A method for evaluating quality of light field image based on spatial and angular measurement

Abstract

The application provides a light field image quality evaluation method based on space and angle measurement, and the method comprises the following steps: Step 1: a multi-band local binary pattern algorithm is used to extract the spatial features of a light field image from a sub-aperture image array of the 4D light field image; Step 2: a feature extraction algorithm based on entropy-weighted local phase quantization is used to extract the features of a microlens image array of the 4D light field image as the angle features of the light field image; Step 3: the spatial features of the light field image and the angle features of the light field image are fused to obtain a one-dimensional feature vector; and Step 4: a support vector regression pooling operation is performed on the one-dimensional feature vector to obtain the quality score of the light field image. The application effectively quantifies the spatial quality and angle consistency of the light field image by performing feature extraction on the sub-aperture image and the macro-pixel respectively, thereby avoiding the problem that the quality of the light field image is restricted by the spatial quality and the angle consistency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of light field image quality evaluation, and particularly relates to a light field image quality evaluation method based on spatial and angular measurement. BACKGROUND

[0002] As a new imaging technology, light field can record the spatial and angular information of a scene simultaneously, so it provides monocular, binocular and depth information of the scene, and thus has more extensive applications. However, the high-dimensional characteristics of light field images bring new challenges to the compression, reconstruction, transmission and visualization of the light field images. In the process of processing the light field images, these algorithms inevitably introduce distortion into the light field images. After the clear light field image is introduced with distortion, the image quality of the light field image is degraded. The quality-degraded light field image will affect its subsequent application and user experience. In order to guide and optimize the light field image processing algorithm and provide better services for users, it is of great research significance to objectively and quantitatively evaluate the light field image quality. It is crucial to design an accurate light field image quality evaluation method to monitor the visual quality of the light field image system.

[0003] In the past few decades, researchers have developed relatively mature methods for accurately evaluating planar images, stereoscopic images and high dynamic range images. However, these methods are designed only according to the characteristics of specific images (planar images, stereoscopic images and high dynamic images). The quality of light field images is mainly affected by spatial quality and angular consistency, and these methods do not take into account the characteristics of light field images, so they cannot accurately estimate the quality of light field images.

[0004] In recent years, in order to accurately evaluate the quality of light field images, Tian et al. designed a multi-derivative feature model (MDFM) to measure the structural similarity between the original light field image and the distorted light field image. Huang et al. proposed a full-reference light field image quality evaluation method based on angular domain curve analysis and scene information statistics. Shi et al. proposed a no-reference light field image quality evaluation (NR-LFQA) method, which combines the natural distribution characteristics of the light field monocular image array and the global and local features of the epipolar plane image to evaluate the quality of the distorted light field image. Zhou et al. developed a tensor-oriented no-reference light field image quality evaluator (Tensor-NLFQ), which generates principal components from four directional sub-aperture view stacks using Tucker decomposition, quantifies the spatial quality of the distorted light field image using the global naturalness and local frequency features of the principal components, and proposes the structural similarity distribution between the principal components and each sub-aperture image in the view stack to evaluate the angular consistency of the distorted light field image.

[0005] Without considering the color dimension, the light field image can be regarded as a 4-dimensional function, which has multiple representations, such as sub-aperture image array, micro-lens image array, and epi-plane image, etc. The above light field image quality evaluation method mainly considers the representation forms of sub-aperture image array and epi-plane image array. Among them, the sub-aperture image array is difficult to effectively represent the angular consistency of the light field image.

[0006] In summary, it is an urgent problem for those skilled in the art to provide a light field image quality evaluation method based on spatial and angular measurement, which can effectively quantify the spatial quality and angular consistency of the light field image, and avoid the quality of the light field image being restricted by the spatial quality and angular consistency. SUMMARY

[0007] The present scheme aims at the above-mentioned problems and needs, and proposes a light field image quality evaluation method based on spatial and angular measurement. The method is a no-reference light field image quality evaluation method, which does not need to use the information of the reference light field image. It considers the multi-channel characteristics of visual perception in the sub-aperture image array, and uses the multi-band local binary pattern to extract the feature information of each sub-aperture image from the sub-aperture image array to quantify the spatial quality of the light field image. In order to capture the angular consistency of the light field image, the weighted local phase quantization algorithm is used to extract the feature on each macro-pixel in the micro-lens image array to represent the angular consistency of the distorted light field image. Finally, the two parts of features are combined to represent the overall quality of the light field image. The above technical problems can be solved by adopting the following technical solutions.

