5D color data. In addition, the high-dimensional characteristics of light field images make it necessary to comprehensively consider spatial quality and angular consistency when evaluating the quality of light field images. Most methods use multi-stage feature extraction separately to measure the deterioration of the spatial quality and angular consistency of light field images.

[0005] In summary, it is an urgent problem for those skilled in the art to provide a no-reference light field image quality evaluation method that can effectively solve the problems brought by the high-dimensionality of light field images and more accurately predict the quality of light field images. Summary of the Invention

[0006] In view of the above problems and requirements, the present solution proposes a no-reference light field image quality evaluation method, which can solve the above technical problems because of the following technical solutions.

[0007] To achieve the above object, the present invention provides the following technical solution: A no-reference light field image quality evaluation method, comprising: Step Step1: Obtain a color light field image to be evaluated, and perform tensor decomposition on the color light field image to obtain three 4D components, where the three 4D components include one luminance component and two chrominance components;

[0008] Step Step2: Rearrange each component into a 3D pseudo-video sequence, and calculate the consecutive frame differences of each pseudo-video;

[0009] Step Step3: Use a feature extraction method based on 3D shearlet transform to extract features from the consecutive frame differences of the pseudo-video;

[0010] Step Step4: Pool and regress the features extracted in Step Step3 to obtain the quality score of the color light field image.

[0011] Further, the tensor decomposition is performed by Tucker decomposition along the color dimension of the 5D color light field image.

[0012] Even further, the decomposition along the color dimension of the 5D color light field image includes:

[0013] Decompose the 5D color light field image according to the formula:

[0014]

[0015] Decompose into a core tensor and multiple orthogonal matrices, where, in the formula represents the color light field image, represents the 5D core tensor, U (i) , i = 1, 2,..., 5 are orthogonal factor matrices, and × i represents the i-mode product;

[0016] Then obtain the decomposition components of the color dimension: where, in the formula represents the decomposition component along the color dimension of the color light field image, and T represents the transpose operation;

[0017] Finally, Define the luminance component Λ of the color light field image F , and Define the chrominance components Λ of the color light field image respectively S and Λ H .

[0018] Furthermore, the step of rearranging each component into a 3D pseudo-video sequence and calculating the consecutive frame differences of each pseudo-video includes:

[0019] Regard each viewpoint in each component as a frame, and arrange each viewpoint of the light field image into the form of a 3D pseudo-video according to the set rules. The pseudo-video of each component is defined as Γ F 、Γ S and Γ H ;

[0020] Then according to the formula: Calculate the amplitude of the difference between adjacent frames in each pseudo-video, where represents the image of the t i th frame, || represents the absolute value operation, represents the amplitude of the difference between the t i th frame and the t i+1 th frame;

[0021] The amplitudes of the differences between all adjacent frames in each pseudo-video constitute the consecutive difference frames of the corresponding pseudo-video. Define the consecutive difference frames of the pseudo-videos of the three components as and

[0022] Furthermore, the steps of extracting features from the consecutive frame differences of the pseudo-video by using the feature extraction method based on 3D shearlet transform include:

[0023] Obtain the 3D blocks in the consecutive difference frames of the pseudo-videos of the three components and , and use the 3D shearlet transform method to perform 3D shearlet transform on each 3D block obtained in and ;

[0024] After the transform, shearlet coefficients in different directions and different scales of each 3D block can be obtained, including a total of 1 low-frequency subband coefficient and 26 high-frequency subband coefficients, which are respectively marked as and where b represents the bth 3D block, and the value range of b is related to the number of specifically segmented 3D blocks. l represents the lth subband coefficient. When l = 1, it represents the low-frequency subband coefficient, and l = 2,..., 27 represents the high-frequency subband coefficients;

[0025] Finally, three types of statistical features are extracted, and the three types of statistical features include the first type of feature, the second type of feature, and the third type of feature.

[0026] Furthermore, the Prewitt operator is used to obtain the spatial dimension of the low-frequency subband coefficients of the corresponding three-dimensional blocks of different components and the corresponding gradient and Then calculate and with and the similarity value of each pixel point, and calculate the average value to quantify the similarity between the gradient of the low-frequency subband coefficients and the high-frequency subband coefficients:

[0027]

[0028] where (s 1 , t 1 , z 1 ) represents the pixel coordinates in the three-dimensional block, and represent the similarity values of the three components in the b-th block; then calculate the average value of the similarity values of each three-dimensional block and denote it as T l F , T l S and T l H , and connect all the eigenvalue into a one-dimensional vector to obtain the first type of feature F Y , where, [] means concatenating scalars or vectors.

