Image quality evaluation method simulating human eye vision subjectivity

Abstract

The application provides a kind of image quality evaluation method simulating human eye vision subject, comprising the following steps: using differential excitation, gradient and convolution neural network to extract three kinds of features of input image, fuse three kinds of features to obtain fusion features, obtain processed features Y through attention mechanism to fusion features, reduce channel number to 1 through a convolution network with convolution kernel size 1 to features Y, and add BatchNorm normalization layer and Relu activation layer, enhance the nonlinear characteristics of the whole network and prevent the explosion or attenuation of gradient to obtain two-dimensional features C', flatten two-dimensional features C' to obtain one-dimensional feature vector N;Finally, one-dimensional feature vector N is input into three-layer MLP for full connection calculation, and the evaluation score of image quality is obtained.The application can more accurately evaluate the image quality of camera shooting from the perspective of human eye vision subject, and improve the effect of image monitoring and data analysis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method for evaluating image quality that simulates the subjective perception of the human eye. Background Technology

[0002] Images captured by automotive camera systems vary significantly depending on the time, weather, and other complex conditions. The difference between the camera's image and human subjective perception differs considerably. However, traditional image quality assessment methods often fail to adequately consider human visual perception, struggling to capture subtle yet crucial features such as texture and rapidly changing scenes. This hinders the effectiveness of image quality evaluation. For example, the CIE 188-2010 standard primarily evaluates automotive headlight illumination based on illuminance, lacking a focus on subjective human perception. While some existing methods exist for assessing image quality from automotive camera systems, they generally only focus on low-level features like brightness and contrast. These methods fail to reflect the complex differences in human perception. Therefore, a more accurate method is needed to evaluate camera image quality from a subjective human perspective, assessing the similarity between the image and human subjective perception to improve image monitoring and data analysis. Summary of the Invention

[0003] The purpose of this invention is to propose a method for evaluating image quality that simulates the subjective perception of the human eye, addressing the problems existing in the prior art.

[0004] The technical solution of the present invention is: the image quality evaluation method simulating human visual subjectivity, comprising the following steps:

[0005] The first step is to use differential activation, gradient, and convolutional neural networks to generate differential activation features F from the input image. de Gradient features are denoted as F g And the fundamental feature F of convolutional neural networks ain Extraction;

[0006] The second step is to fuse the three features of the image obtained in the first step to obtain the fused feature F. fusion2 ;

[0007] The third step is to use the fusion feature F obtained in the second step. fusion2 The attention mechanism is used to make the fused features more biased towards important information, thus obtaining the feature Y after the attention mechanism is processed;

[0008] The fourth step is to reduce the number of channels of the feature Y after the attention mechanism is processed by a convolutional network with a kernel size of 1, and add a BatchNorm normalization layer and a ReLU activation layer to enhance the non-linear characteristics of the entire network and prevent gradient explosion or decay, and finally obtain a 24×24 two-dimensional feature C′.

[0009] The fifth step is to flatten the two-dimensional feature C′ to obtain a one-dimensional feature vector N;

[0010] The sixth step is to input the one-dimensional feature vector N into a three-layer MLP for fully connected computation to obtain the image quality evaluation score.

[0011] A further proposed approach is to include the following specific methods for the first step described above:

[0012] ①The specific method for differential excitation feature extraction of an image is as follows:

[0013] First, Weber difference excitation calculations are performed using three local windows of different sizes: 3×3, 5×5, and 7×7, to obtain features at different scales. The calculation formulas for the Weber local description operator are shown in equations (1-1) and (1-2):

[0014]

[0015]

[0016] Then, the calculated differences are added as scalars, and the ratio between the sum of grayscale differences and the grayscale value of the center pixel is calculated. Next, the arctangent is solved, and finally the Weber difference excitation feature is obtained. The calculation process is shown in equation (1-3):

[0017]

[0018] In equations (1-1) to (1-3), x c x is the center pixel of the window. i For the remaining pixels outside the center, w represents the window of different sizes, using the window function. and Calculate the differential excitation characteristic ξ w (x c Based on three different window sizes, a three-layer pyramid feature map can be obtained. By setting the padding of the window calculation, the size of the three layers of features is adjusted to be the same, thereby obtaining the edges with obvious gray-level changes in the image. Then, simulating the human visual observation mechanism, the differential excitation feature of the image is obtained, denoted as F. de F de The dimensions are 512×512×3, and it contains three layers of features f1, f2, and f3.