[0008] To achieve the above object, the present application provides the following technical scheme: a light field image quality evaluation method based on spatial and angular measurement, comprising: obtaining a 4-dimensional light field image to be evaluated, and using a multi-band local binary pattern algorithm to extract feature information of each sub-aperture image from a sub-aperture image array of the 4-dimensional light field image, and averaging the features of each sub-aperture image as the spatial features of the light field image;

[0009] Step Step2: using a feature extraction algorithm based on entropy weighted local phase quantization to extract the features of the micro-lens image array of the 4-dimensional light field image as the angular features of the light field image;

[0010] Step Step3: fusing the spatial features of the light field image and the angular features of the light field image to obtain a one-dimensional feature vector;

[0011] Step Step4: performing support vector regression pooling operation on the one-dimensional feature vector to obtain the quality score of the light field image.

[0012] Further, the specific steps of extracting the feature information of each sub-aperture image include:

[0013] First, let L be the 4D light field image to be evaluated, and let {I} be the sub-aperture image array. u,v}, Microlens image array {M s,t}, where u and v represent the angular planes of the light field image, and s and t represent the spatial planes of the light field image;

[0014] By convolving each sub-aperture image in the sub-aperture image array using a Gaussian difference filter, the Gaussian difference results of the sub-aperture images are obtained:

[0015]

[0016] in, The standard deviation is represented by σ. i The Gaussian function, with the standard deviation set to σ i =1.6 i-1 Using an approximate Gaussian Laplacian operator, the Lth subaperture image... I The Gaussian difference decomposition is as follows:

[0017]

[0018] L I Setting it to 4 yields four Gaussian difference plots for different frequency bands, respectively. And I is calculated using the Prewitt operator. u,v gradient plot Will Replace with The final feature maps of four different frequency bands were obtained.

[0019] Then, rotation-invariant equivalent local binary pattern encoding is performed on each type of feature map to obtain the corresponding 10 encoding patterns of the encoded map;

[0020] Finally, for each sub-aperture image I u,v Feature maps of 4 different frequency bands Each feature map is encoded using rotation-invariant equivalent local binary patterns. The frequency of each pattern in each feature map is counted to obtain the feature vectors. The vector is obtained by averaging the feature vectors of all sub-aperture images. F Gradient and vector F Gradient The concatenation results in a 40-dimensional feature vector. Feature vector Used to describe the spatial quality of a light field image.

[0021] Furthermore, the specific steps for performing rotation-invariant equivalent local binary pattern encoding on each type of feature map include:

[0022] Define an image I, calculate the center field of each pixel point in the image I:

[0023]

[0024] Where Ic and Ip represent the gray value of the pixel Ic and its surrounding field pixels on I, P and R represent the number of field pixels and the radius of the field, set P = 8 and R = 1, and the threshold function z() is defined as:

[0025]

[0026] According to the formula Calculate the local binary pattern initialization code of the image I ρ() is the number of bit conversion, which is expressed as:

[0027]

[0028] I P,R is the rotation invariant equivalent local binary pattern coding of I, a total of P+2 different types of patterns are obtained, that is, the coding image containing 10 types of coding patterns;

[0029] The coding image I P,R is the same as the amplitude of the corresponding image I lbp pixel value of the same mode is accumulated to obtain the frequency h of each mode Where, represents the coding mode of the coding image I P,R on the pixel Ic, C represents the total number of pixels of the image I, j represents the jth coding mode, j = 0, 1, …, P+1, and the function g() is represented as follows:

[0030]

[0031] Further, the specific steps of extracting the feature information of the microlens image array of the 4D light field image include:

[0032] First, the local phase quantization coding is performed on each macro-pixel M s,t in the light field image L to obtain the coding result;

[0033] Then, the LPQ coding result of the macro-pixel M s,t at each spatial position is defined as is a 256-dimensional feature vector, and the macro-pixel M s,tThe entropy value of the LPQ encoding result is used as a weight and multiplied by the LPQ encoding result of the macro-pixel. Then, the LPQ encoding results of the macro-pixel at each spatial location are accumulated to obtain the LPQ encoding result h of the light field image. lpq , in, h represents the entropy value of the LPQ encoding result of the macro-pixel at spatial location (s,t). lpq It is a 256-dimensional feature vector;

[0034] Then, the frequency results of four adjacent patterns from the 256 coding patterns are summed to obtain a 64-dimensional feature vector F. angular F angular Used to describe the angular consistency of a light field image.

