CN107862678B - Fundus image non-reference quality evaluation method - Google Patents

Abstract

The invention discloses a no-reference quality evaluation method for a fundus image, which includes two processes: a training phase and a testing phase; it takes into account the influence of brightness, naturalness and structural layout on the quality of the fundus image, extracts dark channel specific gravity features, bright The channel proportion feature, non-uniform brightness feature, naturalness quality evaluation score and structural layout index constitute the feature vector, and then use support vector regression to train the feature vector of all fundus images in the training image set to construct a quality prediction model; in the testing phase , by calculating the feature vector of the fundus image used as a test, and according to the quality prediction model constructed in the training stage, the objective quality evaluation prediction value of the fundus image is predicted, because the obtained feature vector information can better reflect the quality of the fundus image Therefore, the correlation between objective evaluation results and subjective perceptions is effectively improved.

Description

A no-reference quality evaluation method for fundus images

technical field

The invention relates to an image quality evaluation method, in particular to a reference-free quality evaluation method of fundus images.

Background technique

The fundus image is captured by a special fundus camera. The fundus image includes the main physiological structures such as the optic disc, macula, and blood vessels in the retina, and is an important type of image in medical imaging. Among them, the optic disc appears as an approximately circular bright color area in the normal fundus image, with the strongest contrast with the background area, which is the starting area of the optic nerve and blood vessels; the macula, because of its rich lutein, is in the fundus image. It appears as a dark area with no vascular structure, and there is an inwardly concave area in the center of the macula called the fovea; the blood vessels start from the optic disc area and extend to the inside of the entire eyeball, showing a tree-like distribution in the entire fundus image , the blood vessels in the optic disc region are the thickest and densest, and extend substantially in the vertical direction.

High-quality fundus images can help ophthalmologists diagnose various fundus diseases, as well as systemic diseases related to retinopathy. However, in the imaging process, there are often problems such as bright illumination, dark illumination, uneven illumination, blurring, low contrast and unreasonable layout, resulting in the acquired fundus images that cannot be used for diagnosis and need to be re-shot, which greatly reduces the efficiency and increase the cost of medical diagnosis. Therefore, it becomes crucial to automatically evaluate the image quality and recommend whether retakes are necessary while taking the fundus image.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a reference-free quality evaluation method for fundus images, which can effectively improve the correlation between objective evaluation results and subjective perception, and can accurately and automatically evaluate the quality of fundus images to determine whether the need for Retake the fundus image.

The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a method for evaluating the quality of fundus images without reference, which is characterized in that it includes two processes: a training phase and a testing phase;

The specific steps of the training phase process are:

①_1. Select N fundus images to form a training image set, denoted as {I _k |1≤k≤N}; among them, N is a positive integer, N>1, k is a positive integer, 1≤k≤N, I _k represents The k-th fundus image in {I _k |1≤k≤N}, the width of each fundus image in {I _k |1≤k≤N} is W, and the height is H;

并计算{I_k|1≤k≤N}中的每幅眼底图像的自然度特征矢量，将I_k的自然度特征矢量记为

计算{I_k|1≤k≤N}中的每幅眼底图像的结构布局特征矢量，将I_k的结构布局特征矢量记为

其中，

的维数为3×1，

的维数为1×1，

的维数为1×1；①_2. Calculate the luminance feature vector of each fundus image in {I _k |1≤k≤N}, and denote the luminance feature vector of I _k as

and calculate the naturalness feature vector of each fundus image in {I _k |1≤k≤N}, and denote the naturalness feature vector of I _k as

Calculate the structural layout feature vector of each fundus image in {I _k |1≤k≤N}, and denote the structural layout feature vector of I _k as

in,

The dimension of is 3 × 1,

The dimension of is 1 × 1,

The dimension is 1 × 1;

其中，F_k的维数为5×1，符号“[ ]”为矢量表示符号，

表示将

和

连接起来形成一个特征矢量，

为

的转置；①_3. Arrange the luminance feature vector, naturalness feature vector and structural layout feature vector of each fundus image in {I _k |1≤k≤N} in sequence to form each of {I _k |1≤k≤N} is the feature vector of a fundus image, and the feature vector of I _k is denoted as F _k ,

Among them, the dimension of F _k is 5 × 1, and the symbol "[ ]" is a vector representation symbol,

means to

and

concatenated to form a feature vector,

for

transpose of ;

其中，f( )为函数表示形式，F用于表示眼底图像的特征矢量，且作为质量预测模型的输入矢量，(w^opt)^T为w^opt的转置，

为F的线性函数；①_4. The respective feature vectors and subjective quality recommendation values of all fundus images in {I _k |1≤k≤N} form a training sample data set, and the training sample data set contains N feature vectors and N subjective quality recommendation values. ; Then use support vector regression as a machine learning method to train all feature vectors in the training sample data set, so that the error between the regression function value obtained after training and the subjective quality recommendation value is the smallest, and the optimal fitting is obtained. The weight vector w ^opt and the optimal bias term b ^opt ; then use the optimal weight vector w ^opt and the optimal bias term b ^opt to construct a quality prediction model, denoted as f(F),

