Saliency map image quality assessment method based on image-based objective parameters

Abstract

The application relates to a SAL image quality evaluation method based on image-based objective parameters, which comprises the following steps: generating an initial gray image of an input SAL echo signal; performing third-order wavelet decomposition on the initial gray image to generate a plurality of image sub-channels; solving corresponding contrast sensitivity function values by using the spatial frequency of the image sub-channels to obtain human visual system response weights; solving a first basic objective parameter of each image sub-channel and a second basic objective parameter of the initial gray image; calculating a human visual system evaluation value of a corresponding imaging result graph of the SAL echo signal according to the human visual system response weights and the first basic objective parameter of the image sub-channels, and solving the quality of the imaging result graph by using the human visual system evaluation value and the second basic objective parameter. The method does not rely on an image database of artificial scoring, and does not need to obtain prior information of a distorted image, the solved basic objective parameter represents the essential and natural characteristics of an image, and has certain universality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a SAL image quality evaluation method based on image basic objective parameters. BACKGROUND

[0002] With the vigorous development of modern technology, images as the carrier of information transmission have a huge and profound impact on our life. Due to the difference in camera quality and the unevenness of light-sensitive elements, distortion will inevitably occur in the process of image transmission, storage and display. In many digital image processing, the quality of the image is an important indicator to measure the performance of the system, and the system adjusts the output of the system parameters through image quality evaluation to improve the performance of the system.

[0003] Synthetic aperture ladar (SAL) is a high-resolution imaging ladar using optical synthetic aperture technology and coherent heterodyne detection technology, which has wide application in target detection and identification. In the process of SAL signal acquisition and transmission, different forms and degrees of degradation will occur, resulting in loss of part of the information and decline of image quality, which will affect the information acquisition ability and subsequent image interpretation. The quality of the image is the main basis for measuring the performance of the imaging system and the advantages and disadvantages of the imaging algorithm, so it is necessary to qualitatively analyze and quantitatively evaluate the SAL image.

[0004] From the aspect of feature extraction, research shows that natural scene images have certain statistical characteristics. The mean subtracted contrast normalized (MSCN) coefficients are extracted from natural scene images, and it is found that the coefficients are Gaussian distributed. The definition of natural scene images in the literature is that the natural images captured by optical cameras are not subject to human post-processing, but are not limited to natural scenes such as trees and sky, so the feature extraction method does not depend on the content of the image itself. When the image is distorted, these statistical characteristics will be destroyed, so these statistical characteristics can be used to measure the distortion degree of the image. According to this theory, Mittal et al. proposed a no-reference image quality assessment algorithm based on spatial statistical characteristics (BRISQUE); the method first extracts quality-sensitive features in the spatial domain: the normalized MSCN coefficients are fitted with a generalized Gaussian distribution model, and the fitted parameters are used as quality-sensitive features; then support vector regression (SVR) is used to learn the mapping relationship between the features and the subjective evaluation score, so as to judge the perceived quality of the image to be tested according to the learned mapping relationship. In 2011, Moorthy proposed a no-reference image quality assessment algorithm based on wavelet domain natural scene statistical characteristics; the method uses pyramid decomposition to decompose the image into different frequency image sub-channels, and solves the natural scene statistical characteristics (NSS) on this basis, and then uses support vector regression (SVR) to map the quality-sensitive features to the perceived quality of the human eye.

[0005] At the level of whether to have reference images, image quality evaluation methods are mainly divided into three categories: full reference type, partial reference type, and no reference type. The full reference type and partial reference type methods in the current mainstream objective image quality evaluation all need reference images as a reference, while in most image quality evaluation tasks, there are no reference images for them to refer to. The main problem faced in practical applications is to evaluate images without reference. Unlike reference image quality evaluation, the challenge of no-reference image quality evaluation lies in the need to estimate the distortion type and degradation level of the image without reference. Most no-reference image quality evaluation algorithms are restricted by two aspects: one type needs to pre-judge the distortion type of the image; the other type needs to be pre-trained based on a specific image database.

[0006] For no-reference type, many scholars propose to extract natural scene statistical characteristics from the spatial domain, wavelet transform domain, etc. Because the image de-mean contrast normalization coefficient tends to be a Gaussian normal distribution, when the image is distorted, this statistical characteristic will be destroyed. In general, the extracted image features do not conform well to the human visual system, and as the terminal for image quality evaluation, the human eye requires that the image quality evaluation conform to human visual perception.

[0007] In practical application scenarios, because the factors that cause image quality degradation are often not just one, it is difficult to obtain prior information about the distortion type of the image. However, algorithms trained based on a specific image database cannot achieve ideal evaluation results when cross-tested on other databases. Since no distortion reference image can be obtained in advance, the focus is on no-reference type, as mentioned above, the natural scene statistical characteristic (NSS) method belongs to the typical no-reference image quality evaluation method. Full reference type methods also have certain reference or contrast reference significance, and scholars often compare the new algorithms proposed with full reference type methods. The most common full reference type image quality evaluation methods are two: (1) Peak Signal-to-Noise Ratio (PSNR); (2) Mean Squared Error (MSE), the expressions of the two methods are:

[0008]

[0009]

[0010] where R represents the reference image, D represents the distorted image, and L represents the maximum pixel value of the image. The unit of PSNR is dB, and the larger the value, the smaller the distortion.