[0035] Furthermore, for each macropixel M in the light field image L s,t The specific steps for performing local phase quantization encoding include:

[0036] Define a macro pixel M, according to the formula Perform a short-time Fourier transform within a 3×3 neighborhood around pixel M(x) of M, where N y For a 3×3 neighborhood, calculate H(u,x) at each of the four frequency points u1=[a,0]. T u2 = [0, a] T u3 = [a, a] T and u4=[a,-a] T The value of , where a is With M set to 3, for each pixel position in the macro pixel, we can obtain: H(x) = [H(u1,x),H(u2,x),H(u3,x),H(u4,x)];

[0037] Separating the real and imaginary parts of H(x) yields a vector W of length 8, where W = [Re(H(u)]. i ,x)),Im(H(u i [,x))], and perform binary encoding on W to obtain the encoded value B(x) of pixel point M(x) at macro pixel M, in, q i This represents the quantization at the i-th position in W. Therefore, there are 256 encoding modes for the pixel M(x) of a macro pixel M, and the value of B(x) is an integer between 0 and 255.

[0038] Furthermore, the specific process of fusing the feature information of each sub-aperture image and the feature information of each macropixel to obtain a one-dimensional feature vector, and then obtaining the quality score of the light field image, is as follows: F...spatial and F angular connect into 104-dimensional feature vector F LFIQA =[F spatial ,F angular ]; and the feature vector F LFIQA input into the support vector regression with radial basis function as the kernel which has been trained, and the regression obtains the quality score Q L .

[0039] From the above technical solutions, the beneficial effects of the present application are:

[0040] 1. The present application is a no-reference light field image quality evaluation method, which does not require any information of the original light field image, and can directly input the light field image waiting for quality evaluation as input, so that the method can be used under conditions where the original light field image cannot be obtained.

[0041] 2. The present application extracts features on the sub-aperture image array and the microlens image array respectively, and quantifies the quality of the light field image more effectively from different representations of the spatial quality and the angle consistency of the light field image.

[0042] 3. The present application provides two algorithms, i.e. the multi-band local binary pattern algorithm and the entropy-weighted local phase quantization algorithm, which provide new ideas for other feature engineering fields.

[0043] In addition to the above-described purposes, features and advantages, the optimal embodiments of the present application will be described in more detail below with reference to the accompanying drawings, so that the features and advantages of the present application can be easily understood. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the description of the embodiments of the present application or the prior art will be briefly introduced below, wherein the drawings are only used to show some embodiments of the present application, and the present application is not limited to the drawings.

[0045] Fig. 1 The specific steps of the present application, a light field image quality evaluation method based on spatial and angular measurement, are shown in the schematic diagram.

[0046] Fig. 2 The structure of the different representations of the 4-dimensional light field image in the present embodiment is shown in the specific steps schematic diagram.

[0047] Fig. 3 The flowchart of the present application, a light field image quality evaluation method based on spatial and angular measurement, is shown in the schematic diagram. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the technical solutions of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below in combination with the drawings of specific embodiments of the present application. The same reference signs in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0049] The present application firstly extracts features of each sub-aperture image in the sub-aperture image array of the light field image based on the characteristics of each sub-aperture image in the sub-aperture image array of the light field image, uses a multi-band local binary pattern algorithm to extract features of each sub-aperture image, and averages the extracted features of each sub-aperture image to obtain features for describing the spatial quality of the light field image. Then, the present application extracts features of each macro-pixel in the microlens image array of the light field image based on the characteristics of each macro-pixel in the microlens image array of the light field image, uses an entropy-weighted local phase quantization algorithm to extract features of each macro-pixel, and uses the obtained features to quantify the angular consistency of the light field image. Finally, the present application combines the two parts of features into a feature vector, and obtains the quality score of the light field image through support vector regression technology. In view of the fact that the quality of the light field image is restricted by the spatial quality and the angular consistency, the present application extracts features from the sub-aperture image array and the microlens image array of the light field image respectively. Each sub-aperture image in the sub-aperture image array of the light field image reflects the spatial information of a certain angle of the light field image, and each macro-pixel in the microlens image array of the light field image reflects the angular information of a certain space of the light field image, so they can fully reflect the spatial quality and the angular consistency of the light field image.