Among them, f( ) is the function representation, F is used to represent the feature vector of the fundus image, and is used as the input vector of the quality prediction model, (w ^opt ) ^T is the transpose of w ^opt ,

is a linear function of F;

The specific steps of the test phase process are:

其中，I_test的宽度为W'，且高度为H'，F_test的维数为5×1，

为F_test的线性函数。2. For any fundus image I _test used as a test, follow the process of step ①_2 to step ①_3, with the same operation, obtain the feature vector of I _test , and denote it as F _test ; then according to the quality prediction model constructed in the training stage to F _test is tested, and the predicted value corresponding to F _test is predicted, and the predicted value is regarded as the quality objective evaluation predicted value of I _test , which is denoted as Q _test ,

Among them, the width of I _test is W', and the height is H', and the dimension of F _test is 5 × 1,

is a linear function of F _test .

的获取过程为：In the mentioned steps ①_2

The acquisition process is:

将

中坐标位置为(x,y)的像素点的像素值记为

其中，1≤x≤W，1≤y≤H，I_k(x,y)表示I_k中坐标位置为(x,y)的像素点的像素值，T_low为暗通道阈值；①_2a1, calculate the dark channel mask image of I _k , denoted as

Will

The pixel value of the pixel whose middle coordinate position is (x, y) is recorded as

Among them, 1≤x≤W, 1≤y≤H, I _k (x, y) represents the pixel value of the pixel point whose coordinate position is (x, y) in I _k , and T _low is the dark channel threshold;

将

中坐标位置为(x,y)的像素点的像素值记为

其中，T_high为亮通道阈值；and calculate the bright channel mask image of I _k , denoted as

Will

The pixel value of the pixel whose middle coordinate position is (x, y) is recorded as

Among them, T _high is the bright channel threshold;

中的所有像素点的像素值的均值，作为I_k的暗通道比重特征，记为

①_2b1, calculation

The mean of the pixel values of all the pixels in , as the dark channel proportion feature of I _k , is denoted as

中的所有像素点的像素值的均值，作为I_k的亮通道比重特征，记为

and calculate

The mean of the pixel values of all the pixels in , as the bright channel proportion feature of I _k , is denoted as

①_2c1. Divide I _k into multiple overlapping sub-blocks with a size of 9 × 9 pixels and a step size of 1 pixel; then randomly select M sub-blocks from all sub-blocks in I _k ; The pixel values of all the pixels in the sub-blocks form a column vector, and the column vector composed of the pixel values of all the pixels in the selected t-th sub-block is denoted as y _t ; wherein, M is a positive integer, 1000≤M≤M ^* , M ^* represents the total number of sub-blocks included in I _k , t is a positive integer, 1≤t≤M, and the dimension of y _t is 81×1;

其中，μ_t表示y_t中的所有元素的值的均值，也即表示选择的第t个子块中的所有像素点的像素值的均值，符号“|| ||”为求欧氏距离符号；①_2d1, calculate the non-uniform brightness feature of I _k , denoted as

Among them, μ _t represents the mean value of the values of all elements in y _t , that is, the mean value of the pixel values of all the pixels in the selected t-th sub-block, and the symbol “|| ||” is the Euclidean distance symbol;

和

按序排列构成的矢量作为I_k的亮度特征矢量

其中，

的维数为3×1，符号“[ ]”为矢量表示符号，

表示将

和

连接起来形成一个矢量，

为

的转置。①_2e1, will

and

The vector formed in order is used as the luminance feature vector of I _k

in,

The dimension of is 3 × 1, the symbol "[ ]" is a vector representation symbol,

means to

and

concatenated to form a vector,

for

transposition of .

的获取过程为：In the mentioned steps ①_2

The acquisition process is:

①_2a2. Select N' fundus images with excellent subjective quality recommendation value to form the training set; then use the natural image quality predictor to extract the original multivariate Gaussian model of the training set from the training set, denoted as (μ, C); among them, N ' is a positive integer, N'>1, μ represents the mean feature of (μ, C), and C represents the covariance matrix feature of (μ, C);

符号

为向下取整操作符号，t'为正整数，1≤t'≤M'，μ_t'表示(μ_t',C_t')的均值特征，C_t'表示(μ_t',C_t')的协方差矩阵特征；①_2b2. Divide I _k into M' non-overlapping sub-blocks with a size of 64×64 pixels; then use a natural image quality predictor to extract each sub-block in I _k from each sub-block in I _k The original _{multivariate Gaussian model} of the

symbol

is the symbol of the round-down operation, t' is a positive integer, 1≤t'≤M', μ _t' represents the mean feature of (μ _t' ,C _t' ), C _t' represents (μ _t' ,C _{t '} ) covariance matrix features;

