Image quality evaluation method, system and readable storage medium

Abstract

The present application discloses an image quality evaluation method, system and readable storage medium, the method includes constructing a training image data set, the training image data set includes real distorted images; constructing an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculation; performing preprocessing operations on each image included in the training image data set according to a preset preprocessing method to obtain corresponding preprocessed training data; inputting the preprocessed training data into the initial network model for training to obtain a trained target network model; performing the preprocessing operation on the real distorted image to be evaluated, and inputting the obtained preprocessed data to be evaluated into the target network model for processing to obtain a corresponding image quality score. The implementation of this method can improve the accuracy of image quality scoring.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to an image quality evaluation method, system and readable storage medium. Background Art

[0002] Generally speaking, image quality evaluation can be simply divided into subjective quality evaluation and objective quality evaluation. The former is not popular because it requires a lot of manpower and material resources as support, while objective quality evaluation is divided into three types: full reference, semi-reference, and no-reference. Full reference and semi-reference require all and part of the reference image information respectively, while no-reference methods do not require reference image information. In addition, as mentioned above, for different computer vision applications and the social and streaming media industries, it is crucial to be able to accurately predict the perceived image quality in the absence of reference images. Therefore, no-reference image quality evaluation methods have attracted much attention and are widely used.

[0003] Most of the existing no-reference methods are based on deep learning. However, the use of convolutional neural networks to extract quality-aware features of images cannot capture long-distance features, and it is necessary to increase the number of network layers to obtain a larger receptive field. Recently, Transformer has performed very well in the field of computer vision. Among them, Swin Transformer integrates the advantages of CNN and Transformer, showing great application prospects. Swin Transformer can use the local attention mechanism to process large-size images, and also has the advantage of the sliding window scheme to establish long-distance dependencies between features. However, there is still room for improvement in the existing Transformer-based methods for image quality evaluation. For example, the features of all channels are treated equally and are not adjusted according to their importance, resulting in low evaluation accuracy. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an image quality evaluation method, system and readable storage medium in view of the technical problem that the evaluation accuracy of the existing technical solutions is not high.

[0005] The technical solution adopted by the present invention to solve the technical problem is: constructing an image quality evaluation method, the method comprising:

[0006] Constructing a training image dataset, wherein the training image dataset includes real distorted images;

[0007] Construct an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculations;

[0008] According to a preset preprocessing method, preprocessing operations are performed on each image included in the training image data set to obtain corresponding preprocessed training data;

[0009] Inputting the preprocessed training data into the initial network model for training to obtain a trained target network model;

[0010] The preprocessing operation is performed on the real distorted image to be evaluated, and the obtained preprocessed data to be evaluated is input into the target network model for processing to obtain a corresponding image quality score.

[0011] The present application discloses an image quality evaluation system, which includes a data set construction module, a model construction module, a data preprocessing module and an image quality scoring module, wherein:

[0012] The data set construction module is used to construct a training image data set, wherein the training image data set includes real distorted images;

[0013] The model building module is used to build an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculation;

[0014] The data preprocessing module is used to perform preprocessing operations on each image included in the training image data set according to a preset preprocessing method to obtain corresponding preprocessed training data;

[0015] The model building module is further used to input the pre-processed training data into the initial network model for training to obtain a trained target network model;

[0016] The image quality scoring module is used to perform the preprocessing operation on the real distorted image to be evaluated, and input the obtained preprocessed data to be evaluated into the target network model for processing to obtain a corresponding image quality score.

[0017] In a third aspect, an embodiment of the present application further provides a readable storage medium, wherein the readable storage medium includes an image quality assessment method program. When the image quality assessment method program is executed by a processor, the steps of an image quality assessment method as described in any one of the above items are implemented.

[0018] The present invention implements an image quality assessment method, system and readable storage medium, and uses deep learning technology with powerful fitting capabilities to further learn quality perception features for high-level features and low-level features extracted by Swin Transformer at different stages. In addition, the channel attention and spatial attention mechanisms are used to allocate and learn weights, and image quality assessment is performed based on the quality perception features learned by the dual attention mechanism. At the same time, an adaptive weighting strategy is used to allocate weights to the scores predicted by the dual attention branches, and quality perception features are integrated, so that the final image quality can be better predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0020] Figure 1 is a flow chart of an image quality evaluation method in one embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the Swin Transformer Block structure that appears with an even number of connections;

[0022] Figure 3 It is a schematic diagram of the network structure of the channel attention branch network;

[0023] Figure 4 It is a schematic diagram of the network structure of the spatial attention branch network;

[0024] Figure 5 It is a system structure diagram of an image quality evaluation system in one embodiment of the present invention. DETAILED DESCRIPTION

[0025] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0026] Please refer to Figure 1 , Figure 1 This is a flowchart of an image quality evaluation method in some embodiments of the present application. Taking the method applied to a computer device (the computer device may be a terminal or a server, and the terminal may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server may be an independent server or a server cluster composed of multiple servers) as an example, the method includes the following steps:

[0027] Step S100: construct a training image dataset, where the training image dataset contains real distorted images.