[0050] As shown in Figs. 1-3 , a light field image quality evaluation method based on spatial and angular measurement specifically includes the following steps:

[0051] Step Step1: obtaining a 4D light field image to be evaluated, and using a multi-band local binary pattern algorithm to extract feature information of each sub-aperture image from a sub-aperture image array of the 4D light field image, and averaging the features of each sub-aperture image as spatial features of the light field image.

[0052] The specific steps of extracting the feature information of each sub-aperture image include:

[0053] Step Step1.1: first, let the 4D light field image to be evaluated be L, and the light field image has a sub-aperture image array and a microlens image array representation form. As shown in Fig. 2 (a), the sub-aperture image array is composed of a plurality of sub-aperture images, and can be represented as {I u,v}, where u and v represent the angular plane of the light field image, meaning that the angular coordinates of the sub-aperture image are (u, v), as shown inFig. 2 As shown in (b), the microlens image array is composed of a plurality of macro-pixels, which can be represented as {M s,t}, wherein s and t represent the spatial plane of the light field image, meaning that the spatial coordinates of the macro-pixel are (s, t);

[0054] Step Step1.2: Convolve each sub-aperture image in the sub-aperture image array with a difference of Gauss filter to obtain the difference of Gauss result of the sub-aperture image:

[0055]

[0056] wherein "*" represents a convolution operator, represents a Gaussian function with a standard deviation of σ i , and the standard deviation is set to σ i = 1.6 i-1 to approximate the Gaussian Laplace operator, and the L I th-order difference of Gauss decomposition of the sub-aperture image is:

[0057]

[0058] Step Step1.3: Set L I to 4 to obtain four difference of Gauss images of different frequency bands, respectively and calculate the gradient image I u,v using the Prewitt operator Replace with to obtain the final four feature maps of different frequency bands

[0059] Step Step1.4: Then, each type of feature map is rotationally invariant local binary pattern (LBP) coded to obtain the corresponding coding image of 10 encoding modes.

[0060] In this embodiment, the specific steps of rotationally invariant local binary pattern coding for each type of feature map include:

[0061] Step Step1.41: Define an image I, and calculate the center field of each pixel point in the image I:

[0062]

[0063] wherein Ic and Ip represent the gray value of the pixel Ic on I and the gray value of the surrounding field pixels, P and R represent the number of field pixels and the radius of the field, P = 8 and R = 1 are set, and the threshold function z() is defined as:

[0064]

[0065] Step Step1.42: Calculate the local binary pattern initialization code of image I according to the formula ρ() is the number of bit conversion, which is expressed as:

[0066]

[0067] I P,R is the rotation invariant equivalent local binary pattern code of I, and a total of P+2 different types of patterns are obtained, that is, the code map containing 10 types of encoding patterns;

[0068] Step Step1.43: The encoding map I P,R is obtained by accumulating the amplitude of the pixel value of the corresponding image I in the same mode. lbp It is a 10-dimensional feature vector, wherein, represents the encoding mode of the encoding map I P,R on the pixel Ic, C represents the total number of pixels of the image I, j represents the jth encoding mode, j=0,1,…,P+1, and the function g() is expressed as follows:

[0069]

[0070] Step Step1.5: Finally, the 4 different frequency band feature maps of each sub-aperture image I u,v are encoded by using the rotation invariant equivalent local binary pattern (LBP), the frequency of each mode appearing in each feature map is counted, and the feature vectors are obtained respectively. The feature vectors of all sub-aperture images are averaged to obtain the vector F Gradient , and the vector F Gradient is concatenated to obtain a 40-dimensional feature vector The feature vector is used to represent the spatial quality of the light field image.

[0071] Step Step2: The feature extraction algorithm based on entropy weighted local phase quantization is used to extract the features of the microlens image array of the 4-dimensional light field image as the angular feature of the light field image, wherein the specific steps of extracting the feature information of the microlens image array of the 4-dimensional light field image include:

[0072] Step Step2.1: First, the local phase quantization encoding is performed on each macro-pixel M s,t in the light field image L to obtain the encoding result.