[0019] ②The specific method for gradient feature extraction of an image is as follows:

[0020] Image gradient feature extraction is performed using the Prewitt operator, and the calculation process is shown in equation (1-4):

[0021]

[0022] The Prewitt operator used is 3×3 in size. The positions in the operator are encoded sequentially from the top left corner as P1 to P9 in a row-first, column-later manner. Formula g x That is, the gradient feature in the vertical direction of the image is obtained by subtracting the sum of the eigenvalues ​​in the first row and the sum of the eigenvalues ​​in the third row in the operator; formula g y This involves subtracting the sum of the first and third column eigenvalues ​​in the operator to obtain the gradient features in the horizontal direction of the image. (The last sentence, "g", appears to be a typo and can be omitted.) x With g y The gradient feature of the image obtained by adding them together is denoted as F. g ;

[0023] ③ The specific method for extracting basic features from the input image using a convolutional neural network is as follows:

[0024] Using the classic ResNet-101 convolutional neural network as the backbone network for CNN feature extraction, the basic features of the intermediate layers of the backbone network are denoted as F. ain Its output size is C×H×W, where C is the number of channels, H is the feature map height, and W is the feature map width.

[0025] A further proposed solution is that the specific methods for the second step mentioned above include:

[0026] ① Using formula (1-5) to define the basic feature F ain With gradient features F g The fusion is performed to obtain the fused feature F. fusion1 :

[0027]

[0028] In the formula, α and β are characteristic weighting coefficients, with α ranging from (0,1) and β = 1 - α; F fusion1 The feature fusion process, specifically the feature F extracted by the convolutional neural network. ain Multiply by the weight coefficient β, the gradient feature is multiplied by the weight coefficient α, then the two parts of the feature are added together and passed through the ResNet-101 network to obtain the final feature map with a size of 14×14×512.

[0029] ②Then use equation (1-6) to transform feature F fusion1 With differential excitation feature F de The fusion is performed to obtain the fused feature F.fusion1 :

[0030] F fusion2 =γF fusion1 +ξf de ∈R (H×W)×C (1-6)

[0031] In the formula, γ and ξ are feature weighting coefficients, with γ ranging from (0,1) and ξ = 1 - γ; the specific process is the feature F extracted in the previous step. ain Multiply by the weight coefficient Y, the gradient feature is multiplied by the weight coefficient ξ, and then the two parts of the feature are added together to obtain the final feature map with a size of 14×14×512.

[0032] A further proposed solution is that the specific methods for the third step mentioned above include:

[0033] ① Using equation (1-7) to fuse the feature F fusion2 Dimensionality reduction is achieved through convolutional neural networks to reduce the computational cost of the attention mechanism:

[0034] F0 = Conv(F fusion2 )∈ H×W×C (1-7)

[0035] ② The Reshape operation is performed using Equation (1-8) to change the feature map into the form required for attention calculation. The size of the X feature obtained after the Reshape operation is (H×W) rows and C columns, where each column represents the information of all channels of a pixel position in the fused and compressed feature map F0.