[0028] Specifically, the computer device will choose to obtain a large-scale natural distortion dataset KonIQ-10k containing real distortions.

[0029] In one embodiment, the large-scale natural distortion dataset contains 10,073 images taken by 1,265 different models of cameras from approximately 100 manufacturers.

[0030] In one of the embodiments, the large-scale natural distortion dataset contains two sizes of images, namely 1024×768 and 512×384, because large-size images can better reflect the quality of original images.

[0031] In the current embodiment, 1024×768 images can be selected as the experimental data set. The training set and the test set can be divided according to a 28 split, that is, 80% of the images in the data set are used for model training, and the remaining 20% ​​of the images are used to test the accuracy of model training.

[0032] Step S200, constructing an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculation.

[0033] Specifically, the training image dataset constructed above will be further input into the initial network model, which will extract image channels and spatial feature information from the input data, and perform image quality assessment based on the image channels and spatial feature information. Finally, the two evaluation results are combined for weighted summation to determine the final image quality assessment prediction score.

[0034] Step S300: performing a preprocessing operation on each image included in the training image data set according to a preset preprocessing method to obtain corresponding preprocessed training data.

[0035] Specifically, the computer device performs preprocessing operations on each image included in the training image data set, including: adjusting the resolution of each image included in the training image data set, specifically scaling it to a resolution of 512×384; then, performing random shearing to control the size of the sheared image to 384×384, so as to meet the input specifications of Swin Transformer and retain more information of the original image that is useful for quality evaluation. The number of random shearing for the same image is set to 5, so that the final score of an image predicted by the model is obtained by averaging the predicted scores of 5 images while meeting the large data volume requirements of Swin Transformer.

[0036] In one embodiment, the final quality assessment score of an image predicted by the model is expressed as:

[0037]

[0038] Where n represents the number of random cuts of the image, Pred j Represents the evaluation score of the slice of the j-th image predicted by the trained model.

[0039] Step S400: input the preprocessed training data into the initial network model for training to obtain a trained target network model.

[0040] Specifically, the mean square error is used as the loss function during network training, where the mean square error is defined as follows:

[0041]

[0042] Among them, B represents the size of a batch during training, x i Represents the prediction score measured by the model, y i Represents the proxy score of the image slice, which is replaced by the label score of the original image.

[0043] Step S500, performing the preprocessing operation on the real distorted image to be evaluated, and inputting the obtained preprocessed data to be evaluated into the target network model for processing to obtain a corresponding image quality score.

[0044] Specifically, when the quality of the real distorted image to be evaluated is evaluated based on the target network model trained in step S400, it is also necessary to follow the operation steps of step S400 and perform the same processing, that is, first scale the image to a size of 512×384, then randomly crop 5 image slices of a size of 384×384, and use the slices as the input of the target network model; wherein the final evaluation score of the image is obtained by averaging the prediction scores of the 5 representing an image, and the specific calculation method is shown in formula (1), which will not be repeated here.

[0045] It should be noted that during the test, the computer equipment uses two indicators, namely the Pearson Linear Correlation Coefficient (PLCC) and the Spearman Rank Correlation Coefficient (SROCC), to measure the accuracy of the assessment. Among them:

[0046] (1) The Pearson linear correlation coefficient (PLCC) is mainly used to measure the prediction accuracy of the objective quality evaluation method. Its expression is:

[0047]

[0048] Where N is the number of images, s i(i=1,…,N) is the subjective score, is the mean of the subjective scores, p i (i=1,…,N) is the predicted objective score, is the mean of the objective scores.

[0049] (2) The Spearman rank correlation coefficient (SROCC) is mainly used to measure the monotonic consistency of the objective quality prediction score. Its value is only related to the ranking of the image in the prediction result, ignoring the relative distance between the predicted value and the true value. Its expression is:

[0050]

[0051] Among them, d i It is the difference between the ranking numbers of image i in the subjective score and the objective score, and N is the number of images.