[0011] The two methods directly solve the difference of pixel values at the same point, and belong to error-sensitive-based image quality evaluation. Since the characteristics of human visual system (HVS) are not considered, such as the high sensitivity of human eyes to low frequency components in spatial frequency, the visual perception of a region is affected by its surrounding adjacent regions, etc., the evaluation results are often inconsistent with the visual effects.

[0012] Yu-jing LIU proposed a general no-reference synthetic aperture radar image (SAR) quality evaluation algorithm based on human visual system. In the paper, discrete wavelet decomposition is used to obtain the weight of human visual system in different image sub-channels, and finally the subjective quality score of synthetic aperture image is obtained.

[0013] No-reference image quality evaluation methods are divided into two categories according to the prior knowledge of distortion type: distortion specified type and non-distortion specified type. In recent years, researchers have invested more research on JPEG distorted images and JP2K distorted images, mainly focusing on the influence of compression and block blur on image quality. However, in practical applications, because there is no prior information about the distortion type, these image quality evaluation algorithms for specific distortion types are limited, so it is necessary to carry out research on non-distortion specified type image quality.

[0014] In terms of feature extraction and feature learning, the existing methods mainly extract natural scene statistical characteristics, then use mathematical models to fit them into feature parameters, and then map the feature parameters to subjective quality scores through support vector regression (SVR). Or use convolutional neural network to integrate feature extraction and regression fitting into one step, so that the feature has self-learning ability and constructs an end-to-end (End-to-End) model. The two methods belong to the field of machine learning, which needs a large number of training samples as training data, and involves image data preprocessing, image segmentation and other operations. The training samples are subjectively evaluated to generate corresponding labels, and the human eye itself has individual differences in generating labels. In the process of training the network, the loss function, network depth, number of convolution kernels and parameter tuning need to be considered, and the actual operation process is relatively complex, and the requirement for computer computing power is also relatively high.

[0015] Image-based objective parameters are important basis for studying and understanding images, and extraction of image objective parameters often focuses on texture, edge, shape and other features. The target features are not extracted completely in accordance with human visual perception, and the final understanding and interpretation of images depend on human visual characteristics, so image quality evaluation based on human visual system will be more conducive to image evaluation. In the field of no-reference image quality evaluation, it is necessary to objectively evaluate the image quality without relying on artificial scoring and prior information of image distortion type. SUMMARY

[0016] In order to solve the above problems existing in the prior art, the present application provides a SAL image quality evaluation method based on image-based objective parameters. The technical problem to be solved by the present application is solved by the following technical scheme:

[0017] The embodiment of the present application provides a SAL image quality evaluation method based on image-based objective parameters, characterized in that it comprises the following steps:

[0018] Step 1, generating an initial gray image of an input SAL echo signal;

[0019] Step 2, performing third-order wavelet decomposition on the initial gray image to generate a plurality of image sub-channels;

[0020] Step 3, solving the contrast sensitivity function value of each image sub-channel by using the spatial frequency of each image sub-channel to obtain the human visual system response weight;

[0021] Step 4, solving the first basis objective parameter of each image sub-channel and the second basis objective parameter of the initial gray image;

[0022] Step 5, calculating the human visual system evaluation value of the corresponding imaging result image of the SAL echo signal according to the human visual system response weight and the first basis objective parameter of each image sub-channel, and solving the quality of the imaging result image by using the human visual system evaluation value and the second basis objective parameter.

[0023] In one embodiment of the present application, the step 1 comprises:

[0024] The SAL echo signal is sequentially subjected to format conversion and numerical normalization to generate the initial gray image.

[0025] In one embodiment of the present application, the step 2 comprises:

[0026] The initial gray image is subjected to three-order wavelet decomposition, and a MATLAB built-in function is called in each decomposition to decompose the image to be decomposed into a horizontal detail coefficient matrix, a vertical detail coefficient matrix, a diagonal detail coefficient matrix and an approximate coefficient matrix, and a decomposed image is obtained by 2 times down-sampling of the image to be decomposed in each decomposition;

[0027] The horizontal detail coefficient matrix, the vertical detail coefficient matrix, the diagonal detail coefficient matrix and the approximate coefficient matrix after each decomposition and the approximate coefficient matrix after the last decomposition are selected as 10 sub-channels after the three-order wavelet decomposition:

[0028] Channels=[cA3,cH3,cV3,cD3,cH2,cV2,cD2,cH1,cV1,cD1]

[0029] Wherein, cA3 is the approximate coefficient matrix after the three-order wavelet decomposition, cHj is the horizontal detail coefficient matrix, cVj is the vertical detail coefficient matrix, cDj is the diagonal detail coefficient matrix, and j is the decomposition layer number.