[0029] Furthermore, the 26 high-frequency subband coefficients of each block in the three components are respectively expanded into one-dimensional vectors; use the formula: to perform generalized Gaussian distribution fitting on each vector, where y represents the high-frequency subband coefficients vectorized into one-dimensional vectors Γ() represents the gamma function; obtain the generalized Gaussian distribution fitting parameters α and σ of the high-frequency subband coefficients corresponding to each three-dimensional block of each component, and average the generalized Gaussian distribution fitting parameters α and σ of the high-frequency subband coefficients of all three-dimensional blocks of each component respectively as the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of each component, where the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of the luminance component include and the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of the two chrominance components respectively include and And Finally, the fitting parameters of the generalized Gaussian distribution of the same kind with different components are cascaded into a one-dimensional vector, that is And The second type of feature F is obtained α And the third type of feature F β .

[0030] Furthermore, the pooling and regression of the features extracted in step Step3 to obtain the quality score of the color light field image include: for the features F Y , F α And F σ Perform principal component analysis dimensionality reduction to obtain the cropped features F′ Y , F′ α And F′ σ , their dimensions are 32, 24, and 16 respectively, and they are concatenated into a one-dimensional feature vector Then the feature vector F LFIQA Is input into the trained support vector regression with a radial basis function as the kernel to regress and obtain the quality score Q of the distorted light field image predict .

[0031] As can be seen from the above technical solutions, the beneficial effects of the present invention are as follows:

[0032] 1. The method proposed by the present invention is a no-reference light field image quality evaluation method. This method does not require the original light field image and can directly obtain the quality score of the image by using the distorted image as the input, and can be used under certain conditions where the original light field image cannot be obtained.

[0033] 2. The color space transformation based on tensor decomposition proposed by the present invention, based on the properties of tensor decomposition, can decompose the color dimension of any high-dimensional color data and obtain the corresponding luminance and chrominance components.

[0034] 3. The present invention transforms the 4D light field image into a 3D pseudo-video, and can obtain the spatial and angular information of the light field image without using high-dimensional transformation; and designs three statistical features to effectively represent the quality of the light field image.

[0035] In addition to the purposes, features, and advantages described above, the optimal embodiments of the present invention will be described in more detail below in conjunction with the accompanying drawings, so as to easily understand the features and advantages of the present invention. Description of the Drawings

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments of the present invention or the prior art. Among them, the drawings are only used to show some embodiments of the present invention, rather than limiting all embodiments of the present invention thereto.

[0037] Figure 1 It is a schematic diagram of the specific steps of a method for evaluating the quality of a light field image without a reference light field according to the present invention.

[0038] Figure 2 It is a schematic diagram of the specific steps for calculating the consecutive frame differences of each pseudo-video in this embodiment.

[0039] Figure 3 It is a schematic diagram of the specific steps for feature extraction of the consecutive frame differences of the pseudo-video in this embodiment.

[0040] Figure 4 It is a schematic diagram of the specific steps for pooling and regression of the features of the extracted consecutive frame differences in this embodiment.

[0041] Figure 5 It is a schematic diagram of the arrangement order for reshaping the 4D light field image data into 3D pseudo-video data in this embodiment. Detailed implementation manners

[0042] In order to make the objectives, technical solutions and advantages of the technical solutions of the present invention clearer, the following will clearly and completely describe the technical solutions of the embodiments of the present invention with reference to the drawings of the specific embodiments of the present invention. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] First, based on tensor decomposition, the 5D color light field image is decomposed along the color dimension to obtain a luminance component and two chrominance components, both of which are 4D data. Then, each component is reshaped into a 3D pseudo-video, and the difference between adjacent frames is calculated to obtain the continuous difference frames of the pseudo-video. Subsequently, the three-dimensional shearlet transform is used to transform the continuous difference frames of the pseudo-video into the shearlet domain to extract three types of statistical features to represent the distortion of the light field image. Finally, the features of the same category between each component are cropped through principal component analysis, and the cropped features are concatenated into a one-dimensional feature vector, and the quality score of the light field image is obtained through the support vector regression technique. Considering that traditional color space transformation cannot directly perform color conversion on high-dimensional color data, a method based on tensor decomposition is proposed to decompose the color dimension of high-dimensional data to obtain the luminance and chrominance information of high-dimensional data. To simultaneously capture the spatial quality and angular consistency of the light field image, the light field image is also arranged in the form of a 3D pseudo-video, and the three-dimensional shearlet transform is used to extract the spatial and angular information of the light field image.