其中，(μ-μ_t')^T为(μ-μ_t')的转置，

为

的逆；①_2c2. According to (μ, C) and the original multivariate Gaussian model of each sub-block in I _k , calculate the naturalness quality evaluation score of each sub-block in I _k , and calculate the naturalness of the t'th sub-block in I _k The quality evaluation score is recorded as q _t' ,

where (μ-μ _t' ) ^T is the transpose of (μ-μ _t' ),

for

the inverse of ;

然后直接将

作为I_k的自然度特征矢量

其中，

的维数为1×1。①_2d2, calculate the naturalness quality evaluation score of I _k , denoted as

then directly

Naturalness feature vector as I _k

in,

The dimension is 1×1.

的获取过程为：In the mentioned steps ①_2

The acquisition process is:

θ表示Log-Gabor滤波器的方向因子，

e_ω,θ(x,y)为G_ω,θ(x,y)的实部，o_ω,θ(x,y)为G_ω,θ(x,y)的虚部，符号“j”为虚数表示符号；①-2a3, use Log-Gabor filter to filter I _k , and obtain the frequency response of each pixel point in I _k under different center frequencies and different direction factors, and set the coordinate position in I _k as (x, y) The frequency response of a pixel at a center frequency of ω and a direction factor of θ is recorded as G _{ω, θ} (x, y), G _{ω, θ} (x, y) = e _{ω, θ} (x, y) + jo _{ω ,θ} (x,y); where 1≤x≤W, 1≤y≤H, ω represents the center frequency of the Log-Gabor filter,

θ represents the direction factor of the Log-Gabor filter,

e _{ω, θ} (x, y) is the real part of G _{ω, θ} (x, y), o _{ω, θ} (x, y) is the imaginary part of G _{ω, θ} (x, y), symbol "j" sign for imaginary numbers;

其中，

①_2b3. Calculate the phase consistency map of I _k , denoted as {PC _k (x, y)}, and denote the pixel value of the pixel point whose coordinate position is (x, y) in {PC _k (x, y)} as PC _k (x,y),

in,

其中，T_PC为二值化阈值；①_2c3. Calculate the binary blood vessel map of I _k , denoted as {B _k (x, y)}, and denote the pixel value of the pixel point whose coordinate position is (x, y) in {B _k (x, y)} as B _k (x,y),

Among them, T _PC is the binarization threshold;

其中，

表示求取使得

的值最小时的(x',y')，δ表示水平偏移位置，ε表示垂直偏移位置，

表示I_k中以坐标位置(x',y')为中心、半径为100像素的圆形区域，B_k(x'+δ,y'+ε)表示{B_k(x,y)}中坐标位置为(x'+δ,y'+ε)的像素点的像素值；①_2d3, calculate the center position of the optic disc of I _k , denoted as

in,

means to ask for

(x', y') when the value of is the smallest, δ represents the horizontal offset position, ε represents the vertical offset position,

Represents a circular area with a coordinate position (x', y') as the center and a radius of 100 pixels in I _k , and B _k (x'+δ, y'+ε) represents {B _k (x, y)} in The pixel value of the pixel whose coordinate position is (x'+δ, y'+ε);

是否落在规定的区域

内，如果是，则令S_k＝1，否则，令S_k＝0；再直接将S_k作为I_k的结构布局特征矢量

其中，

的维数为1×1。①_2e3, let S _k represent the structural layout index of I _k ; then judge

Whether it falls in the specified area

If yes, let _Sk = 1, otherwise, let _Sk = 0; then directly use _Sk as the structural layout feature vector of I _k

in,

The dimension is 1×1.

定义为：将I_k划分成8×8个互不重叠的尺寸大小为

像素的区域，将I_k中的第p个区域记为

然后水平扫描I_k，找出I_k中的第25个区域

第26个区域

第27个区域

第30个区域

第31个区域

第32个区域

第33个区域

第34个区域

第35个区域

第38个区域

第39个区域

第40个区域

再将找出的所有区域构成的集合作为规定的区域

其中，符号

为向下取整操作符号，p为正整数，1≤p≤64。In the step ①_2e3 described, the specified area

Defined as: dividing I _k into 8 × 8 non-overlapping sizes as

The area of pixels, denote the p-th area in I _k as

Then scan I _k horizontally to find the 25th region in I _k

26th area

27th area

30th area

31st area

32nd area

33rd area

34th area

35th area

38th area

39th area

40th area

Then take the set of all the found areas as the specified area

Among them, the symbol

In order to round down the operation symbol, p is a positive integer, 1≤p≤64.