[0036] X = Reshape(F0) ∈ (H×W)×C (1-8)

[0037] ③ An attention mechanism is used to improve the performance of image quality assessment. The X features are input into the self-attention layer to calculate the attention matrix, as shown in the following formula:

[0038] Q = W Q ∈ (x)x (1-9)

[0039] K = W Q ∈ (x)x (1-10)

[0040] V = W V ∈ (x)x (1-11)

[0041] In the formula, Q is the query matrix, K is the key matrix, and V is the value matrix. All three matrices are obtained through linear transformations of the X feature and are special mappings of the X feature to represent relevant information at specific locations in the feature map; the weight matrix W...Q W K W V The attention is dynamically adjusted, and then the overall attention H is calculated as shown in equation (1-12):

[0042]

[0043] The overall attention H is calculated as follows: the transpose of the key matrix K is multiplied by the query matrix Q to obtain the cross-relationship matrix between features at each location in the feature map; then, the cross-relationship matrix is ​​normalized by dividing by the square root of dimension d to obtain the attention weight matrix; subsequently, a softmax operation is performed on the obtained attention weight matrix to normalize the attention weights, making them satisfy the properties of a probability distribution, thus transforming the input vector into a probability distribution form. In this way, each weight becomes a value between 0 and 1, and the sum of all weights is 1, ensuring that the model does not favor any particular part when fusing information, but rather obtains information evenly from different features; finally, the value matrix V is multiplied by the attention weight matrix to obtain the overall attention H.

[0044] ④ The obtained overall attention H is reshaped using equation (1-13) to obtain the feature Y:

[0045] Y = Reshape(H) ∈ H×W×C (1-13).

[0046] A further proposed solution is as follows: The specific method for the fourth step mentioned above is:

[0047] The two-dimensional feature C′ is calculated using equation (1-14):

[0048] C' = Relu(BN(Conv) 1×1 (Y)))∈ H×W×1 (1-14).

[0049] In the formula, Conv 1×1 (·) represents a convolution operation with a kernel size of 1, which compresses the acquired H×W×C features from C to 1 in the channel dimension. BN represents the BatchNorm layer, which aims to keep the input of each neural network layer in the same distribution and prevent the gradient vanishing of the lower neural network during backpropagation. ReLU is an activation function that reduces the number of parameters and increases nonlinear relationships, thereby enhancing the expressive power of the network.

[0050] A further proposed solution is as follows: The specific method for step five above is:

[0051] Using equation (1-15), the obtained two-dimensional feature C′ is flattened to obtain a one-dimensional feature vector N of 1×(H×W):

[0052] N=flatten(C′)∈ 1×(H×W) (1-15)

[0053] In the formula, flatten(·) represents the flattening operation, which unfolds the feature C′ with size H×W×1 into the feature N with size 1×(H×W).

[0054] This invention offers several advantages: it comprehensively utilizes Weber descriptors, differential excitation feature extraction, gradient direction, and other techniques, while incorporating deep learning's attention mechanism to achieve more accurate image quality assessment. Specifically, by introducing Weber descriptors to capture the structural and textural information of images, it can better measure image quality. The differential excitation feature extraction module effectively captures rapidly changing scenes in images, such as traffic moments and sudden braking, improving the sensitivity of image quality assessment. The introduction of gradient direction enhances the diversity of image features, making the assessment more comprehensive. Furthermore, the deep learning attention mechanism enables convolutional neural networks to focus on key areas related to human visual perception, such as key objects, light sources, and texture features, when evaluating image quality. This helps to more accurately simulate the human eye's perception process, thus significantly improving the reliability of image quality assessment results compared to existing technologies. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0057] (Example 1)

[0058] The image quality evaluation method simulating human visual subjectivity in this embodiment uses multi-scale Weber excitation, differential excitation to extract features, gradient feature extraction, and deep learning neural networks to evaluate the gap between the image captured by the camera and human visual perception, and finally evaluates whether the image conforms to human visual subjectivity.

[0059] See Figure 1 The image quality evaluation method simulating human visual subjectivity in this embodiment is implemented according to the following steps:

[0060] The first step involves extracting three features from the input image using differential activation, gradient, and convolutional neural networks, as detailed below:

[0061] ①The specific method for differential excitation feature extraction of an image is as follows:

[0062] First, Weber difference excitation calculations are performed using three local windows of different sizes to obtain features at different scales. The different window sizes used are 3×3, 5×5, and 7×7. The calculation formulas for the Weber local description operator are shown in equations (1-1) and (1-2).