[0030] In an embodiment of the present application, the step 3 comprises:

[0031] The spatial frequency of each image sub-channel is calculated by using the image gray value:

[0032]

[0033]

[0034] Wherein, RF is the row frequency of the image sub-channel, CF is the column frequency of the image sub-channel, [m, n] is the row number and column number, I(i, j) is the image gray value of the element coordinate (i, j), and SF is the spatial frequency of the image sub-channel.

[0035] The spatial frequency value is normalized to 0-50 to obtain the normalized spatial frequency.

[0036] The contrast sensitivity function value of the image sub-channel is solved by using the normalized spatial frequency and is subjected to numerical normalization to obtain the human visual system response weight of each sub-channel, wherein the calculation formula of the contrast sensitivity function value is:

[0037] H(f)=2.6×(0.0192+0.114×f)×exp(-(0.114×f) 1.1 )

[0038] Wherein, is the normalized spatial frequency, f x and f y are the normalized spatial frequencies in the horizontal and vertical directions.

[0039] In one embodiment of the present application, the step 4 comprises:

[0040] numerically normalizing each of the image sub-channels to make the image sub-channels into gray-scale images;

[0041] solving a first basic objective parameter of each of the image sub-channels and a second basic objective parameter of the initial gray-scale image.

[0042] In one embodiment of the present application, the first basic objective parameter comprises: mean, variance, equivalent number of looks, radiometric resolution, image gradient amplitude spectrum and first image entropy, and the second basic objective parameter comprises: contrast based on gray-scale co-occurrence matrix, second image entropy and signal-to-noise ratio.

[0043] In one embodiment of the present application, the solving method of the mean comprises: summing up the gray-scale values of each pixel point of the image sub-channels and then averaging to obtain the mean of the image sub-channels.

[0044] The solving method of the variance comprises: subtracting the mean from the gray-scale value of each pixel point of the image sub-channels, squaring the result, then traversing the squared values to sum up and dividing by the total number of pixel points to obtain the variance of the image sub-channels.

[0045] The solving method of the equivalent number of looks comprises: dividing the mean of the image sub-channels by the variance and then squaring to obtain the equivalent number of looks of the image sub-channels:

[0046]

[0047] wherein ENL is the equivalent number of looks, I is the gray-scale image, μ is the mean of the image, σ 2 is the variance of the image.

[0048] The solving method of the radiometric resolution comprises: calculating the radiometric resolution using the equivalent number of looks:

[0049]

[0050] wherein ENL is the equivalent number of looks, Γ is the radiometric resolution.

[0051] The solving formula of the image gradient amplitude spectrum is:

[0052]

[0053]

[0054] wherein grad H (i,j) is the image horizontal gradient spectrum, grad V(i,j) is an image vertical gradient spectrum, grad(i,j) is an image gradient amplitude spectrum, I(i,j) is a gray image, and G(x,y; sigma) is a Gaussian filter with a standard deviation of sigma;

[0055] The solution method of the contrast based on the gray level co-occurrence matrix is that the initial gray image is compressed by gray level compression, 256 levels are compressed into 16 levels, and then the frequency of occurrence of two pixel points with gray values of i and j in a target direction is counted to obtain the contrast based on the gray level co-occurrence matrix in the corresponding direction:

[0056]

[0057] wherein L is a gray level (L=16), P d (i,j) is the frequency of occurrence of two pixel points with gray values of i and j at a spatial position distance of d;

[0058] The solution formulae of the first image entropy and the second image entropy are as follows:

[0059]

[0060] wherein i is a gray value of a pixel point, P(i) is the ratio of the number of pixel points with a gray value of i to the total number of pixel points, and n is the total number of gray levels;

[0061] The solution formula of the signal-to-noise ratio is as follows:

[0062] SNR new = SNR x 1.04-7

[0063]

[0064] wherein sv is a signal variance, nv is a noise variance, SNR new is the corrected signal-to-noise ratio, and SNR is the uncorrected signal-to-noise ratio.

[0065] In an embodiment of the present application, the step 5 comprises:

[0066] The human visual system evaluation value of the corresponding imaging result map of the SAL echo signal is calculated according to the human visual system response weight of each image sub-channel and the mean, variance, equivalent visual number, radiation resolution, image gradient amplitude spectrum and first image entropy of the plurality of image sub-channels;

[0067] The quality of the imaging result map is solved by using the human visual system evaluation value and the contrast based on the gray level co-occurrence matrix, the signal-to-noise ratio and the second image entropy.

[0068] In an embodiment of the present application, the human visual system evaluation value is:

[0069]

[0070] wherein n is the number of image bits (n = 8), i is the sub-channel, ω i is the HVS response weight on sub-channel i, μ i is the mean of sub-channel i, is the variance of sub-channel i, ENL i is the equivalent noise of sub-channel i, Γ i is the radiometric resolution of sub-channel i, grad i is the image gradient magnitude spectrum of sub-channel i, Entropy i is the first image entropy of sub-channel i.

[0071] In an embodiment of the present application, the quality of the imaging result map is:

[0072]

[0073] wherein Quality is the quality of the gray scale image to be tested, HVSNR is the human visual system evaluation value, Con is the contrast based on the gray scale co-occurrence matrix, SNR is the signal-to-noise ratio, and Entropy is the second image entropy.