Compared with the prior art, the advantages of the present invention are:

The method of the invention takes into account the influence of brightness, naturalness and structural layout on the quality of the fundus image, and extracts the dark channel proportion feature, the bright channel proportion feature, the non-uniform brightness feature, the naturalness quality evaluation score and the structural layout index to form a feature vector, Then use support vector regression to train the feature vectors of all fundus images in the training image set to construct a quality prediction model; in the testing phase, by calculating the feature vectors of the fundus images used for testing, and according to the quality prediction model constructed in the training phase, The predicted value of the objective evaluation of the quality of the fundus image is obtained by predicting, because the obtained feature vector information can better reflect the quality change of the fundus image, thus effectively improving the correlation between the objective evaluation result and the subjective perception.

Description of drawings

FIG. 1 is a block diagram of the overall implementation of the method of the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the embodiments of the accompanying drawings.

The overall implementation block diagram of a fundus image quality evaluation method without reference proposed by the present invention is shown in FIG. 1 , which includes two processes: a training phase and a testing phase.

The specific steps of the training phase process are:

①_1. Select N fundus images to form a training image set, denoted as {I _k |1≤k≤N}; among them, N is a positive integer, N>1, if N=1000, k is a positive integer, 1≤k ≤N, I _k represents the k-th fundus image in {I _k |1≤k≤N}, and each fundus image in {I _k |1≤k≤N} has a width of W and a height of H.

In this embodiment, a part of the fundus images in the fundus image database established by Ningbo University is randomly selected to constitute the training image set.

并计算{I_k|1≤k≤N}中的每幅眼底图像的自然度特征矢量，将I_k的自然度特征矢量记为

计算{I_k|1≤k≤N}中的每幅眼底图像的结构布局特征矢量，将I_k的结构布局特征矢量记为

其中，

的维数为3×1，

的维数为1×1，

的维数为1×1。①_2. Calculate the luminance feature vector of each fundus image in {I _k |1≤k≤N}, and denote the luminance feature vector of I _k as

and calculate the naturalness feature vector of each fundus image in {I _k |1≤k≤N}, and denote the naturalness feature vector of I _k as

Calculate the structural layout feature vector of each fundus image in {I _k |1≤k≤N}, and denote the structural layout feature vector of I _k as

in,

The dimension of is 3 × 1,

The dimension of is 1 × 1,

The dimension is 1×1.

的获取过程为：In this embodiment, in steps ①_2

The acquisition process is:

将

中坐标位置为(x,y)的像素点的像素值记为

其中，1≤x≤W，1≤y≤H，I_k(x,y)表示I_k中坐标位置为(x,y)的像素点的像素值，T_low为暗通道阈值，在本实施例中取T_low＝50。①_2a1, calculate the dark channel mask image of I _k , denoted as

Will

The pixel value of the pixel whose middle coordinate position is (x, y) is recorded as

Among them, 1≤x≤W, 1≤y≤H, I _k (x, y) represents the pixel value of the pixel point whose coordinate position is (x, y) in I _k , T _low is the dark channel threshold, in this implementation Take T _low =50 in the example.

将

中坐标位置为(x,y)的像素点的像素值记为

其中，T_high为亮通道阈值，在本实施例中取T_high＝240。and calculate the bright channel mask image of I _k , denoted as

Will

The pixel value of the pixel whose middle coordinate position is (x, y) is recorded as

Among them, T _high is the threshold value of the bright channel, and in this embodiment, T _high =240 is taken.

中的所有像素点的像素值的均值，作为I_k的暗通道比重特征，记为

①_2b1, calculation

The mean of the pixel values of all the pixels in , as the dark channel proportion feature of I _k , is denoted as

中的所有像素点的像素值的均值，作为I_k的亮通道比重特征，记为

and calculate

The mean of the pixel values of all the pixels in , as the bright channel proportion feature of I _k , is denoted as

①_2c1. Divide I _k into multiple overlapping sub-blocks with a size of 9 × 9 pixels and a step size of 1 pixel; then randomly select M sub-blocks from all sub-blocks in I _k ; The pixel values of all the pixels in the sub-blocks form a column vector, and the column vector composed of the pixel values of all the pixels in the selected t-th sub-block is denoted as y _t ; wherein, M is a positive integer, 1000≤M≤M ^* , M ^* represents the total number of sub-blocks included in I _k , in this embodiment, M=1000, t is a positive integer, 1≤t≤M, and the dimension of y _t is 81×1.

其中，μ_t表示y_t中的所有元素的值的均值，也即表示选择的第t个子块中的所有像素点的像素值的均值，符号“|| ||”为求欧氏距离符号。①_2d1, calculate the non-uniform brightness feature of I _k , denoted as

Among them, μ _t represents the mean of the values of all elements in y _t , that is, the mean of the pixel values of all the pixels in the selected t-th sub-block, and the symbol “|| ||” is the Euclidean distance symbol.

