A no-reference color image quality assessment method based on color and structural distortion

Abstract

The application is suitable for the field of image processing, and provides a no-reference color image quality evaluation method based on color and structure distortion, comprising the following steps: performing a blocking operation on a color image to form a color distortion image block data set, and forming a training set and a test set; a group of two-dimensional Scharr detection operators are constructed, and convolution processing is performed on the input image block by using the two-dimensional Scharr detection operators to obtain gradient feature information of the distortion image block; the obtained image block data set is converted into an HSV color space; a no-reference image quality evaluation convolutional neural network model based on color and structure distortion is constructed according to the existing image block data set; and the divided training set is input into the constructed no-reference image quality evaluation convolutional neural network model to obtain a trained no-reference image quality evaluation model. The application has reasonable design, the simulated quality score and the real quality score have strong correlation and small error, and is worth promoting.

Description

Technical Field

[0001] This invention belongs to the field of image processing, and particularly relates to a method for evaluating the quality of no-reference color images based on color and structural distortion. Background Technology

[0002] Currently, commonly used objective image quality assessment methods are divided into three categories: full-reference image quality assessment (FR-IQA), half-reference image quality assessment (RR-IQA), and no-reference image quality assessment (NR-IQA). The first two types of methods must use the original undistorted image as a reference for quality assessment. However, in many practical applications, the original reference image is difficult or impossible to obtain. Therefore, it is particularly important to establish a stable and universal no-reference image quality assessment (NR-IQA) method that conforms to the human visual system.

[0003] With the rapid development of information technology, high-resolution color images have become the mainstream for transmission. Most users enhance their photos before publishing or sharing them. However, inexperienced users often encounter unknown color and structural distortions during image processing. In recent years, many scholars have proposed numerous methods for no-reference quality assessment of color images. However, current image quality assessment methods mainly target images with distortions such as blurriness, JPEG compression, and noise, with relatively little research on quality assessment for color-distorted images. Existing objective natural image quality assessment methods are mostly for grayscale images, such as SSIM (Structure Similarity) and GMSD (Gradient Magnitude Similarity Deviation), while work on color image quality assessment is relatively limited.

[0004] In the process of color image quality assessment, two methods are commonly used. The first is to convert the color image to grayscale and then evaluate it using conventional methods. This is the most common method in practical image processing; however, these methods only consider structural distortion and do not fully account for the impact of color distortion on image quality. The second method uses color information in the RGB color space for quality evaluation. For example, the paper "A Feature-Enriched Completely Blind Image Quality Evaluator" separates and quantizes the RGB three channels of a color image to perceive color differences, achieving relatively good results, but there is still room for improvement.

[0005] These algorithms still have the following drawbacks: 1) Most algorithms are based on the error between pixels and do not consider the visual characteristics of the human eye. Therefore, the experimental results in many cases are inconsistent with human subjective perception, and these methods are only applicable to grayscale images but not to color images; 2) Most quality assessment algorithms for color images are studied on the RGB channels. However, the RGB color space does not conform to human visual perception of color, which will lead to deviations in image visual quality.

[0006] Currently, most image quality assessment methods do not specifically study color images, but mainly target common distortion types such as noise, compression, and blur. However, color images lose a lot of color information during transformation. When these methods assess the quality of color-distorted images, the results differ significantly from subjective quality scores.

[0007] Therefore, in view of the above situation, there is an urgent need to develop a no-reference color image quality assessment method based on color and structural distortion to overcome the shortcomings in current practical applications. Summary of the Invention

[0008] The purpose of this invention is to provide a no-reference color image quality assessment method based on color and structural distortion, aiming to solve the problems mentioned in the background above.

[0009] This invention is implemented as follows: a no-reference color image quality assessment method based on color and structural distortion includes the following steps:

[0010] Step 1: Perform a block operation on the color image, cropping it into a set of image blocks of the same size to form a color distortion image block dataset, which constitutes the training set and the test set;

[0011] Step 2: Construct a set of two-dimensional Scharr detection operators and use them to perform convolution processing on the input image patches to obtain the gradient feature information of the distorted image patches and generate gradient image patches;

[0012] Step 3: Convert the image patch dataset obtained in Step 1 to the HSV color space, and separate the H channel and S channel in the HSV color space;

[0013] Step 4: Based on the existing image patch dataset, construct a no-reference image quality assessment convolutional neural network model based on color and structural distortion;

[0014] Step 5: Input the training sets divided in Steps 1, 2 and 3 into the no-reference image quality assessment convolutional neural network model constructed in Step 4 to obtain the trained no-reference image quality assessment model; then, the no-reference image quality assessment model calculates the quality prediction values ​​of all sub-blocks segmented from the images in the test set.