[0074] Compared with the prior art, the present application has the following beneficial effects:

[0075] The present application aims at the evaluation of the quality of SAL images, is based on the initial gray scale image generated from SAL echo data, considers that the image quality evaluation result needs to be combined with human visual perception, performs third-order wavelet decomposition on the initial gray scale image to simulate the human visual system, gives each sub-channel a corresponding visual sensitivity weight by using a contrast sensitivity function, and solves the basic objective parameters of the image on the sub-channels of the decomposed image, so as to obtain the human visual system evaluation value. This method does not rely on an image database of artificial scoring, nor does it need to obtain prior information of distorted images. The solved basic objective parameters represent the essential and natural characteristics of the image, have a good distinction and interpretation of the image quality of different degrees of blurring and defocusing under the same scene, have no requirements and restrictions on the type of image distortion, have a certain universality, have a relatively simple operation algorithm, and have a relatively small amount of calculation. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 Fig. 1 is a flowchart of a SAL image quality evaluation method based on image basic objective parameters according to an embodiment of the present application;

[0077] Figure 2 Fig. 2 is a flowchart of another SAL image quality evaluation method based on image basic objective parameters according to an embodiment of the present application;

[0078] Figures 3a-3d An imaging result graph after processing of measured data provided by the embodiment of the present application;

[0079] Figures 4a-4b A relationship curve graph of the human visual system evaluation value HVSNR and the quality with the defocus degree of the distorted image provided by the embodiment of the present application. DETAILED DESCRIPTION

[0080] The present application will be further described in detail below in combination with specific embodiments, but the embodiments of the present application are not limited thereto.

[0081] Embodiment one

[0082] The present embodiment aims at the problem that the synthetic aperture laser radar (SAL) image in the prior art lacks an image database of artificial scoring and prior information of image distortion types, and proposes a SAL image quality evaluation method based on image basic objective parameters.

[0083] See Figure 1 and Figure 2 , Figure 1 A flowchart of a SAL image quality evaluation method based on image basic objective parameters provided by the embodiment of the present application, Figure 2 A flowchart of another SAL image quality evaluation method based on image basic objective parameters provided by the embodiment of the present application.

[0084] The SAL image quality evaluation method based on image basic objective parameters of the present embodiment comprises the following steps:

[0085] S1, generating an initial gray scale image of an input SAL echo signal.

[0086] Specifically, the complex matrix of the input measured SAL echo signal is sequentially subjected to format conversion and numerical normalization to generate the initial gray scale image.

[0087] Firstly, the SAL echo signal is subjected to format conversion. Specifically, the input SAL echo signal is a complex signal, and when the SAL echo signal is loaded into MATLAB, a complex matrix is obtained through format conversion, and then the absolute value of each element of the complex matrix is taken to transform the complex matrix into a real matrix.

[0088] Then, when the numerical span range of the real matrix exceeds 255 and a large number of double-precision decimals are distributed, the numerical normalization method is needed to map it to the interval of 0-255, and the gray scale value of each pixel point needs to be rounded to make it become an 8-bit gray scale image, so as to obtain the initial gray scale image.

[0089] S2, performing third-order wavelet decomposition on the initial gray-scale image to generate a plurality of image sub-channels. Specifically, the method comprises the following steps:

[0090] S21, performing third-order wavelet decomposition on the initial gray-scale image, and calling a MATLAB built-in function at each decomposition step to decompose the image to be decomposed into a horizontal detail coefficient matrix, a vertical detail coefficient matrix, a diagonal detail coefficient matrix, and an approximation coefficient matrix, and at each decomposition step, performing 2 times down-sampling on the image to be decomposed to obtain a decomposed image.

[0091] Specifically, the human visual system is good at capturing low-frequency and medium-frequency information in an image, and is not sensitive to high-frequency information of the image. This visual frequency perception characteristic can be simulated by wavelet transform.

[0092] The initial gray-scale image is divided into image sub-channels with different frequencies by performing third-order wavelet decomposition on the initial gray-scale image. At each decomposition step, low-frequency information of a previous order is decomposed into low-frequency information and high-frequency information of a next order by using a low-pass filter and a high-pass filter. In the third-order wavelet decomposition, a wavelet base of 'haar' is used, and at each decomposition step, a MATLAB built-in function 'detcoef2' is called to obtain three detail coefficient matrices of the order, i.e., a horizontal detail coefficient matrix, a vertical detail coefficient matrix, and a diagonal detail coefficient matrix; and a MATLAB built-in function 'appcoef2' is called to obtain an approximation coefficient matrix of the order. The approximation coefficient matrix represents low-frequency information of the image, and the detail coefficient matrix represents high-frequency information of the image. Meanwhile, each decomposed image is obtained by performing 2 times down-sampling on the image to be decomposed, i.e., the 2 times down-sampling operation is integrated into the filtering process, which can be expressed by a mathematical formula as follows:

[0093]

[0094] wherein c is a low-frequency sub-band, d is a high-frequency sub-band, j is a decomposition order, h(n) and g(n) are a low-frequency filter and a high-frequency filter respectively, k is a pixel point in a two-dimensional image plane during traversal, and Z is a two-dimensional plane image.