[0063]

[0064]

[0065] Then, the calculated differences are added as scalars, and the ratio between the sum of grayscale differences and the grayscale value of the center pixel is calculated. Next, the arctangent is solved, and finally the Weber difference excitation feature is obtained. The calculation process is shown in equation (1-3):

[0066]

[0067] In equations (1-1) to (1-3), x c x is the center pixel of the window. i For the remaining pixels outside the center, w represents the window of different sizes, using the window function. and Calculate the differential excitation characteristic ξ w (x c Based on three different window sizes, a three-layer pyramid feature map can be obtained. By setting the padding of the window calculation, the dimensions of the three layers of features are adjusted to be the same, thereby obtaining the edges with obvious gray-level changes in the image. This simulates the human visual observation mechanism, thus obtaining the differential excitation features of the image, denoted as F. de F de The dimensions are 512×512×3, which contains three layers of features f1, f2, and f3.

[0068] ②The specific method for gradient feature extraction of an image is as follows:

[0069] Image gradient feature extraction is performed using the Prewitt operator, and the calculation process is shown in equation (1-4).

[0070]

[0071] The Prewitt operator used is 3×3 in size. The positions in the operator are encoded sequentially from the top left corner as P1 to P9 in a row-first, column-later manner. Formula g x That is, the gradient feature in the vertical direction of the image is obtained by subtracting the sum of the eigenvalues ​​in the first row and the sum of the eigenvalues ​​in the third row in the operator; formula g y This involves subtracting the sum of the first and third column eigenvalues ​​in the operator to obtain the gradient features in the horizontal direction of the image. (The last sentence, "g", appears to be a typo and can be omitted.) x With g yThe gradient features of the image are obtained by summing them, denoted as F. g .

[0072] ③ The specific method for extracting basic features from the input image using a convolutional neural network is as follows:

[0073] Using the classic ResNet-101 convolutional neural network as the backbone network for CNN (Convolutional Neural Network) feature extraction, the basic features of the intermediate layers of the backbone network are denoted as F. ain Its output size is C×H×W, where C is the number of channels, H is the feature map height, and W is the feature map width.

[0074] The second step is to fuse the three features of the image obtained in the first step, as follows:

[0075] ① Using formula (1-5) to define the basic feature F ain With gradient features F g The fusion is performed to obtain the fused feature F. fusion1 :

[0076]

[0077] In the formula, α and β are characteristic weighting coefficients, with α ranging from (0,1) and β = 1 - α; F fusio1 The feature fusion process, specifically the feature F extracted by the convolutional neural network. ain Multiply by the weight coefficient β, the gradient feature is multiplied by the weight coefficient α, then the two feature parts are added together and passed through the ResNet-101 network to obtain the final feature map with a size of 14×14×512.

[0078] ②Use equation (1-6) to define feature F fusion1 With differential excitation feature F de The fusion is performed to obtain the fused feature F. fusion2 :

[0079] F fusion2 =γF fusion1 +ξf de ∈R (H×W)×C (1-6)

[0080] In the formula, γ and ξ are feature weighting coefficients, with γ ranging from (0,1) and ξ = 1 - γ; the specific process is the feature F extracted in the previous step. ain Multiply by the weight coefficient Y, the gradient feature is multiplied by the weight coefficient ξ, and then the two parts of the feature are added together to obtain the final feature map with a size of 14×14×512.