[0015] A further technical solution involves dividing the color image into 32×32 image blocks in step 1 according to a non-overlapping partitioning method; randomly selecting 80% of the data from the color distortion image block dataset as the training set, and the remaining 20% ​​as the test set.

[0016] Further technical solutions, step 2 specifically includes the following steps:

[0017] Step 2.1: Input the distorted image block I(x);

[0018] Step 2.2: Establish a set of two-dimensional Scharr detection operators:

[0019]

[0020] Step 2.3: Using the two-dimensional Scharr detection operator established in Step 2.2, perform convolution calculation on the distorted image patch I(x) input in Step 2.1 to obtain the gradient information of the distorted image patch I(x) in two-dimensional space; the calculation formula for image patch I(x) through Scharr convolution is:

[0021]

[0022] Among them, G P (x) and G Q (x) represents the gradient of the image along the horizontal and vertical directions, respectively. The convolution operation is represented by the formula for calculating the global gradient of image patch I(x):

[0023]

[0024] Further technical solutions, step 3 specifically includes the following steps:

[0025] Step 3.1: Convert image patch I(x) from RGB color space to HSV color space, using the following formula:

[0026]

[0027]

[0028] V = max;

[0029] Where R, G, and B are the RGB color values ​​of the image point, respectively, and max and min are the maximum and minimum values ​​among the three values ​​of R, G, and B.

[0030] Further technical solutions, step 4 specifically includes the following steps:

[0031] Step 4.1: The network model consists of subnetwork M and subnetwork N; subnetwork M consists of three convolutional layers and three max pooling layers; subnetwork N consists of five convolutional layers and five max pooling layers.

[0032] Among them, the max pooling layer can retain more texture information. The calculation formula for the max pooling layer is:

[0033]

[0034] Among them, y kij This indicates the rectangular region R related to the k-th feature map. ij Max pooling output value, x kpq Represents a rectangular region R ij The element located at (p, q);

[0035] Step 4.2: The no-reference image quality assessment network model uses the L1 loss function. The formula for calculating the L1 loss function of the network model is as follows:

[0036]

[0037] Where N represents the number of images processed in a batch, p i Let q represent the prediction score of the i-th image patch. i This represents the true label score of the i-th image patch;

[0038] Step 4.3: The no-reference image quality assessment network model is optimized using the Adam optimizer, which adaptively adjusts parameters such as the learning rate to accelerate the model's convergence speed.

[0039] Further technical solutions, step 5 specifically includes the following steps:

[0040] Step 5.1: Use the original distorted image patch, gradient image patch, H-channel image patch and S-channel image patch from the training set obtained in Steps 1, 2 and 3 as input to the sub-network M;

[0041] Step 5.2: After the above four types of image patch data pass through the first convolutional layer and pooling layer of sub-network M, four feature vectors are obtained. These four feature vectors are then fused and passed through sub-network N to enhance the extraction of global features. The formula for fusing these four feature vectors is as follows:

[0042] X = concat(X) i X g X h X s );

[0043] Among them, X i X g Xh and X s These represent the feature vectors obtained after the first convolutional pooling layer of the original image patch, gradient image patch, H-channel image patch, and S-channel image patch, respectively. X represents the fused features, and concat is the feature fusion operation.

[0044] Step 5.3: Input the X obtained in Step 5.2 into the sub-network M. The extracted global features are then processed again through a fully connected neural network to reduce the dimensionality and obtain the feature vector Y. In this process, the fully connected layer uses matrix multiplication to transform the high-dimensional input feature data into low-dimensional sample labels, retaining the useful information and eliminating the spatial relationships between features. The calculation formula for the output of its l-th layer is shown below:

[0045] a l =σ(W l a l-1 +b l );

[0046] Among them, a l-1 W represents the output data of layer l-1. l b represents the weight parameters of the l-th layer. l This represents the offset of the l-th layer;

[0047] Step 5.4: The dimensionality-reduced feature vector Y is fused with the structurally distorted high-dimensional features and the color-distorted high-dimensional features. The fused features are then passed through a linear regression layer.