[0095] S22, selecting the horizontal detail coefficient matrix, the vertical detail coefficient matrix, the diagonal detail coefficient matrix, and the approximation coefficient matrix of each decomposition order, and the approximation coefficient matrix of the last decomposition order as 10 sub-channels after the third-order wavelet decomposition.

[0096] Specifically, because four coefficient matrices with the same size are obtained at each decomposition step, 12 coefficient matrices are obtained after the third-order decomposition. Three detail coefficient matrices of each decomposition order and the approximation coefficient matrix of the last decomposition order are selected from the 12 coefficient matrices as 10 coefficient matrices as 10 image sub-channels after the third-order wavelet transform of the initial gray-scale image. The 10 coefficient matrices can be expressed by an array as follows:

[0097] Channels = [cA3, cH3, cV3, cD3, cH2, cV2, cD2, cH1, cV1, cD1]

[0098] Wherein, cA3 is the approximate coefficient matrix of the third-order wavelet decomposition, cHj is the horizontal detail coefficient matrix, cVj is the vertical detail coefficient matrix, cDj is the diagonal detail coefficient matrix, and j is the decomposition layer.

[0099] In this embodiment, the input initial gray image is sequentially subjected to three single-scale wavelet decompositions and direct third-order wavelet decomposition using the same wavelet basis, and the same result is obtained.

[0100] In this embodiment, the third-order wavelet decomposition is performed on the image to simulate the sensitivity of the human visual system to different frequency components, so that the final evaluation result can be well combined with the human visual perception.

[0101] S3, the spatial frequency of each image sub-channel is used to solve the contrast sensitivity function value of the corresponding image sub-channel, and the human visual system response weight is obtained. Specifically, it includes the following steps:

[0102] S31, the spatial frequency of each image sub-channel is calculated using the image gray value.

[0103] Specifically, an input initial gray image is decomposed into 10 image sub-channels with different spatial frequencies through third-order wavelet transform. The spatial frequency of each image sub-channel is expressed by the following mathematical formula:

[0104]

[0105]

[0106] Wherein, RF is the row frequency of the image sub-channel, CF is the column frequency of the image sub-channel, [m, n] is the row and column number, I(i,j) is the image gray value of element coordinate (i,j), and SF is the spatial frequency of the image sub-channel.

[0107] S32, the spatial frequency value is normalized to 0-50 to obtain the normalized spatial frequency.

[0108] Specifically, when SF=50, the contrast sensitivity function (CSF) curve is very close to 0, so the spatial frequency SF in the range of (0, 50) needs to be normalized to 0-50, so as to obtain the normalized spatial frequency; then the average value of the CSF of each image sub-channel is solved as the frequency response of the human visual system model.

[0109] S33, solving the contrast sensitivity function value of the image sub-channel by using the normalized spatial frequency and performing numerical normalization, to obtain the human visual system response weight of each sub-channel.

[0110] Specifically, the contrast sensitivity function value CSF corresponding to each image sub-channel is solved and numerically normalized, and is taken as the HVS response weight of each sub-channel, that is, the response of the human visual system to different frequency components is obtained. The contrast sensitivity function is mathematically expressed as follows:

[0111] H(f) = 2.6 x (0.0192 + 0.114 x f) x exp(-(0.114 x f) 1.1 )

[0112] wherein, is the normalized spatial frequency, f x and f y are the normalized spatial frequencies in the horizontal and vertical directions.

[0113] S4, solving the first basic objective parameter of each image sub-channel and the second basic objective parameter of the initial gray image. Specifically, it comprises the following steps:

[0114] S41, performing numerical normalization on each image sub-channel, so that the image sub-channel becomes a gray image.

[0115] Specifically, the numerical value of some elements in the coefficient matrix obtained by three-order wavelet decomposition is in the order of 10 -6 , and if the numerical value is not normalized, the result accuracy will be seriously lost. Therefore, before solving the first basic objective parameter of each image sub-channel, the 10 coefficient matrices need to be numerically normalized to become gray images.

[0116] S42, solving the first basic objective parameter of each image sub-channel and the second basic objective parameter of the initial gray image.

[0117] Specifically, the first basic objective parameter includes mean, variance, equivalent number of views, radiation resolution, image gradient amplitude spectrum and first image entropy, and the second basic objective parameter includes contrast based on gray level co-occurrence matrix, second image entropy and signal-to-noise ratio. In this embodiment, the first basic objective parameter is solved based on the gray image of the image sub-channel, and the second basic objective parameter is solved based on the initial gray image.

[0118] 1) solving the mean and variance of the image sub-channel.

[0119] Specifically, the mean value μ of the image sub-channel is obtained by summing and averaging the gray scale values of each pixel point of the image sub-channel. Then, the squared value of the difference between the gray scale value of each pixel point of the image sub-channel and the mean value is obtained, and then the squared value is summed and divided by the total number of pixel points to obtain the variance σ of the image sub-channel 2 .

[0120] 2) Solving the equivalent number of looks and the radiation resolution of the image sub-channel.