和

按序排列构成的矢量作为I_k的亮度特征矢量

其中，

的维数为3×1，符号“[ ]”为矢量表示符号，

表示将

和

连接起来形成一个矢量，

为

的转置。①_2e1, will

and

The vector formed in order is used as the luminance feature vector of I _k

in,

The dimension of is 3 × 1, the symbol "[ ]" is a vector representation symbol,

means to

and

concatenated to form a vector,

for

transposition of .

的获取过程为：In this embodiment, in steps ①_2

The acquisition process is:

①_2a2. Select N' fundus images with excellent subjective quality recommendation value to form the training set; then use the existing Natural Image Quality Evaluator (NIQE) to extract the original multivariate Gaussian (Pristine Gaussian) of the training set from the training set Multivariate Gaussian, MVG) model, denoted as (μ, C); among them, N' is a positive integer, N'>1, in this embodiment, N'=100, μ represents the mean feature of (μ, C), C represents the covariance matrix feature of (μ, C).

符号

为向下取整操作符号，t'为正整数，1≤t'≤M'，μ_t'表示(μ_t',C_t')的均值特征，C_t'表示(μ_t',C_t')的协方差矩阵特征。 _{①_2b2} . Divide I _k into M' non-overlapping sub-blocks with a size of 64 × 64 pixels; The original multivariate Gaussian model of each sub-block in _Ik is extracted from the block, and the original multivariate Gaussian model of the t'th sub-block in _Ik is recorded as (μ _t' ,C _t' ); where, M' is positive integer,

symbol

is the symbol of the round-down operation, t' is a positive integer, 1≤t'≤M', μ _t' represents the mean feature of (μ _t' ,C _t' ), C _t' represents (μ _t' ,C _{t '} ) of the covariance matrix feature.

其中，(μ-μ_t')^T为(μ-μ_t')的转置，

为

的逆。①_2c2. According to (μ, C) and the original multivariate Gaussian model of each sub-block in I _k , calculate the naturalness quality evaluation score of each sub-block in I _k , and calculate the naturalness of the t'th sub-block in I _k The quality evaluation score is recorded as q _t' ,

where (μ-μ _t' ) ^T is the transpose of (μ-μ _t' ),

for

inverse of .

然后直接将

作为I_k的自然度特征矢量

其中，

的维数为1×1。①_2d2, calculate the naturalness quality evaluation score of I _k , denoted as

then directly

Naturalness feature vector as I _k

in,

The dimension is 1×1.

的获取过程为：In this embodiment, in steps ①_2

The acquisition process is:

θ表示Log-Gabor滤波器的方向因子，

e_ω,θ(x,y)为G_ω,θ(x,y)的实部，o_ω,θ(x,y)为G_ω,θ(x,y)的虚部，符号“j”为虚数表示符号。①-2a3, use Log-Gabor filter to filter I _k , and obtain the frequency response of each pixel point in I _k under different center frequencies and different direction factors, and set the coordinate position in I _k as (x, y) The frequency response of a pixel at a center frequency of ω and a direction factor of θ is recorded as G _{ω, θ} (x, y), G _{ω, θ} (x, y) = e _{ω, θ} (x, y) + jo _{ω ,θ} (x,y); where 1≤x≤W, 1≤y≤H, ω represents the center frequency of the Log-Gabor filter,

θ represents the direction factor of the Log-Gabor filter,

e _{ω, θ} (x, y) is the real part of G _{ω, θ} (x, y), o _{ω, θ} (x, y) is the imaginary part of G _{ω, θ} (x, y), symbol "j" Signs for imaginary numbers.

其中，

①_2b3. Calculate the phase consistency map of I _k , denoted as {PC _k (x, y)}, and denote the pixel value of the pixel point whose coordinate position is (x, y) in {PC _k (x, y)} as PC _k (x,y),

in,

其中，T_PC为二值化阈值，在本实施例中取T_PC＝0.45。①_2c3. Calculate the binary blood vessel map of I _k , denoted as {B _k (x, y)}, and denote the pixel value of the pixel point whose coordinate position is (x, y) in {B _k (x, y)} as B _k (x,y),

Wherein, T _PC is a binarization threshold, and in this embodiment, T _PC =0.45.

其中，

表示求取使得

的值最小时的(x',y')，δ表示水平偏移位置，ε表示垂直偏移位置，

表示I_k中以坐标位置(x',y')为中心、半径为100像素的圆形区域，B_k(x'+δ,y'+ε)表示{B_k(x,y)}中坐标位置为(x'+δ,y'+ε)的像素点的像素值。①_2d3, calculate the center position of the optic disc of I _k , denoted as

in,

means to ask for

(x', y') when the value of is the smallest, δ represents the horizontal offset position, ε represents the vertical offset position,

Represents a circular area with a coordinate position (x', y') as the center and a radius of 100 pixels in I _k , and B _k (x'+δ, y'+ε) represents {B _k (x, y)} in The pixel value of the pixel whose coordinate position is (x'+δ, y'+ε).