[0081] The third step is to obtain the fusion feature F fusion2The attention mechanism is used to make features more biased towards important information, thus obtaining the feature Y after attention mechanism processing; the specific method is as follows:

[0082] ① Using equation (1-7) to fuse the feature F fusion2 Dimensionality reduction is achieved through convolutional neural networks to reduce the computational cost of the attention mechanism:

[0083] F0 = Conv(F fsion2 )∈ H×W×C (1-7)

[0084] ② The Reshape operation is performed using Equation (1-8) to change the feature map into the form required for attention calculation. The size of the X feature obtained after the Reshape operation is (H×W) rows and C columns, where each column represents the information of all channels of a pixel position in the fused and compressed feature map F0.

[0085] X = Reshape(F0) ∈ (H×W)×C (1-8)

[0086] ③ An attention mechanism is used to improve the performance of image quality assessment. The fused feature X is input into the self-attention layer to calculate the attention matrix, as shown in the following formula:

[0087] Q = W Q ∈ (x)x (1-9)

[0088] K = W Q ∈ (x)x (1-10)

[0089] V = W V ∈ (x)x (1-11)

[0090] In the formula, Q, K, and V are all obtained through linear transformations of X, and are special mappings of feature X to represent relevant information at specific locations in the feature map. The query matrix Q guides the model to focus on specific parts of the image. Through the query vector obtained by linear transformation with the input data, the model can determine which locations or feature points to pay more attention to. The key matrix K corresponds to the query and represents the source of information the model focuses on. In the attention mechanism, the key vector is obtained through linear transformation with the input data and matched with the query vector to calculate attention weights. The value matrix V provides data containing actual information, which is weighted and summed according to the attention weights to generate the final output. The weight matrix W... Q W K W VThis can be dynamically adjusted through attention, followed by the overall attention H-related calculation as shown in equation (1-12):

[0091]

[0092] The overall attention mechanism calculation process is as follows: The transpose of the key matrix K is multiplied by the query matrix Q to obtain the cross-relationship matrix between features at each location in the feature map. Then, the cross-relationship matrix is ​​normalized by dividing by the square root of dimension d to obtain the attention weight matrix. This operation aims to prevent model overfitting and accelerate the training process, making the attention mechanism more stable and avoiding numerical problems caused by excessively large feature dimensions. Next, a softmax operation is performed on the obtained attention weight matrix. The main function of the softmax function is to normalize the attention weights to satisfy the properties of a probability distribution, transforming the input vector into a probability distribution form. This results in each weight becoming a value between 0 and 1, and the sum of all weights is 1, ensuring that the model does not favor any particular part during information fusion and obtains information evenly from different features. Finally, the value matrix V is multiplied by the attention weight matrix to obtain the output H.

[0093] ④ Perform a reshape operation on the obtained output H using equation (1-13) to obtain the feature Y:

[0094] Y = Reshape(H) ∈ H×W×C (1-13)

[0095] The fourth step involves passing the obtained feature Y through a convolutional network with a kernel size of 1 to reduce the number of channels to 1, and adding a BatchNorm normalization layer and a ReLU activation layer to enhance the nonlinear characteristics of the entire network and prevent gradient explosion or decay. Finally, a 24×24 two-dimensional feature C′ is obtained, as shown in equation (1-14).

[0096] C' = Relu(BN(Conv) 1×1 (Y)))∈ H×W×1 (1-14)

[0097] In the formula, Conv 1×1 (·) represents a convolution operation with a kernel size of 1, which compresses the acquired H×W×C features from C to 1 in the channel dimension. BN represents the BatchNorm layer, which aims to keep the input of each neural network layer in the same distribution and prevent the gradient vanishing of the lower neural network during backpropagation. ReLU is an activation function that reduces the number of parameters and increases nonlinear relationships, thereby enhancing the expressive power of the network.

[0098] The fifth step is to flatten the two-dimensional feature C′ to obtain a one-dimensional feature vector, specifically:

[0099] Using equation (1-15), the obtained two-dimensional feature C′ is flattened to obtain a one-dimensional feature vector N of 1×(H×W):

[0100] N=flatten(C′)∈ 1×(H×W) (1-15)

[0101] In the formula, flatten(·) represents the flattening operation, which unfolds the feature C′ with size H×W×1 into the feature N with size 1×(H×W);

[0102] The sixth step is to input the one-dimensional feature vector N into a three-layer MLP for fully connected computation to obtain the image quality evaluation score.