[0121] Specifically, the equivalent number of looks of the image sub-channel is obtained by squaring the mean value divided by the variance:

[0122]

[0123] where ENL is the equivalent number of looks, I is the gray scale image, μ is the mean value of the image, σ 2 is the variance of the image.

[0124] The radiation resolution of the image sub-channel is solved as a parameter of the equivalent number of looks, which is mathematically expressed as follows:

[0125]

[0126] where ENL is the equivalent number of looks, Γ is the radiation resolution.

[0127] 3) Solving the image gradient amplitude spectrum of the image sub-channel.

[0128] Specifically, the gray scale image of each image sub-channel is convolved with a Gaussian filter, and then the horizontal and vertical gradients of the image are obtained along the horizontal and vertical directions, respectively, to obtain the horizontal gradient spectrum and the vertical gradient spectrum of the image, and further to obtain the image gradient amplitude spectrum, which is mathematically expressed as follows:

[0129]

[0130]

[0131] where grad H (i,j) is the horizontal gradient spectrum of the image, grad V (i,j) is the vertical gradient spectrum of the image, grad(i,j) is the image gradient amplitude spectrum, I(i,j) is the gray scale image, and G(x,y;σ) is the Gaussian filter with a standard deviation of σ.

[0132] 4) Solving the contrast of the initial gray scale image based on the gray scale co-occurrence matrix.

[0133] Specifically, in order to reduce the amount of calculation, the initial gray scale image is compressed in gray scale, and 256 levels are compressed into 16 levels, i.e. the original 256 levels are re-quantized in every 16 levels. The frequency of two pixel points with gray values of i and j in the target direction (0°, 45°, 90°, 135°) is counted to obtain the contrast of the gray scale co-occurrence matrix in the corresponding direction. The contrast of the gray scale co-occurrence matrix is expressed by the following mathematical formula:

[0134]

[0135] wherein, L is the gray scale (L = 16), P d (i,j) is the frequency of two pixel points with gray values of i and j in the spatial position distance d.

[0136] 5) solving the first image entropy of the image sub-channel and the second image entropy of the initial gray scale image.

[0137] Specifically, the mathematical formula of the first image entropy and the second image entropy is as follows:

[0138]

[0139] wherein, i is the gray value of the pixel point, P(i) is the ratio of the pixel number with the gray value of i to the total number of image pixels, and n is the total number of gray scales.

[0140] 6) solving the signal-to-noise ratio of the initial gray scale image.

[0141] Specifically, the initial gray scale image is divided into 3*3 image blocks, and when the number of rows and columns of the input initial gray scale image is not an integer multiple of 3, it needs to be zero-padded, and then the local variance of all image blocks is solved, wherein the maximum value is the signal variance and the minimum value is the noise variance. The method for solving the local variance is as described in 1) solving the variance of the image sub-channel, and the signal-to-noise ratio is expressed by the following mathematical formula:

[0142]

[0143] wherein, sv is the signal variance, and nv is the noise variance.

[0144] Experience shows that the signal-to-noise ratio needs to be corrected to be more consistent with the actual situation, which is expressed by the following mathematical formula:

[0145] SNR new = SNR x 1.04-7

[0146] wherein, SNR new is the corrected signal-to-noise ratio, and SNR is the uncorrected signal-to-noise ratio.

[0147] S5, calculating a human visual system evaluation value of the imaging result map corresponding to the SAL echo signal according to the human visual system response weight of each of the image sub-channels and the first basic objective parameter, and solving the quality of the imaging result map by using the human visual system evaluation value and the second basic objective parameter. Specifically, the step includes:

[0148] S51, calculating a human visual system evaluation value of the imaging result map corresponding to the SAL echo signal according to the human visual system response weight of each of the image sub-channels and the mean, variance, equivalent number of views, radiation resolution, image gradient amplitude spectrum and first image entropy of the plurality of image sub-channels.

[0149] Specifically, the human visual system evaluation value is:

[0150]

[0151] wherein n is the number of image bits (n=8), i is a sub-channel, ω i is the HVS response weight on the sub-channel i, μ i is the mean of the sub-channel i, is the variance of the sub-channel i, ENL i is the equivalent number of views of the sub-channel i, Γ i is the radiation resolution of the sub-channel i, grad i is the image gradient amplitude spectrum of the sub-channel i, Entropy i is the first image entropy of the sub-channel i.

[0152] S52, solving the quality of the imaging result map by using the human visual system evaluation value and the contrast based on the gray level co-occurrence matrix, the signal-to-noise ratio, and the second image entropy.

[0153] Specifically, it is verified that the human visual system evaluation value, the contrast, and the signal-to-noise ratio show a downward trend with the decay of image quality, and the image entropy shows an upward trend with the decay of image quality. Physical quantities reflecting the same trend are classified, and physical quantities representing different trends are compared. The sum of the human visual system evaluation value, the contrast, and the signal-to-noise ratio of the image is divided by the second image entropy to obtain a parameter Quality representing the quality of the gray image. The mathematical formula is as follows:

[0154]

[0155] wherein Quality is the quality of the gray image to be measured, HVSNR is the human visual system evaluation value, Con is the contrast based on the gray level co-occurrence matrix, SNR is the signal-to-noise ratio, and Entropy is the second image entropy.