[0044] As Figures 1 to 5 shown, a no-reference light field image quality assessment method based on tensor decomposition and three-dimensional shearlet transform specifically includes the following steps: Step Step1: Obtain the color light field image to be evaluated, and perform tensor decomposition on the color light field image to obtain three 4D components, where the three 4D components include a luminance component and two chrominance components. Among them, the tensor decomposition is performed along the color dimension of the 5D color light field image using the Tucker decomposition method.

[0045] Specifically, the decomposition along the color dimension of the 5D color light field image includes:

[0046] Decompose the 5D color light field image according to the formula:

[0047]

[0048] into a core tensor and multiple orthogonal matrices, where in the formula represents the color light field image, represents the 5D core tensor, U (i) , i = 1, 2,..., 5 are orthogonal factor matrices, and × i represents the i-mode product;

[0049] Then, obtain the decomposition components of the color dimension: where in the formula represents the decomposition component along the color dimension of the color light field image, and T represents the transpose operation;

[0050] Finally, define as the luminance component Λ F of the color light field image, and are respectively defined as the chromaticity components Λ of the color light field image S and Λ H , both of which are 4D tensors.

[0051] In this embodiment, for a given 5D color light field image using tensor decomposition, specifically Tucker decomposition in tensor decomposition, the color light field image is decomposed into a core tensor and multiple orthogonal factor matrices, where (u, v) represents the angular coordinates of the viewpoints in the light field image, (s, t) represents the spatial coordinates of the viewpoints, and c represents the three color channels.

[0052] Step Step2: Rearrange each component into a 3D pseudo-video sequence and calculate the consecutive frame differences of each pseudo-video.

[0053] The specific steps of Step Step2 include:

[0054] Step Step2.1: Treat each viewpoint in each component as a frame, and arrange each viewpoint of the light field image into the form of a 3D pseudo-video according to the rules as Figure 5 shown. The pseudo-videos of each component are defined as Γ F , Γ S and Γ H ; as Figure 5 shown, the color light field image is regarded as a 5D tensor. When ignoring the color dimension, the light field image can be regarded as a 4D tensor. The color light field image is regarded as a 5D tensor. When ignoring the color dimension, the light field image can be regarded as a 4D tensor, denoted as Λ = {Λ(u, v, s, t)}. The light field image can be regarded as a multi-view array, where (s, t) represents the spatial coordinates, which are the spatial positions of the pixels in an image, and (u, v) are the angular coordinates, representing the angular positions of the image.

[0055] Step Step2.2: Then according to the formula: calculate the magnitude of the difference between adjacent frames in each pseudo-video, where, represents the image of the t i th frame, || represents the absolute value operation, represents the magnitude of the difference between the t i th frame and the t i+1 th frame;

[0056] Step Step2.3: The magnitudes of the differences between all adjacent frames in each pseudo-video constitute the consecutive difference frames of the corresponding pseudo-video. The consecutive difference frames of the pseudo-videos of the three components are respectively defined as and

[0057] Step Step3: Use a feature extraction method based on three-dimensional shearlet transform to extract features from the consecutive frame differences of the pseudo-video.

[0058] The specific steps of Step Step3 are as follows:

[0059] Step Step3.1: Obtain the consecutive difference frames of the pseudo-video of three components and in the three-dimensional blocks, and use the three-dimensional shearlet transform method to perform three-dimensional shearlet transform on each three-dimensional block obtained from and ;

[0060] In this embodiment, the process of obtaining three-dimensional blocks is to divide the consecutive difference frames of the pseudo-video of three components and into three-dimensional blocks of size 80×80×80. Whether there is overlap between blocks is determined according to and the original resolution.

[0061] Step Step3.2: After the transformation, shearlet coefficients of different directions and different scales of each three-dimensional block can be obtained, including a total of 1 low-frequency subband coefficient and 26 high-frequency subband coefficients, which are respectively marked as and where b represents the b-th three-dimensional block, and the value range of b is related to the number of specifically segmented three-dimensional blocks. l represents the l-th subband coefficient. When l = 1, it represents the low-frequency subband coefficient, and l = 2,..., 27 represents the high-frequency subband coefficients;

[0062] Step Step3.3: Finally, extract three types of statistical features, and the three types of statistical features include the first type of feature, the second type of feature, and the third type of feature.