是否落在规定的区域

内，如果是，则令S_k＝1，否则，令S_k＝0；再直接将S_k作为I_k的结构布局特征矢量

其中，

的维数为1×1。①_2e3, let S _k represent the structural layout index of I _k ; then judge

Whether it falls in the specified area

If yes, let _Sk = 1, otherwise, let _Sk = 0; then directly use _Sk as the structural layout feature vector of I _k

in,

The dimension is 1×1.

定义为：将I_k划分成8×8个互不重叠的尺寸大小为

像素的区域，将I_k中的第p个区域记为

然后水平扫描I_k，找出I_k中的第25个区域

第26个区域

第27个区域

第30个区域

第31个区域

第32个区域

第33个区域

第34个区域

第35个区域

第38个区域

第39个区域

第40个区域

再将找出的所有区域构成的集合作为规定的区域

其中，符号

为向下取整操作符号，p为正整数，1≤p≤64。In this embodiment, in steps ①_2e3, the specified area is

Defined as: dividing I _k into 8 × 8 non-overlapping sizes as

The area of pixels, denote the p-th area in I _k as

Then scan I _k horizontally to find the 25th region in I _k

26th area

27th area

30th area

31st area

32nd area

33rd area

34th area

35th area

38th area

39th area

40th area

Then take the set of all the found areas as the specified area

Among them, the symbol

In order to round down the operation symbol, p is a positive integer, 1≤p≤64.

其中，F_k的维数为5×1，符号“[ ]”为矢量表示符号，

表示将

和

连接起来形成一个特征矢量，

为

的转置。①_3. Arrange the luminance feature vector, naturalness feature vector and structural layout feature vector of each fundus image in {I _k |1≤k≤N} in sequence to form each of {I _k |1≤k≤N} is the feature vector of a fundus image, and the feature vector of I _k is denoted as F _k ,

Among them, the dimension of F _k is 5 × 1, and the symbol "[ ]" is a vector representation symbol,

means to

and

concatenated to form a feature vector,

for

transposition of .

其中，f( )为函数表示形式，F用于表示眼底图像的特征矢量，且作为质量预测模型的输入矢量，(w^opt)^T为w^opt的转置，

为F的线性函数。①_4. The respective feature vectors and subjective quality recommendation values of all fundus images in {I _k |1≤k≤N} form a training sample data set, and the training sample data set contains N feature vectors and N subjective quality recommendation values. ; Then use support vector regression as a machine learning method to train all feature vectors in the training sample data set, so that the error between the regression function value obtained after training and the subjective quality recommendation value is the smallest, and the optimal fitting is obtained. The weight vector w ^opt and the optimal bias term b ^opt ; then use the optimal weight vector w ^opt and the optimal bias term b ^opt to construct a quality prediction model, denoted as f(F),

Among them, f( ) is the function representation, F is used to represent the feature vector of the fundus image, and is used as the input vector of the quality prediction model, (w ^opt ) ^T is the transpose of w ^opt ,

is a linear function of F.

The specific steps of the test phase process are:

其中，I_test的宽度为W'，且高度为H'，W'可与W相同或不相同，H'可与H相同或不相同，F_test的维数为5×1，

为F_test的线性函数。2. For any fundus image I _test used as a test, follow the process of step ①_2 to step ①_3, with the same operation, obtain the feature vector of I _test , and denote it as F _test ; then according to the quality prediction model constructed in the training stage to F _test is tested, and the predicted value corresponding to F _test is predicted, and the predicted value is regarded as the quality objective evaluation predicted value of I _test , which is denoted as Q _test ,

Among them, the width of I _test is W', and the height is H', W' can be the same or different from W, H' can be the same or different from H, and the dimension of F _test is 5 × 1,

is a linear function of F _test .

In order to further illustrate the feasibility and effectiveness of the method of the present invention, the method of the present invention is tested.

In this embodiment, the fundus image database established by Ningbo University is used as the image database. The fundus image database established by Ningbo University includes 1000 fundus images, and each fundus image is assigned a subjective quality recommendation value of 1 or 0, where 1 represents The recommendation quality is excellent, and 0 means the recommendation quality is poor. The present invention utilizes 4 commonly used indicators to assess the quality of classification, namely sensitivity (Sensitivity), specificity (Specificity), accuracy (Accuracy), area under the ROC curve (AUC). The closer the area under the ROC curve is to 100%, the better the correlation between the objective evaluation result of the method of the present invention and the subjective quality recommendation value. Table 1 shows the correlation between the quality objective evaluation prediction value obtained by the method of the present invention and the subjective quality recommendation value. As can be seen from Table 1, even if the fundus images of different proportions are used to form the training image set, the method of the present invention is used to form the training image set. The correlation between the objective evaluation prediction value of the obtained fundus image quality and the subjective quality recommendation value is very high, which is sufficient to demonstrate the effectiveness of the method of the present invention.