[0156] See Figures 3a-3d andFigures 4a-4b , Figures 3a-3d The imaging result graph after processing the measured data provided by the embodiment of the application is shown in the following table: Figure 3a The imaging result graph of the original SAL echo data is shown in the following table: Figure 3b The imaging result graph after adding a phase error of π / 8 in the original SAL echo data is shown in the following table: Figure 3c The imaging result graph after adding a phase error of π / 4 in the original SAL echo data is shown in the following table: Figure 3d The imaging result graph after adding a phase error of π / 2 in the original SAL echo data is shown in the following table: Figures 4a-4b The relationship curve between the human visual system evaluation value HVSNR and the defocus degree of the distorted image, and the relationship curve between the Quality and the defocus degree of the distorted image are shown in the following table: Figure 4a The evaluation results of the SAL imaging result graph in FIG. 3 by using the human visual system evaluation value HVSNR method are shown in the following table: Figure 4b The evaluation results of the SAL imaging result graph in FIG. 3 by using the Quality method are shown in the following table: Figure 4a Figure 4b It can be seen from the above tables that the human visual system evaluation value HVSNR and the Quality decrease with the increase of the defocus degree of the distorted image, which verifies the conclusion that the image quality decreases with the increase of the defocus degree of the image.

[0157] The embodiment does not use the network training idea, but solves the objective evaluation indicators related to image quality, such as contrast, information entropy, signal-to-noise ratio, and equivalent visual number, for the data normalized as a gray image. Considering that the evaluation terminal is related to the human eye visual system, and the human eye is more sensitive to low-frequency information in the image and not sensitive to high-frequency information such as noise, the embodiment uses multi-scale two-dimensional wavelet (third-order wavelet) decomposition to simulate the human visual system to divide the image into multiple image sub-channels, and uses a contrast sensitivity function to give each sub-channel a corresponding visual sensitivity weight, and finally obtains the quality of the entire image by weighted calculation. The embodiment has good discrimination and interpretation of the image quality of images with different blur and defocus degrees under the same scene, and has no requirements and restrictions on the image distortion type, and the operation algorithm is relatively simple and the calculation amount is relatively small.

[0158] ​The embodiment is directed to the evaluation of SAL image quality, based on the initial gray-scale image generated by SAL echo data, considering that the image quality evaluation result needs to be combined with human visual perception, the initial gray-scale image is subjected to three-order wavelet decomposition to simulate the human visual system, the contrast sensitivity function is used to give each sub-channel corresponding visual sensitivity weight, and the image basic objective parameter is solved on the decomposed image sub-channel, so as to obtain the human visual system evaluation value. The method mainly uses wavelet decomposition to simulate the response degree of the human visual system to different frequency components to evaluate the synthetic aperture laser radar image, does not rely on the image database of artificial scoring, and does not need to obtain the prior information of distorted images, only extracts the objective evaluation index representing the image quality in the spatial domain, then combines the contrast sensitivity function related to the frequency response to evaluate the image quality, and solves the image basic objective parameter from the pixel level using the pixel gray value, the solved basic objective parameter represents the essential and natural characteristics of the image, and does not change with the change of the data test set, has good distinction and interpretation for the image quality of different blur and defocus degrees in the same scene, has no requirement and limitation for the image distortion type, has certain universality, the operation algorithm is relatively simple, the calculation amount is small, and it is a method for effectively evaluating the image quality without relying on the image database of artificial scoring and the prior information of the image distortion type.

[0159] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, and all of them should be regarded as falling within the protection scope of the present application.

Claims

1. A SAL image quality assessment method based on fundamental objective parameters of an image, characterized in that, Including the following steps: Step 1: Generate an initial grayscale image of the input SAL echo signal; Step 2: Perform third-order wavelet decomposition on the initial grayscale image to generate multiple image sub-channels; Step 3: Calculate the contrast sensitivity function value of the corresponding image sub-channel using the spatial frequency of each image sub-channel to obtain the human visual system response weight; include: The spatial frequency of each image sub-channel is calculated using the image grayscale value: in, This represents the row frequency of the image sub-channel. The column frequency of the image sub-channel. For the number of rows and columns, element coordinates The image grayscale value, Spatial frequency of image sub-channels; The spatial frequency value is normalized to a range of 0 to 50 to obtain the normalized spatial frequency. The contrast sensitivity function values ​​of the image sub-channels are calculated using the normalized spatial frequency and then normalized to obtain the human visual system response weights for each sub-channel. The formula for calculating the contrast sensitivity function value is as follows: in, For normalized spatial frequency, and These are the normalized spatial frequencies in the horizontal and vertical directions. Step 4: Solve for the first basic objective parameters of each image sub-channel and the second basic objective parameters of the initial grayscale image; the first basic objective parameters include: mean, variance, equivalent number of views, radiometric resolution, image gradient magnitude spectrum and first image entropy; the second basic objective parameters include: contrast based on the gray-level co-occurrence matrix, second image entropy and signal-to-noise ratio. Step 5: Calculate the human visual system evaluation value of the imaging result map corresponding to the SAL echo signal based on the human visual system response weight of each image sub-channel and the first basic objective parameter, and use the human visual system evaluation value and the second basic objective parameter to solve for the quality of the imaging result map.