[0063] The process of obtaining the first type of feature includes: using the Prewitt operator to obtain the spatial dimensions of the low-frequency subband coefficients of the corresponding three-dimensional blocks of different components

[0064] and the corresponding gradients

[0065] and Then calculate and with and the similarity values of each pixel point, and obtain the average value to quantify the similarity between the gradient of the low-frequency subband coefficient and the high-frequency subband coefficient:

[0066]

[0067] Among them, (s 1 , t 1 , z 1 ) represents the pixel coordinates in the three-dimensional block, and represent the similarity values of the three components in the b-th block; then, the average value of the similarity values of each three-dimensional block is calculated and denoted as T l F , T l S and T l H , and all the eigenvalue vectors are concatenated into a one-dimensional vector to obtain the first type of feature f Y , Among them, [] represents concatenating scalars or vectors.

[0068] The process of obtaining the second and third types of features includes: expanding the 26 high-frequency subband coefficients of each block in the three components into one-dimensional vectors; using the formula: to perform generalized Gaussian distribution fitting on each vector, where y represents the high-frequency subband coefficients vectorized into a one-dimensional vector Γ() represents the gamma function; the generalized Gaussian distribution fitting parameters α and σ of the high-frequency subband coefficients corresponding to each three-dimensional block of each component are obtained, and the generalized Gaussian distribution fitting parameters α and σ of the high-frequency subband coefficients of all three-dimensional blocks of each component are respectively averaged as the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of each component. Among them, the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of the luminance component include and The generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of the two chrominance components respectively include and and Finally, the same-type generalized Gaussian distribution fitting parameters of different components are concatenated into a one-dimensional vector, that is and to obtain the second type of feature F α and the third type of feature F β .

[0069] Step Step4: Pool and regress the features extracted in Step Step3 to obtain the quality score of the color light field image.

[0070] The specific steps of Step Step4 include:

[0071] Step Step4.1: Perform principal component analysis dimensionality reduction on the same-type features F Y , F α and F σ of each component to obtain the cropped features F′ Y , F′ α and F′σ , their dimensions are 32, 24, and 16 respectively, and they are concatenated into a one-dimensional feature vector F LFIQA = [F′ Y , F′ α , F′ β ; Step Step4.2: Then the feature vector F LFIQA is input into the trained support vector regression with a radial basis function as the kernel to regress and obtain the quality score Q of the distorted light field image predict .

[0072] In this embodiment, a computer with an Intel Core-i5 central processing unit and 8G bytes of memory is used, and a no-reference light field image quality evaluation method based on tensor decomposition and three-dimensional shearlet transform is programmed in Matlab language to implement the method of the present invention.

[0073] The specific operating hardware and programming language of the method of this application are not limited, and it can be completed in any language.

[0074] It should be noted that the embodiments described in the present invention are only the preferred ways to implement the present invention. Any obvious modifications that belong to the overall concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A no-reference light field image quality evaluation method, characterized in that, it includes the following steps: Step 1: Obtain the color light field image to be evaluated, and perform tensor decomposition on the color light field image to obtain three 4D components, where the three 4D components include one luminance component and two chrominance components; Step 2: Rearrange each component into a 3D pseudo-video sequence, and calculate the consecutive frame differences of each pseudo-video; Step 3: Use a feature extraction method based on 3D shearlet transform to extract features from the consecutive frame differences of the pseudo-video, and finally extract three types of statistical features, where the three types of statistical features include the first type of feature, the second type of feature, and the third type of feature; Step 4: Perform pooling and regression on the features extracted in Step 3 to obtain the quality score of the color light field image.

2. The no-reference light field image quality evaluation method according to claim 1, characterized in that, the tensor decomposition is performed by Tucker decomposition along the color dimension of the 5D color light field image.

3. The no-reference light field image quality evaluation method according to claim 2, characterized in that, the decomposition along the color dimension of the 5D color light field image includes: decompose the 5D color light field image according to the formula: Decompose into a core tensor and multiple orthogonal matrices, where, in the formula represents a color light field image, represents a 5-dimensional core tensor, and U (i) , i = 1, 2,..., 5 are orthogonal matrices, and × i represents the i-mode product; Obtain the decomposed components of the color dimension again: wherein, in the formula represents the decomposed components along the color dimension of the color light field image, and T represents the transpose operation; Finally, is defined as the luminance component Λ of the color light field image F , and are respectively defined as the chrominance components Λ S and Λ H .