[0073] ​​In the present embodiment, the LPQ encoding of each macro-pixel M in the light field image L is defined as s,t The specific steps of local phase quantization encoding include:

[0074] Step Step2.11: define a macro-pixel M, and according to the formula Perform short-time Fourier transform in the 3x3 field around the pixel point M(x) of M, where N y represents the 3x3 field, and the values of H(u, x) at four frequency points u1=[a, 0] T , u2=[0, a] T , u3=[a, a] T and u4=[a, -a] T are calculated respectively, where a is M is empirically set to 3, and for each pixel position in the macro-pixel, H(x)=[H(u1, x), H(u2, x), H(u3, x), H(u4, x)] can be obtained;

[0075] Step Step2.12: separate the real and imaginary parts of H(x) to obtain a vector W with a length of 8, W=[Re(H(u i , x)), Im(H(u i , x))], and binary encode W to obtain the encoding value B(x) of the pixel point M(x) of the macro-pixel M, wherein, q i represents the quantization of the i-th position in W, therefore, there are 256 encoding modes for the pixel point M(x) of a macro-pixel M, i.e. B(x) takes an integer value between 0 and 255.

[0076] Step Step2.2: according to the above process, define the LPQ encoding result of each spatial position macro-pixel M s,t as is a 256-dimensional feature vector, and the entropy value of the LPQ encoding result of each spatial position macro-pixel M s,t is taken as the weight, multiplied by the LPQ encoding result of the macro-pixel, and then the LPQ encoding result of each spatial position macro-pixel is accumulated as the LPQ encoding result h lpq of the light field image, wherein, represents the entropy value of the LPQ encoding result of the macro-pixel at the spatial position (s, t), h lpq is a 256-dimensional feature vector;

[0077] Step Step2.3: Then, the frequency results of 4 adjacent modes in 256 encoding modes are accumulated, for example, the frequency results of modes "0", "1", "2" and "3" are added, the frequency results of modes "4", "5", "6" and "7" are added, and so on, so as to finally obtain a 64-dimensional feature vector F angular angular for describing the angular consistency of the light field image.

[0078] Step Step3: The spatial feature of the light field image and the angular feature of the light field image are fused to obtain a one-dimensional feature vector;

[0079] Step Step4: After support vector regression pooling operation is performed on the one-dimensional feature vector, a quality score of the light field image is obtained.

[0080] Specifically, the feature information of each sub-aperture image and the feature information of each macro-pixel are fused to obtain a one-dimensional feature vector, and then the specific process of obtaining the quality score of the light field image is as follows: F spatial and F angular are connected to form a 104-dimensional feature vector F LFIQA = [F spatial , F angular ]; and the feature vector F LFIQA is input into the support vector regression with a radial basis function as a kernel which has been trained, and the quality score Q L of the light field image L to be evaluated is obtained by regression.

[0081] It should be noted that the embodiments of the present application are only preferred modes for realizing the present application, and any modifications which belong to the overall concept of the present application and are only obvious should be within the protection scope of the present application.​

Claims

1. A method for light field image quality assessment based on spatial and angular measurements, characterized in that, The method comprises the following steps: Step Step1: obtaining a 4D light field image to be evaluated, and extracting feature information of each sub-aperture image from a sub-aperture image array of the 4D light field image by using a multi-band local binary pattern algorithm, and averaging the feature of each sub-aperture image as a spatial feature of the light field image; Step Step2: extracting features of a microlens image array of the 4D light field image as an angle feature of the light field image by using a feature extraction algorithm based on entropy-weighted local phase quantization; Step Step3: fusing the spatial feature of the light field image and the angle feature of the light field image to obtain a one-dimensional feature vector; Step Step4: obtaining a quality score of the light field image after performing a support vector regression pooling operation on the one-dimensional feature vector; The specific steps of extracting the feature information of each sub-aperture image comprise: First, let's denote the 4D light field image to be evaluated as , the sub-aperture image array as I u,v , the microlens image array as M s,t , where u and v represent the angular planes of the light field image, s and t represent the spatial planes of the light field image; convolving each sub-aperture image in the sub-aperture image array by using a Gaussian difference filter to obtain a Gaussian difference result of the sub-aperture image: wherein, represents a Gaussian function with a standard deviation of is set to to approximate a Gaussian Laplacian, then the sub-aperture image is the first L I level Gaussian difference decomposition is: ； will be set to 4, and four different frequency band Gaussian difference maps will be obtained, respectively L I Set to 4, obtain 4 different frequency band Gaussian difference map, respectively , and use Prewitt operator to calculate I u,v Gradient map , replace With , get the final four different frequency band feature map ; Then, each type of feature map is subjected to rotationally invariant equivalent local binary pattern coding to obtain an encoding map corresponding to 10 types of encoding modes; Finally, the feature vectors of each sub-aperture image I u,v are obtained by encoding the feature maps of the 4 different frequency bands respectively using the rotation-invariant equivalent local binary pattern, and counting the frequency of each pattern appearing in each feature map The feature vectors of all sub-aperture images are averaged to obtain the vector and the vector is concatenated to obtain a 40-dimensional feature vector The feature vector is used to represent the spatial quality of the light field image; The specific steps of extracting the feature information of the microlens image array of the 4D light field image comprise: First, each macro-pixel in the light field image L is encoded by local phase quantization to obtain an encoding result. M s,t First, each macro-pixel in the light field image L is encoded by local phase quantization to obtain an encoding result. LPQ encoding result of each macro-pixel at each spatial position M s,t is defined as , is a 256-dimensional feature vector, and the entropy value of the LPQ encoding result of each macro-pixel at each spatial position M s,t is multiplied by the LPQ encoding result of the macro-pixel, and then the LPQ encoding result of each macro-pixel at each spatial position is accumulated as the LPQ encoding result of the light field image h lpq , wherein represents the entropy value of the LPQ encoding result of the macro-pixel at the spatial position (s, t), h lpq is a 256-dimensional feature vector; Then, the frequency results of the 4 adjacent modes among the 256 encoding modes are accumulated, and finally a 64-dimensional feature vector is obtained F angular , F angular Angular consistency for describing light field images.