[0052] In combination with formula (3)-formula (4), it should be noted that the closer the values ​​of PLCC and SROCC are to 1, the better the effect of the method proposed in this application.

[0053] As can be seen from the above, an image quality evaluation method disclosed in the present application utilizes deep learning technology with powerful fitting capabilities to further learn quality-aware features for high-level features and low-level features extracted by Swin Transformer at different stages. In addition, the channel attention and spatial attention mechanisms are used to allocate and learn weights, and image quality evaluation is performed based on the quality-aware features learned by the dual attention mechanism. At the same time, an adaptive weighting strategy is used to allocate weights to the scores predicted by the dual attention branches, and quality-aware features are integrated, so as to better predict the final image quality.

[0054] In one embodiment, the initial network model includes a feature extraction network, and a dual attention prediction and adaptive weighting network, wherein:

[0055] The feature extraction network is used to perform multi-scale feature extraction on the image encoded into a feature vector after slicing to obtain a corresponding feature map, wherein the multi-scale feature extraction includes shallow feature extraction, middle feature extraction and deep feature extraction.

[0056] The dual attention prediction and adaptive weighted network is used to perform weighted calculation of image quality evaluation based on channel feature information and spatial feature information extracted from the obtained feature map.

[0057] The above embodiment obtains features of four different stages by extracting features from the input image to be evaluated, and realizes modeling of low-level feature and high-level feature space and channel information through two branches. It combines multi-scale feature fusion, dual attention mechanism and adaptive weighting strategy to make an evaluation of image quality that is consistent with the subjective perception of the human eye, thereby improving the evaluation accuracy.

[0058] In one embodiment, performing multi-scale feature extraction by the feature extraction network includes the following steps:

[0059] In the shallow feature extraction stage, i.e. the first stage, a Linear Embedding module for feature mapping through linear operations and two consecutive Swin Transformer Block modules are combined for shallow feature extraction. The vector input to the Swin Transformer Block module with an odd sequence number will first be normalized within the layer, and the standard multi-head self-attention processing will be performed within the standard window where the vector is located. The processing result will then be added to the input and then normalized at the layer. The Swin Transformer Block module with an even sequence number uses a multi-head self-attention processing mechanism based on a sliding window for feature extraction.

[0060] Specifically, the Linear Embedding module performs feature mapping for linear operations, the convolution kernel size is H / 4*W / 4, and the total number of feature maps is 96. Among them, the Swin Transformer Block structure that appears in an even number of connections is as follows Figure 2 shown.

[0061] In one embodiment, (1) when performing multi-head self-attention processing (W-MSA), the calculation formula includes:

[0062] MultiHead(Q,K,V)=Concat(head1,…,head h )W1; (5)

[0063] head i =Attention(Q i ,K i ,V i ); (6)

[0064] Among them, head i represents the i-th head, W1 represents the weight matrix, Q is the query matrix corresponding to the input vector, K is the key matrix, V is the value matrix, d kis the dimension of the input vector; the calculation formula of the softmax activation function includes: S j is the softmax value of the jth component in the input vector, e j is the value of the jth component.

[0065] (2) The calculation of two consecutive Swin Transformer Blocks is as follows:

[0066]

[0067]

[0068]

[0069]

[0070] in, and They respectively represent the output features of the W-MSA layer and the MLP layer of the Swin Transformer Block module with odd numbers. W-MSA represents window-based multi-head attention, and SW-MSA represents moving window multi-head attention.

[0071] In the middle-level feature extraction stage, i.e., the second stage, a Patch Merging module for reducing the size of feature vectors to obtain hierarchical features and two consecutive Swin Transformer Block modules are used to extract middle-level features.

[0072] Specifically, after the first and second stage Patch Merging modules, the feature map size is Wherein, H and W are the height and width of the image respectively, and C is the number of channels passing through the first stage. In one embodiment, the value of C can be 128, and the same applies below.

[0073] The deep feature extraction stage is divided into two sub-stages. In the first deep feature extraction stage, i.e., the third stage, a Patch Merging module and 18 consecutive Swin Transformer Block modules are combined to perform the deep feature extraction of this stage. In the second deep feature extraction stage, i.e., the fourth stage, a Patch Merging module and 2 consecutive Swin Transformer Block modules are combined to perform the deep feature extraction of this stage.