[0103] In this step, MLP stands for Multilayer Perceptron, also known as the MLP algorithm, or feedforward neural network or artificial neural network. It is existing technology and will not be elaborated further.

[0104] The above embodiments are descriptions of specific implementations of the present invention, and not limitations thereof. Those skilled in the art can make various modifications and changes without departing from the spirit and scope of the present invention to obtain corresponding equivalent technical solutions. Therefore, all equivalent technical solutions should be included in the patent protection scope of the present invention.

Claims

1. A method for simulating subjective human visual perception in image quality evaluation, characterized in that, Includes the following steps: The first step is to use differential activation, gradient, and convolutional neural networks to generate differential activation features F from the input image. de Gradient features are denoted as F g And the fundamental feature F of convolutional neural networks ain Extraction; The second step is to fuse the three features of the image obtained in the first step to obtain the fused feature F. fusion2 ; The third step is to use the fusion feature F obtained in the second step. fusion2 The attention mechanism is used to obtain the feature Y after processing. The fourth step is to reduce the number of channels of the feature Y after the attention mechanism is processed by a convolutional network with a kernel size of 1, and add a BatchNorm normalization layer and a ReLU activation layer to finally obtain a 24×24 two-dimensional feature C′. The fifth step is to flatten the two-dimensional feature C′ to obtain a one-dimensional feature vector N; The sixth step is to input the one-dimensional feature vector N into a three-layer MLP for fully connected computation to obtain the image quality evaluation score. The specific method for differential excitation feature extraction of the image in the first step is as follows: First, Weber difference excitation calculations are performed using three local windows of different sizes: 3×3, 5×5, and 7×7, to obtain features at different scales. The calculation formulas for the Weber local description operator are shown in equations (1-1) and (1-2): ； ； Then, the calculated differences are added as scalars, and the ratio between the sum of grayscale differences and the grayscale value of the center pixel is calculated. Next, the arctangent is solved, and finally the Weber difference excitation feature is obtained. The calculation process is shown in equation (1-3): ； In equations (1-1) to (1-3), x c x is the center pixel of the window. i For the remaining pixels outside the center, w represents the window of different sizes, using the window function. and Calculate the differential excitation characteristic ξw(x) c Based on three different window sizes, a three-layer pyramid feature map can be obtained. By setting the padding of the window calculation, the size of the three layers of features is adjusted to be the same, thereby obtaining the edges with obvious gray-level changes in the image. Then, simulating the human visual observation mechanism, the differential excitation feature of the image is obtained, denoted as F. de F de The dimensions are 512×512×3, which contains three layers of features f1, f2, and f3.

2. The image quality evaluation method simulating subjective human visual perception according to claim 1, characterized in that, The specific methods of the first step include: The specific method for gradient feature extraction from an image is as follows: Image gradient feature extraction is performed using the Prewitt operator, and the calculation process is shown in equation (1-4): ； ； The Prewitt operator used is 3×3 in size. The positions in the operator are encoded sequentially from the top left corner as P1 to P9 in a row-first, column-later manner. Formula g x That is, the gradient feature in the vertical direction of the image is obtained by subtracting the sum of the eigenvalues ​​in the first row and the sum of the eigenvalues ​​in the third row in the operator; formula g y That is, the difference between the sum of the first column of eigenvalues ​​and the sum of the third column of eigenvalues ​​in the operator is used to obtain the gradient features in the horizontal direction of the image. x With g y The gradient feature of the image obtained by adding them together is denoted as Fg; ③ The specific method for extracting basic features from the input image using a convolutional neural network is as follows: Using the classic ResNet-101 convolutional neural network as the backbone network for CNN feature extraction, the basic features of the intermediate layers of the backbone network are denoted as F. ain Its output size is C×H×W, where C is the number of channels, H is the feature map height, and W is the feature map width.