2. The SAL image quality assessment method based on fundamental objective parameters of an image according to claim 1, characterized in that, Step 1 includes: The SAL echo signal is sequentially format-converted and numerically normalized to generate the initial grayscale image.

3. The SAL image quality evaluation method based on fundamental objective parameters of an image according to claim 1, characterized in that, Step 2 includes: The initial grayscale image is subjected to third-order wavelet decomposition. During each decomposition, MATLAB built-in functions are called to decompose the image into horizontal detail coefficient matrix, vertical detail coefficient matrix, diagonal detail coefficient matrix and approximation coefficient matrix. At the same time, the image to be decomposed is downsampled by 2 times during each decomposition to obtain the decomposed image. The horizontal detail coefficient matrix, vertical detail coefficient matrix, diagonal detail coefficient matrix, and approximation coefficient matrix after each decomposition, as well as the approximation coefficient matrix after the last decomposition, are selected as the 10 sub-channels after the third-order wavelet decomposition: Channels = [cA3, cH3, cV3, cD3, cH2, cV2, cD2, cH1, cV1, cD1] Where cA3 is the approximate coefficient matrix after third-order wavelet decomposition, cHj is the horizontal detail coefficient matrix, cVj is the vertical detail coefficient matrix, cDj is the diagonal detail coefficient matrix, and j is the decomposition level.

4. The SAL image quality evaluation method based on fundamental objective parameters of an image according to claim 1, characterized in that, Step 4 includes: The image sub-channels are numerically normalized to convert them into grayscale images; Solve for the first fundamental objective parameter of each of the image sub-channels and the second fundamental objective parameter of the initial grayscale image.

5. The SAL image quality evaluation method based on fundamental objective parameters of an image according to claim 1, characterized in that, The method for calculating the mean includes: summing the gray values ​​of each pixel in the image sub-channel and then averaging them to obtain the mean of the image sub-channel; The method for calculating the variance includes: taking the difference between the gray value of each pixel in the image sub-channel and the mean, squaring the difference, summing the squared values, and then dividing by the total number of pixels to obtain the variance of the image sub-channel. The method for calculating the equivalent number of views includes: dividing the mean of the image sub-channel by the variance and then squaring the result to obtain the equivalent number of views of the image sub-channel. in, For equivalent apparent number, It is a grayscale image. The mean of the image. The variance of the image; The method for determining the radiometric resolution includes: calculating the radiometric resolution using the equivalent apparent number. in, For equivalent apparent number, For radiometric resolution; The formula for solving the image gradient magnitude spectrum is: in, The horizontal gradient spectrum of the image. For the vertical gradient spectrum of the image, The image gradient magnitude spectrum, It is a grayscale image. The standard deviation is Gaussian filter; The method for solving the contrast based on the gray-level co-occurrence matrix is ​​as follows: The initial gray-level image is compressed from 256 levels to 16 levels. Then, the gray-level values ​​of two pixels in the target direction are statistically analyzed. and The frequency of occurrence is used to obtain the contrast ratio based on the gray-level co-occurrence matrix in the corresponding direction: in, Grayscale =16, Spatial location distance is The gray values ​​of the two pixels are respectively and Frequency of occurrence; The formulas for calculating the first image entropy and the second image entropy are both: in, The grayscale value of a pixel. The grayscale value is The ratio of the number of pixels to the total number of pixels in the image, where n is the total number of gray levels; The formula for calculating the signal-to-noise ratio is: in, For signal variance, For noise variance, This is the corrected signal-to-noise ratio. This represents the signal-to-noise ratio before correction.

6. The SAL image quality evaluation method based on fundamental objective parameters of an image according to claim 1, characterized in that, Step 5 includes: The human visual system evaluation value of the imaging result map corresponding to the SAL echo signal is calculated based on the human visual system response weight of each image sub-channel and the mean, variance, equivalent number of views, radiometric resolution, image gradient magnitude spectrum and first image entropy of the plurality of image sub-channels. The quality of the imaging result image is determined using the human visual system evaluation value, the contrast based on the gray-level co-occurrence matrix, the signal-to-noise ratio, and the second image entropy.

7. The SAL image quality evaluation method based on fundamental objective parameters of an image according to claim 6, characterized in that, The human visual system evaluation value is: in, The image bit depth is n=8. For sub-channels, Sub-channel HVS response weights on Sub-channel The mean, Sub-channel variance Sub-channel The equivalent number of views, Sub-channel radiometric resolution, Sub-channel Image gradient magnitude spectrum Sub-channel The first image entropy.

8. The SAL image quality evaluation method based on fundamental objective parameters of an image according to claim 6, characterized in that, The quality of the imaging result image is: in, To determine the quality of the grayscale image to be tested, As an evaluation value for the human visual system, For contrast based on the gray-level co-occurrence matrix, For signal-to-noise ratio, This is the entropy of the second image.

Images