[0048] Step 5.5: In the linear regression layer, the network will go through multiple iterations. In each iteration, a small batch of image patch training samples is read and a set of predictions is obtained through the network model. After calculating the loss, the network starts backpropagation and stores the gradient of each parameter. At the same time, the network calls the optimization algorithm Adam to update the model parameters.

[0049] Step 5.6: After the above steps, the trained no-reference image quality assessment network model is obtained;

[0050] Step 5.7: Input the test set obtained in Steps 1, 2, and 3 into the no-reference image quality assessment model based on color and structural distortion. Then, the no-reference image quality assessment model calculates the predicted quality values ​​of all sub-blocks segmented from the image in the test set. The predicted quality score Q of the image is obtained by averaging all scores. The calculation formula is as follows:

[0051]

[0052] Where N represents the number of sub-blocks after segmentation of the test image, p i This represents the predicted quality value of the i-th image patch.

[0053] The present invention provides a no-reference color image quality assessment method based on color and structural distortion, which has the following beneficial effects:

[0054] 1) This method is an end-to-end optimization network model algorithm for color images. Compared with the divide-and-conquer strategy, the "end-to-end" network does not require human intervention throughout the entire learning process, overcoming the inaccuracies and incompleteness caused by the need for manual feature extraction in traditional machine learning algorithms. At the same time, the "end-to-end" learning approach has the advantage of synergistic efficiency and is more likely to obtain the global optimum.

[0055] 2) This method is a no-reference image quality assessment algorithm. By using a trained convolutional neural network framework, it is possible to assess the quality of distorted images without a reference image (original image).

[0056] 3) This method introduces the Scharr operator, which amplifies the weight coefficients in the filter to increase the difference between pixel values, and is also very effective for weak edge extraction in the image.

[0057] 4) This method fully considers the impact of unstructured distortions, such as color distortion, on image quality, and introduces the HSV color space consistent with the human visual system. In this space, the luminance component (V) is independent of the image's color information, while the hue (H) and saturation (S) closely approximate how the human eye perceives color. Therefore, this invention separates the hue (H) and saturation (S) components in the HSV color space of an image for quality evaluation, achieving results consistent with the human visual system.

[0058] 5) This method fully considers the impact of high- and low-level features on image quality evaluation. It utilizes a well-designed network model to extract low-dimensional and high-dimensional features related to structural and color distortion, and then fuses these features before inputting them into a fully connected neural network for quality regression prediction. The results show a strong correlation and small error between the simulated quality score and the actual quality score. Attached Figure Description

[0059] Figure 1 A flowchart of a no-reference color image quality assessment method based on color and structural distortion provided in an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the network structure of the no-reference color image quality assessment method based on color and structural distortion provided in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0063] like Figure 1-2 As shown, an embodiment of the present invention provides a method for evaluating the quality of a no-reference color image based on color and structural distortion, comprising the following steps:

[0064] Step 1: Perform a block operation on the color image, cropping it into a set of image blocks of the same size to form a color distortion image block dataset, which constitutes the training set and the test set;

[0065] Step 2: Construct a set of two-dimensional Scharr (edge ​​extraction) detection operators and use them to perform convolution processing on the input image patches to obtain the gradient feature information of the distorted image patches and generate gradient image patches;

[0066] Step 3: Convert the (RGB) image patch dataset obtained in Step 1 to the HSV (Hue, Saturation, Value) color space, and separate the H (hue) channel and S (saturation) channel in the HSV color space.

[0067] Step 4: Based on the existing image patch dataset, construct a no-reference image quality assessment convolutional neural network model based on color and structural distortion;

[0068] Step 5: Input the training sets divided in Steps 1, 2 and 3 into the no-reference image quality assessment convolutional neural network model constructed in Step 4 to obtain the trained no-reference image quality assessment model; then, the no-reference image quality assessment model calculates the quality prediction values ​​of all sub-blocks segmented from the images in the test set.

[0069] In a preferred embodiment of the present invention, in step 1, the color image is divided into 32×32 image blocks according to a non-overlapping division method; 80% of the data in the color distortion image block dataset is randomly selected as the training set, and the remaining 20% ​​of the data is used as the test set.

[0070] In a preferred embodiment of the present invention, in step 2, since gradient information is more sensitive to image blurring and compression, using multi-feature fusion will improve the performance of IQA (image-based interactive question answering). Normally, the gradient image represents the details of the input. However, given a distorted image, the distortion of high-frequency components often contains the most important features for identifying image degradation. Therefore, in practice, using both the image and gradient as input is more likely to describe image quality.