2. The spatial and angular measurement based light field image quality evaluation method of claim 1, wherein, The specific steps of performing rotationally invariant equivalent local binary pattern coding on each type of feature map comprise: defining an image I, and calculating a center field of each pixel point in the image I: ， where Icand Ipdenote the gray value of the pixel Icand its surrounding neighborhood pixels on I, P and R denote the number of neighborhood pixels and the radius of the neighborhood, P = 8 and R = 1 are set, and the threshold function is defined as: ； According to the formula The local binary pattern initialization encoding of the image I is calculated , The number of bit transitions, denoted as: ； is a rotationally invariant equivalent local binary pattern coding of I, resulting in a total of P+2 different types of patterns, i.e. a coding map comprising 10 coding patterns; The encoding map is denoted by The amplitude of the pixel values in the corresponding image I of the same pattern is accumulated to obtain the frequency of each pattern , wherein, The encoding map is denoted by The encoding pattern on the pixel Ic, C represents the total number of pixels of the image I, j represents the jth encoding pattern, j = 0, 1, …, P + 1, and the function is denoted as follows: 。 3.The spatial and angular measurement based light field image quality evaluation method of claim 1, wherein, said each macro-pixel in the light field image L M s,t The specific steps of local phase quantization encoding include: Defining a macro-pixel M , according to the formula In M the 3 M 3 domain around the pixel point x ( ) Short-time Fourier transform is performed, wherein, N y represents the 3 3 domain, and the values of H ( u , x ) at four frequency points u 1=[ a ,0] T , u 2=[0, a ] T , u 3=[ a , a ] T and u 4=[ a ,- a ] T are calculated, wherein, a is , M is set to 3, and for each pixel position in the macro-pixel, the following can be obtained: ; Separating the real and imaginary parts of H(x) to obtain a vector W of length 8, and binary encoding W to obtain the encoding value B(x) of the pixel point M(x) of the macro-pixel M, wherein, , q i denotes quantization of the quantity in the i-th position of W, thus, there are 256 encoding modes for the pixel point M(x) of one macro-pixel M, and B(x) takes integer values ranging from 0 to 255.

4. The spatial and angular measurement based light field image quality evaluation method of claim 3, wherein, The feature information of each sub-aperture image and the feature information of each macro-pixel are fused to obtain a one-dimensional feature vector, and then the specific process of obtaining the quality score of the light field image is as follows: the feature vector is input into the support vector regression with the radial basis function as the kernel which has been trained, and the regression obtains the quality score L of the light field image to be evaluated. F spatial and F angular connected into a 104-dimensional feature vector F LFIQA [ F spatial , F angular ]; and the feature vector F LFIQA is input into the support vector regression with the radial basis function as the kernel which has been trained, and the regression obtains the quality score L of the light field image to be evaluated. Q L .

Images