Table 1 The correlation between the objective quality evaluation prediction value obtained by the method of the present invention and the subjective quality recommendation value

index 80% fundus image 60% fundus image 40% fundus image 20% fundus image Sensitivity 94.55% 94.58% 94.41% 94.01% Specificity 89.11% 88.82% 87.99% 84.38% Accuracy 92.07% 91.95% 91.49% 89.62% AUC 91.88% 91.68% 91.20% 89.21%

Claims (4)

1. a fundus image without reference quality evaluation method, is characterized in that comprising two processes of training stage and testing stage; The specific steps of the training phase process are: ①_1. Select N fundus images to form a training image set, denoted as {I _k |1≤k≤N}; among them, N is a positive integer, N>1, k is a positive integer, 1≤k≤N, I _k represents The k-th fundus image in {I _k |1≤k≤N}, the width of each fundus image in {I _k |1≤k≤N} is W, and the height is H; ①_2. Calculate the luminance feature vector of each fundus image in {I _k |1≤k≤N}, and denote the luminance feature vector of I _k as

and calculate the naturalness feature vector of each fundus image in {I _k |1≤k≤N}, and denote the naturalness feature vector of I _k as

Calculate the structural layout feature vector of each fundus image in {I _k |1≤k≤N}, and denote the structural layout feature vector of I _k as

in,

The dimension of is 3 × 1,

The dimension of is 1 × 1,

The dimension is 1 × 1; In the mentioned steps ①_2

The acquisition process is: ①_2a2. Select N' fundus images with excellent subjective quality recommendation value to form the training set; then use the natural image quality predictor to extract the original multivariate Gaussian model of the training set from the training set, denoted as (μ, C); among them, N ' is a positive integer, N'>1, μ represents the mean feature of (μ, C), and C represents the covariance matrix feature of (μ, C); ①_2b2. Divide I _k into M' non-overlapping sub-blocks with a size of 64×64 pixels; then use a natural image quality predictor to extract each sub-block in I _k from each sub-block in I _k The original _{multivariate Gaussian model} of the

symbol

is the symbol of the round-down operation, t' is a positive integer, 1≤t'≤M', μ _t' represents the mean feature of (μ _t' ,C _t' ), C _t' represents (μ _t' ,C _{t '} ) covariance matrix features; ①_2c2. According to (μ, C) and the original multivariate Gaussian model of each sub-block in I _k , calculate the naturalness quality evaluation score of each sub-block in I _k , and calculate the naturalness of the t'th sub-block in I _k The quality evaluation score is recorded as q _t' ,

where (μ-μ _t' ) ^T is the transpose of (μ-μ _t' ),

for

the inverse of ; ①_2d2, calculate the naturalness quality evaluation score of I _k , denoted as

then directly

Naturalness feature vector as I _k

in,

The dimension is 1 × 1; ①_3. Arrange the luminance feature vector, naturalness feature vector and structural layout feature vector of each fundus image in {I _k |1≤k≤N} in sequence to form each of {I _k |1≤k≤N} is the feature vector of a fundus image, and the feature vector of I _k is denoted as F _k ,

Among them, the dimension of F _k is 5 × 1, and the symbol "[]" is a vector representation symbol,

means to

and

concatenated to form a feature vector,

for

transpose of ; ①_4. The respective feature vectors and subjective quality recommendation values of all fundus images in {I _k |1≤k≤N} form a training sample data set, and the training sample data set contains N feature vectors and N subjective quality recommendation values. ; Then use support vector regression as a machine learning method to train all feature vectors in the training sample data set, so that the error between the regression function value obtained after training and the subjective quality recommendation value is the smallest, and the optimal fitting is obtained. The weight vector w ^opt and the optimal bias term b ^opt ; then use the optimal weight vector w ^opt and the optimal bias term b ^opt to construct a quality prediction model, denoted as f(F),

Among them, f() is the function representation, F is used to represent the feature vector of the fundus image, and is used as the input vector of the quality prediction model, (w ^opt ) ^T is the transpose of w ^opt ,

is a linear function of F; The specific steps of the test phase process are: 2. For any fundus image I _test used as a test, follow the process of step ①_2 to step ①_3, with the same operation, obtain the feature vector of I _test , and denote it as F _test ; then according to the quality prediction model constructed in the training stage to F _test is tested, and the predicted value corresponding to F _test is predicted, and the predicted value is regarded as the quality objective evaluation predicted value of I _test , which is denoted as Q _test ,