[0071] Step 2 specifically includes the following steps:

[0072] Step 2.1: Input the distorted image block I(x);

[0073] Step 2.2: Establish a set of two-dimensional Scharr detection operators:

[0074]

[0075] Step 2.3: Using the two-dimensional Scharr detection operator established in Step 2.2, perform convolution calculation on the distorted image patch I(x) input in Step 2.1 to obtain the gradient information of the distorted image patch I(x) in two-dimensional space; the calculation formula for image patch I(x) through Scharr convolution is:

[0076]

[0077] Among them, G P (x) and G Q (x) represents the gradient of the image along the horizontal and vertical directions, respectively. The convolution operation is represented by the formula for calculating the global gradient of image patch I(x):

[0078]

[0079] In a preferred embodiment of the present invention, in step 3, the color and intensity of pixels are the original and primary information reflecting image quality. Most methods have achieved good performance by using them. Therefore, it is a natural idea to use color images to solve the IQA problem. The present invention extracts the H and S channels from the HSV channels, which are most relevant to the human visual system, to obtain more color features, and extracts various features from the H and S channel images through a convolutional neural network.

[0080] Step 3 specifically includes the following steps:

[0081] Step 3.1: Convert image patch I(x) from RGB color space to HSV color space, using the following formula:

[0082]

[0083]

[0084] V = max;

[0085] Where R, G, and B are the RGB color values ​​of the image point, respectively, and max and min are the maximum and minimum values ​​among the three values ​​of R, G, and B.

[0086] In a preferred embodiment of the present invention, in step 4, a network model based on HSV convolutional neural network is designed, which consists of multiple single-channel CNN (convolutional neural network) models. The training result of each CNN can be regarded as a deep feature extractor. The multiple CNN models can extract quality feature representations of multi-color component images, more comprehensively describing image quality distortion features, especially image color information. The features extracted from multiple network models are synergistically fused, and the performance of the image quality predictor is further improved through the idea of ​​ensemble learning.

[0087] Step 4 specifically includes the following steps:

[0088] Step 4.1: The network model consists of subnetwork M and subnetwork N; subnetwork M consists of three convolutional layers and three max pooling layers; subnetwork N consists of five convolutional layers and five max pooling layers.

[0089] Among them, the max pooling layer can retain more texture information. The calculation formula for the max pooling layer is:

[0090]

[0091] Among them, y kij This indicates the rectangular region R related to the k-th feature map. ij Max pooling output value, x kpq Represents a rectangular region R ij The element located at (p, q).

[0092] Step 4.2: The no-reference image quality assessment network model uses the L1 loss function. The formula for calculating the L1 loss function of the network model is as follows:

[0093]

[0094] Where N represents the number of images processed in a batch, p i Let q represent the prediction score of the i-th image patch. i This represents the true label score of the i-th image patch.

[0095] Step 4.3: The no-reference image quality assessment network model is optimized using the Adam optimizer, which adaptively adjusts parameters such as the learning rate to accelerate the model's convergence speed.

[0096] In a preferred embodiment of the present invention, step 5 specifically includes the following steps:

[0097] Step 5.1: Use the original distorted image patch, gradient image patch, H-channel image patch and S-channel image patch from the training set obtained in Steps 1, 2 and 3 as input to the sub-network M.

[0098] Step 5.2: After the above four types of image patch data pass through the first convolutional layer and pooling layer of sub-network M, four feature vectors are obtained. These four feature vectors are then fused and passed through sub-network N to enhance the extraction of global features. The formula for fusing these four feature vectors is as follows:

[0099] X = concat(X) i X g X h X s );

[0100] Among them, X i X g X h and X s These represent the feature vectors obtained after the first convolutional pooling layer for the original image block, gradient image block, H-channel image block, and S-channel image block, respectively. X represents the fused features, and concat is the feature fusion operation.

[0101] Step 5.3: Input the X obtained in Step 5.2 into the sub-network M. The extracted global features are then processed again through a fully connected neural network to reduce the dimensionality and obtain the feature vector Y. In this process, the fully connected layer uses matrix multiplication to transform the high-dimensional input feature data into low-dimensional sample labels, retaining the useful information and eliminating the spatial relationships between features. The calculation formula for the output of its l-th layer is shown below:

[0102] a l =σ(W l a l-1 +b l );

[0103] Among them, a l-1 W represents the output data of layer l-1. l b represents the weight parameters of the l-th layer. l This represents the offset of the l-th layer.