[0074] Specifically, after the second and third stage Patch Merging modules, the feature map size is Then, after passing through the Patch Merging modules in the third and fourth stages, the feature map size is Finally, the feature map size obtained after the fourth stage is

[0075] The above embodiment comprehensively utilizes the Swin Transformer network with strong fitting ability to perform feature extraction, obtains features of four different stages, and makes up for the defect that the quality perception feature does not fit the prediction score well enough while considering the feature importance of different channels of the deep network.

[0076] In one embodiment, the dual attention prediction and adaptive weighting network includes a channel attention branch network and a spatial attention branch network, wherein:

[0077] The channel attention branch network includes a channel attention module for extracting and processing channel feature information of a feature map, and a fully connected layer for performing image quality assessment in combination with the obtained channel feature information.

[0078] Specifically, the network structure of the channel attention branch network can be referred to Figure 3 .

[0079] The spatial attention branch network includes a spatial attention module for extracting and processing spatial feature information of a feature map, and a fully connected layer for combining the obtained spatial feature information to perform image quality assessment.

[0080] Specifically, the network structure of the spatial attention branch network can be referred to Figure 4 .

[0081] In the above-mentioned embodiment, channel attention and spatial attention mechanisms are used to allocate and learn weights respectively, and image quality is evaluated based on the quality perception features learned by the dual attention mechanism. At the same time, an adaptive weighting strategy is used to allocate weights to the scores predicted by the dual attention branches, and quality perception features are integrated to better predict the final image quality.

[0082] In one of the embodiments, image quality assessment is performed based on channel feature information extracted from the obtained feature map, including: performing a convolution on each of the four multi-scale feature maps extracted by the feature extraction network to obtain a first feature map with a uniform number of corresponding channels; inputting each of the first feature maps into the channel attention branch network, and the channel attention branch network extracts a channel attention feature map from the first feature map; and performing image quality assessment based on the extracted channel attention feature map.

[0083] Specifically, the feature extraction network and the channel attention branch of the dual attention prediction and adaptive weighted network are connected by convolving the multi-scale four-stage feature maps extracted by the feature extraction network so that their number of channels is unified to 256, and then the features of these four stages are input into the channel attention-based branch for image quality assessment.

[0084] In one embodiment, the channel attention feature map extracted by the channel attention branch network is represented as CA:

[0085] CA=s{FC[AvgPool(P i )]+FC1[MaxPool(P i )]},i=1,2,3,4; (11)

[0086] Among them, P i represents the input feature vector, i represents the four stages of the feature extraction network based on Swin Transformer, AvgPool(P i ) represents the average pooling operation on the input feature vector, MaxPool(P i ) represents the maximum pooling operation on the input feature vector, FC and FC1 represent the fully connected layer, and s{*} represents the sigmoid nonlinear activation function.

[0087] The first image quality assessment result predicted by the channel attention branch network is denoted as Pred1:

[0088]

[0089] Among them, FC2 represents the fully connected layer, Represents the dot product operation.

[0090] In one embodiment, image quality evaluation is performed based on spatial feature information extracted from the obtained feature map, including: adjusting the sizes of the feature maps of the last three stages extracted by the feature extraction network respectively to obtain a second feature map whose image size is consistent with the feature map of the first stage; splicing together each of the second feature maps and the feature map of the first stage, and inputting the obtained new feature map into the spatial attention branch network, the spatial attention branch network extracts the spatial attention feature map from the new feature map, and image quality evaluation is performed based on the extracted spatial attention feature map.

[0091] Specifically, the feature extraction network and the spatial attention branch of the dual attention prediction and adaptive weighting network are connected in such a way that the last three of the four-stage feature maps of the multi-scale extracted by the feature extraction network are upsampled to the same size as the feature map of the first stage and then spliced ​​together to obtain a new feature map. Afterwards, the new feature map is input into the branch based on spatial attention for image quality evaluation.

[0092] In one embodiment, the obtained new feature map is represented as:

[0093] F new =Concat[F1,Unsample(F2),Unsample(F3),Unsample(F4)]; (13)

[0094] Among them, Concat represents the concatenation operation between feature vectors, Unsample represents the upsampling operation, and F1, F2, F3, and F4 represent the outputs of the four stages of Swin Transformer respectively.