3. The image quality evaluation method simulating subjective human visual perception according to claim 1, characterized in that, The specific methods of the second step include: ① The basic feature Fain and the gradient feature Fg are fused using equation (1-5) to obtain the fused feature F. fusion1 : ； In the formula, α and β are characteristic weighting coefficients, with α ranging from (0, 1) and β = 1 - α; F fusion1 The feature fusion process is as follows: the feature Fain extracted by the convolutional neural network is multiplied by the weight coefficient β, the gradient feature is multiplied by the weight coefficient α, the two parts of the feature are then added together, and then passed through the ResNet-101 network to obtain a feature map with a final size of 14×14×512. ②Then use equation (1-6) to transform feature F fusion1 With differential excitation feature F de The fusion is performed to obtain the fused feature F. fusion2 : ； In the formula, γ and ξ are feature weighting coefficients, γ takes values ​​in the range (0, 1), and ξ = 1 - γ; the specific process is the feature F extracted in the previous step. fusion1 Multiply by the weight coefficient γ, the differential excitation feature F de Multiply by the weight coefficient ξ, and then add the two feature parts together to obtain a feature map with a final size of 14×14×512.

4. The image quality evaluation method simulating subjective human visual perception according to claim 1, characterized in that, The specific methods for the third step include: ① Using equation (1-7) to fuse the feature F fusion2 Dimensionality reduction is achieved through convolutional neural networks to reduce the computational cost of the attention mechanism: ； ② The Reshape operation is performed using Equation (1-8) to change the feature map into the form required for attention calculation. After the Reshape operation, the size of the X feature is (H×W) rows and C columns, where each column represents the information of all channels at a pixel position in the fused and compressed feature map F0. ； ③ An attention mechanism is used to improve the performance of image quality assessment. The X features are input into the self-attention layer to calculate the attention matrix, as shown in the following formula: ； ； ； In the formula, Q is the query matrix, K is the key matrix, and V is the value matrix. All three matrices are obtained through linear transformations of the X feature and are special mappings of the X feature to represent relevant information at specific locations in the feature map; the weight matrix W... Q W K W V The attention is dynamically adjusted, and then the overall attention H is calculated as shown in equation (1-12): ； The overall attention H is calculated as follows: the transpose of the key matrix K is multiplied by the query matrix Q to obtain the cross-relationship matrix between features at each location in the feature map; then, the cross-relationship matrix is ​​normalized by dividing it by the square root of dimension d. The attention weight matrix is ​​obtained, and then a softmax operation is performed on the obtained attention weight matrix to normalize the attention weights. The input vector is transformed into a probability distribution form to satisfy the properties of a probability distribution, ensuring that each weight... All values ​​are converted to values ​​between 0 and 1, and the sum of all weights is 1, ensuring that the model does not favor any particular part when performing information fusion, and obtains information evenly from different features; finally, the value matrix V is multiplied by the attention weight matrix to obtain the overall attention H; ④ The obtained overall attention H is reshaped using equation (1-13) to obtain the feature Y: 。 5. The image quality evaluation method simulating subjective human visual perception according to claim 1, characterized in that, The specific method for the fourth step is as follows: The two-dimensional feature C′ is calculated using equation (1-14): ； In the formula, Conv 1×1 (·) represents a convolution operation with a kernel size of 1, which compresses the obtained H×W×C features from C to 1 in the channel dimension; BN represents the BatchNorm layer; and ReLU is the activation function.

6. The image quality evaluation method simulating subjective human visual perception according to claim 1, characterized in that, The specific method for the fifth step is as follows: Using equation (1-15), the obtained two-dimensional feature C′ is flattened to obtain a one-dimensional feature vector N of 1×(H×W): ； In the formula, flatten(·) represents the flattening operation, which unfolds the feature C′ with size H×W×1 into the feature N with size 1×(H×W).

Images