[0104] Step 5.4: The dimensionality-reduced feature vector Y is fused with the structurally distorted high-dimensional features and the color-distorted high-dimensional features. The fused features are then passed through a linear regression layer.

[0105] Step 5.5: In the linear regression layer, the network will go through multiple iterations. In each iteration, a small batch of image patch training samples is read and a set of predictions is obtained through the network model. After calculating the loss, the network starts backpropagation and stores the gradient of each parameter. At the same time, the network calls the optimization algorithm Adam to update the model parameters.

[0106] Step 5.6: After the above steps, the trained no-reference image quality assessment network model is obtained.

[0107] Step 5.7: Input the test set obtained in Steps 1, 2, and 3 into the no-reference image quality assessment model based on color and structural distortion. Then, the no-reference image quality assessment model calculates the predicted quality values ​​of all sub-blocks segmented from the image in the test set. The predicted quality score Q of the image is obtained by averaging all scores. The calculation formula is as follows:

[0108]

[0109] Where N represents the number of sub-blocks after segmentation of the test image, p i This represents the predicted quality value of the i-th image patch.

[0110] in addition, Figure 1 This paper demonstrates the workflow for color image quality assessment using multi-depth CNNs. First, the color image is divided into a set of image patch datasets. Second, a general CNN architecture is adopted and improved to learn useful feature representations from the image patch dataset. Each image patch serves as the input to a single CNN, thus forming a multi-CNN model. Third, hue (H) and saturation (S) of the HSV color model are extracted. Then, multiple output feature vectors are fused into depth features. Finally, a nonlinear regression model is constructed to map the extracted depth features to visual quality scores.

[0111] The above embodiments of the present invention provide a no-reference color image quality assessment method based on color and structural distortion, which has the following beneficial effects:

[0112] 1) This method is an end-to-end optimization network model algorithm for color images. Compared with the divide-and-conquer strategy, the "end-to-end" network does not require human intervention throughout the entire learning process, overcoming the inaccuracies and incompleteness caused by the need for manual feature extraction in traditional machine learning algorithms. At the same time, the "end-to-end" learning approach has the advantage of synergistic efficiency and is more likely to obtain the global optimum.

[0113] 2) This method is a no-reference image quality assessment algorithm. By using a trained convolutional neural network framework, it is possible to assess the quality of distorted images without a reference image (original image).

[0114] 3) Regarding structural distortion, in order to effectively extract weak edges, it is necessary to increase the difference between pixel values. Therefore, the Scharr operator is introduced. The Scharr operator amplifies the weight coefficients in the filter to increase the difference between pixel values, which is also very effective for extracting weak edges in the image.

[0115] 4) While most no-reference quality assessment methods only consider structural distortion, this method fully considers the impact of non-structural distortion, such as color distortion, on image quality, and introduces the HSV color space consistent with the human visual system. In this space, the luminance component (V) is independent of the image's color information, while hue (H) and saturation (S) closely approximate how the human eye perceives color. Therefore, this invention separates the hue (H) and saturation (S) components in the HSV color space of an image for quality assessment, achieving results consistent with the human visual system.

[0116] 5) This method fully considers the impact of high- and low-level features on image quality evaluation. It utilizes a well-designed network model to extract low-dimensional and high-dimensional features related to structural and color distortion, and then fuses these features before inputting them into a fully connected neural network for quality regression prediction. The results show a strong correlation and small error between the simulated quality score and the actual quality score.