Among them, the width of I _test is W', and the height is H', and the dimension of F _test is 5 × 1,

is a linear function of F _test . 2. a kind of fundus image no-reference quality evaluation method according to claim 1, is characterized in that in described step ①-2

The acquisition process is: ①_2a1, calculate the dark channel mask image of I _k , denoted as

Will

The pixel value of the pixel whose middle coordinate position is (x, y) is recorded as

Among them, 1≤x≤W, 1≤y≤H, I _k (x, y) represents the pixel value of the pixel point whose coordinate position is (x, y) in I _k , and T _low is the dark channel threshold; and calculate the bright channel mask image of I _k , denoted as

Will

The pixel value of the pixel whose middle coordinate position is (x, y) is recorded as

Among them, T _high is the bright channel threshold; ①_2b1, calculation

The mean of the pixel values of all the pixels in , as the dark channel proportion feature of I _k , is denoted as

and calculate

The mean of the pixel values of all the pixels in , as the bright channel proportion feature of I _k , is denoted as

①_2c1. Divide I _k into multiple overlapping sub-blocks with a size of 9 × 9 pixels and a step size of 1 pixel; then randomly select M sub-blocks from all sub-blocks in I _k ; The pixel values of all the pixels in the sub-blocks form a column vector, and the column vector composed of the pixel values of all the pixels in the selected t-th sub-block is denoted as y _t ; wherein, M is a positive integer, 1000≤M≤M ^* , M ^* represents the total number of sub-blocks included in I _k , t is a positive integer, 1≤t≤M, and the dimension of y _t is 81×1; ①_2d1, calculate the non-uniform brightness feature of I _k , denoted as

Among them, μ _t represents the mean value of the values of all elements in y _t , that is, the mean value of the pixel values of all the pixels in the selected t-th sub-block, and the symbol “|| ||” is the Euclidean distance symbol; ①_2e1, will

and

The vector formed in order is used as the luminance feature vector of I _k

in,

The dimension of is 3 × 1, the symbol "[]" is a vector representation symbol,

means to

and

concatenated to form a vector,

for

transposition of . 3. a kind of fundus image no-reference quality evaluation method according to claim 1, is characterized in that in described step ①-2

The acquisition process is: ①-2a3, use Log-Gabor filter to filter I _k , and obtain the frequency response of each pixel point in I _k under different center frequencies and different direction factors, and set the coordinate position in I _k as (x, y) The frequency response of a pixel at a center frequency of ω and a direction factor of θ is recorded as G _{ω, θ} (x, y), G _{ω, θ} (x, y) = e _{ω, θ} (x, y) + jo _{ω ,θ} (x,y); where 1≤x≤W, 1≤y≤H, ω represents the center frequency of the Log-Gabor filter,

θ represents the direction factor of the Log-Gabor filter,

e _{ω, θ} (x, y) is the real part of G _{ω, θ} (x, y), o _{ω, θ} (x, y) is the imaginary part of G _{ω, θ} (x, y), symbol "j" sign for imaginary numbers; ①_2b3. Calculate the phase consistency map of I _k , denoted as {PC _k (x, y)}, and denote the pixel value of the pixel point whose coordinate position is (x, y) in {PC _k (x, y)} as PC _k (x,y),

in,

①_2c3. Calculate the binary blood vessel map of I _k , denoted as {B _k (x, y)}, and denote the pixel value of the pixel point whose coordinate position is (x, y) in {B _k (x, y)} as B _k (x,y),

Among them, T _PC is the binarization threshold; ①_2d3, calculate the center position of the optic disc of I _k , denoted as

in,

means to ask for

(x', y') when the value of is the smallest, δ represents the horizontal offset position, ε represents the vertical offset position,

Represents a circular area with a coordinate position (x', y') as the center and a radius of 100 pixels in I _k , and B _k (x'+δ, y'+ε) represents {B _k (x, y)} in The pixel value of the pixel whose coordinate position is (x'+δ, y'+ε); ①_2e3, let S _k represent the structural layout index of I _k ; then judge

Whether it falls in the specified area

If yes, let _Sk = 1, otherwise, let _Sk = 0; then directly use _Sk as the structural layout feature vector of I _k

in,

The dimension is 1×1. 4. a kind of fundus image no-reference quality evaluation method according to claim 3, is characterized in that in described step ①-2e3, the specified area

Defined as: dividing I _k into 8 × 8 non-overlapping sizes as

The area of pixels, denote the p-th area in I _k as

Then scan I _k horizontally to find the 25th region in I _k

26th area

27th area

30th area

31st area

32nd area

33rd area

34th area

35th area

38th area

39th area

40th area

Then take the set of all the found areas as the specified area

Among them, the symbol

In order to round down the operation symbol, p is a positive integer, 1≤p≤64.

Images