4. The no-reference light field image quality evaluation method according to claim 1, characterized in that, the rearrangement of each component into a 3D pseudo-video sequence and the calculation of the consecutive frame differences of each pseudo-video include: Regarding each viewpoint in each component as a frame, arranging each viewpoint of the light field image in the form of a 3D pseudo-video according to the set rules, and the pseudo-video of each component is defined as Γ F , Γ S and Γ H ; Then, according to the formula: calculate the magnitude of the difference between adjacent frames in each pseudo-video, where represents the image of the t i -th frame, || represents the absolute value operation, represents the magnitude of the difference between the t i -th frame and the t i+1 +1-th frame; The magnitudes of the differences between all adjacent frames in each pseudo-video constitute the continuous difference frames corresponding to the pseudo-video, and the continuous difference frames of the pseudo-videos of the three components are respectively defined as and 5. The no-reference light field image quality evaluation method according to claim 4, characterized in that, the use of a feature extraction method based on 3D shearlet transform to extract features from the consecutive frame differences of the pseudo-video includes: Obtain the pseudo-video consecutive difference frames of three components and the three-dimensional blocks in, and use the three-dimensional shearlet transform method to perform three-dimensional shearlet transform on each three-dimensional block obtained in and ; After transformation, shearlet coefficients with different directions and different scales of each 3D block can be obtained, including a total of 1 low-frequency subband coefficient and 26 high-frequency subband coefficients, which are respectively labeled as and where b represents the b-th 3D block, and the value range of b is related to the number of specifically segmented 3D blocks. l represents the l-th subband coefficient. When l = 1, it represents the low-frequency subband coefficient, and when l = 2,..., 27, it represents the high-frequency subband coefficients; finally extract three types of statistical features, where the three types of statistical features include the first type of feature, the second type of feature, and the third type of feature.

6. The no-reference light field image quality evaluation method according to claim 5, characterized in that, Use the Prewitt operator to obtain the spatial dimension of the low-frequency subband coefficients of the 3D blocks corresponding to different components and the corresponding gradient and Then calculate and with and the similarity value of each pixel point, and calculate the average value to quantify the similarity between the gradient of the low-frequency subband coefficients and the high-frequency subband coefficients: Among them, (s 1 , t 1 , z 1 ) represents the pixel coordinates in a three-dimensional block, and represent the similarity values of three components in the b-th block; then the average value of the similarity values of each three-dimensional block is obtained and denoted as T l F , T l S and T l H , and all the eigenvalue are concatenated into a one-dimensional vector to obtain the first type of feature F Y , Among them, [] represents concatenating scalars or vectors.

7. The no-reference light field image quality evaluation method according to claim 5, characterized in that, Expand the 26 high-frequency subband coefficients of each three-dimensional block in the three components into one-dimensional vectors respectively; use the formula: Perform generalized Gaussian distribution fitting on each vector, where y represents the high-frequency subband coefficients vectorized into one-dimensional vectors Γ() represents the gamma function; obtain the generalized Gaussian distribution fitting parameters α and σ of the high-frequency subband coefficients corresponding to each three-dimensional block of each component, and average the generalized Gaussian distribution fitting parameters α and σ of the high-frequency subband coefficients of all three-dimensional blocks of each component respectively as the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of each component, where the generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of the luminance component include and The generalized Gaussian distribution fitting parameters of the high-frequency subband coefficients of the two chrominance components respectively include and and Finally, cascade the same-kind generalized Gaussian distribution fitting parameters of different components into one-dimensional vectors, that is and Obtain the second type of feature F α and the third type of feature F β .

8. The no-reference light field image quality evaluation method according to claim 1, characterized in that, Pooling and regression of the features extracted in Step 3 to obtain the quality score of the color light field image includes: for the features F of each same-component class Y 、F α and F σ perform principal component analysis for dimensionality reduction to obtain the cropped features F′ Y 、F′ α and F′ σ , their dimensions are 32, 24, and 16 respectively, and connect them into a one-dimensional feature vector Then input the feature vector F LFIQA into the pre-trained support vector regression with a radial basis function as the kernel to regress and obtain the quality score Q of the distorted light field image predict .

Images