[0117] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0118] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for evaluating the quality of a no-reference color image based on color and structural distortion, characterized in that, Includes the following steps: Step 1: Perform a block operation on the color image, cropping it into a set of image blocks of the same size to form a color distortion image block dataset, which constitutes the training set and the test set; Step 2: Construct a set of two-dimensional Scharr detection operators and use them to perform convolution processing on the input image patches to obtain the gradient feature information of the distorted image patches and generate gradient image patches; Step 3: Convert the image patch dataset obtained in Step 1 to the HSV color space, and separate the H channel and S channel in the HSV color space; Step 4: Based on the existing image patch dataset, construct a no-reference image quality assessment convolutional neural network model based on color and structural distortion; Step 5: Input the training sets divided in Steps 1, 2 and 3 into the no-reference image quality assessment convolutional neural network model constructed in Step 4 to obtain the trained no-reference image quality assessment model; then, the no-reference image quality assessment model calculates the quality prediction values ​​of all sub-blocks segmented from the images in the test set. Step 4 specifically includes the following steps: Step 4.1: The network model consists of subnetwork M and subnetwork N; subnetwork M consists of three convolutional layers and three max pooling layers; subnetwork N consists of five convolutional layers and five max pooling layers. Among them, the max pooling layer can retain more texture information. The calculation formula for the max pooling layer is: ； wherein, represents a maximum pooling output value related to the k-th feature map in a rectangular region , represents an element located at (p, q) in a rectangular region ; Step 4.2: The no-reference image quality assessment network model uses the L1 loss function. The formula for calculating the L1 loss function of the network model is as follows: ； wherein, denotes the number of images at batch time, denotes the prediction score of the -th image block, denotes the true label score of the -th image block; Step 4.3: The no-reference image quality assessment network model is optimized using the Adam optimizer, which adaptively adjusts the learning rate parameter to accelerate the model convergence speed; Step 5 specifically includes the following steps: Step 5.1: Use the original distorted image patch, gradient image patch, H-channel image patch and S-channel image patch from the training set obtained in Steps 1, 2 and 3 as input to subnetwork N; Step 5.2: After the four types of image patch data from Step 5.1 pass through the first convolutional and pooling layers of sub-network N, four feature vectors are obtained. These four feature vectors are then fused and passed through sub-network M to enhance global feature extraction. The formula for fusing these four feature vectors is as follows: ； in, , , and These represent the feature vectors obtained after the first convolutional pooling layer for the original image patch, gradient image patch, H-channel image patch, and S-channel image patch, respectively. Indicates the characteristics after fusion. It is a feature fusion operation; Step 5.3: Take the product obtained in Step 5.2 The global features extracted from the input subnetwork M are then passed through a fully connected neural network for dimensionality reduction to obtain the feature vector. In this process, the fully connected layer uses matrix multiplication to transform the high-dimensional input feature data into low-dimensional sample labels, retaining useful information and eliminating spatial relationships between features. The calculation formula for the output of its l-th layer is shown below: ； in, Indicates the first The output data of the layer, Indicates the first Layer weight parameters, Indicates the first Layer offset; Step 5.4: Feature vector after dimension reduction The structural distortion high-dimensional feature and the color distortion high-dimensional feature are fused, and the fused feature passes through a linear regression layer. Step 5.5: In the linear regression layer, the network will go through multiple iterations. In each iteration, a small batch of image patch training samples is read and a set of predictions is obtained through the network model. After calculating the loss, the network starts backpropagation and stores the gradient of each parameter. At the same time, the network calls the optimization algorithm Adam to update the model parameters. Step 5.6: After the above steps, the trained no-reference image quality assessment network model is obtained; Step 5.7: Input the test set obtained in Steps 1, 2, and 3 into the no-reference image quality assessment model based on color and structural distortion. Then, the no-reference image quality assessment model calculates the predicted quality values ​​of all sub-blocks segmented from the image in the test set. The predicted quality scores of the image are obtained by averaging all the predicted quality values. The calculation formula is as follows: ； Where N represents the number of sub-blocks after segmentation of the test image. Indicates the first Quality prediction values ​​for each image patch.

2. The color and structural distortion based no-reference color image quality assessment method according to claim 1, characterized in that, In step 1, the color image is divided into 32×32 image blocks according to a non-overlapping division method; 80% of the data in the color distortion image patch dataset was randomly selected as the training set, and the remaining 20% ​​was used as the test set.

3. The color and structural distortion based no-reference color image quality assessment method according to claim 2, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Inputting the distorted image block ; Step 2.2: Establish a set of two-dimensional Scharr detection operators: ； Step 2.3: Use the two-dimensional Scharr detection operator established in Step 2.2 to process the distorted image patch input in Step 2.

1. Perform convolution calculations to obtain distorted image patches. Gradient information in two-dimensional space; for image patches The formula for calculating it using Scharr convolution is: ； in, and These represent the gradients of the image along the horizontal and vertical directions, respectively. Represents convolution operation; image patch The formula for calculating the overall gradient is: 。 4. The no-reference color image quality assessment method based on color and structural distortion according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Convert the image block Convert from RGB color space to HSV color space with the following formula: ； ； Wherein, R, G, B are respectively the score of RGB color quantity of the image point; max and min are the maximum and minimum of the three values of R, G, B.

Images