[0095] In one embodiment, the spatial attention feature map extracted by the spatial attention branch network is represented as SA:

[0096] SA=s{Conv(1,(7,7),1)[AvgPool(P),MaxPool(P)]}; (14)

[0097] Among them, Conv(1,(7,7),1) indicates that the number of output channels is 1, and a convolution operation with a convolution kernel of 7×7 and a stride of 1 is used. The second image quality assessment result predicted by the spatial attention branch network is expressed as Pred2:

[0098]

[0099] The weighted calculation of image quality evaluation based on the channel feature information and spatial feature information extracted from the obtained feature map includes: combining the obtained first image quality evaluation result Pred1 and the second image quality evaluation result Pred2, and performing weighted calculation of image quality evaluation by the following formula:

[0100] Pred=w1*Pred1+w2*Pred2; (16)

[0101] Wherein, w1 and w2 represent weighting coefficients, which can be determined according to adaptive parameters that can be learned during the training process, and Pred represents the weighted calculation result of the obtained image quality evaluation.

[0102] Please refer to Figure 5The present application discloses an image quality evaluation system 500, which includes a data set construction module 501, a model construction module 502, a data preprocessing module 503 and an image quality scoring module 504, wherein:

[0103] The data set construction module 501 is used to construct a training image data set, and the training image data set includes real distorted images.

[0104] The model building module 502 is used to build an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculation.

[0105] The data preprocessing module 503 is used to perform preprocessing operations on each image included in the training image data set according to a preset preprocessing method to obtain corresponding preprocessed training data.

[0106] The model building module 502 is also used to input the pre-processed training data into the initial network model for training to obtain a trained target network model.

[0107] The image quality scoring module 504 is used to perform the preprocessing operation on the real distorted image to be evaluated, and input the obtained preprocessed data to be evaluated into the target network model for processing to obtain a corresponding image quality score.

[0108] In one embodiment, each module in the system 500 is also used to implement the method in any optional implementation of the above embodiment.

[0109] As can be seen from the above, the image quality evaluation system disclosed in the present application uses deep learning technology with strong fitting ability to further learn quality perception features for high-level features and low-level features extracted by Swin Transformer at different stages. In addition, the channel attention and spatial attention mechanisms are used to allocate and learn weights respectively, and image quality evaluation is performed based on the quality perception features learned by the dual attention mechanism. At the same time, an adaptive weighting strategy is used to allocate weights to the scores predicted by the dual attention branches, and quality perception features are integrated, so as to better predict the final image quality.

[0110] The embodiment of the present application provides a readable storage medium, and when the computer program is executed by the processor, the method in any optional implementation of the above embodiment is executed. Wherein, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, disk or optical disk.

[0111] The above-mentioned readable storage medium uses deep learning technology with powerful fitting capabilities to further learn quality-aware features for high-level features and low-level features extracted by SwinTransformer at different stages. In addition, the channel attention and spatial attention mechanisms are used to allocate and learn weights, and image quality evaluation is performed based on the quality-aware features learned by the dual attention mechanism. At the same time, an adaptive weighting strategy is used to allocate weights to the scores predicted by the dual attention branches, and quality-aware features are integrated to better predict the final image quality.

[0112] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for evaluating image quality, characterized in that: The method comprises: Constructing a training image dataset, wherein the training image dataset includes real distorted images; Construct an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculations; According to a preset preprocessing method, preprocessing operations are performed on each image included in the training image data set to obtain corresponding preprocessed training data; Inputting the preprocessed training data into the initial network model for training to obtain a trained target network model; Performing the preprocessing operation on the real distorted image to be evaluated, and inputting the obtained preprocessed data to be evaluated into the target network model for processing to obtain a corresponding image quality score; The initial network model includes a feature extraction network, a dual attention prediction and an adaptive weighting network, wherein: The feature extraction network is used to perform multi-scale feature extraction on the image encoded into the feature vector after slicing to obtain the corresponding feature map, wherein the multi-scale feature extraction includes shallow feature extraction, middle feature extraction and deep feature extraction; Performing multi-scale feature extraction through the feature extraction network includes the following steps: In the shallow feature extraction stage, i.e. the first stage, a LinearEmbedding module for feature mapping through linear operations and two consecutive Swin Transformer Block modules are combined for shallow feature extraction. The vectors input to the Swin Transformer Block modules with odd numbers will first be normalized within the layer, and the standard multi-head self-attention processing will be performed within the standard window where the vector is located. The processing result will then be added to the input and then normalized. The Swin Transformer Block modules with even numbers use a multi-head self-attention processing mechanism based on a sliding window for feature extraction. In the second stage of mid-level feature extraction, a Patch Merging module is used to reduce the size of feature vectors to obtain hierarchical features, and two consecutive Swin Transformer Block modules are used to extract mid-level features. The deep feature extraction stage is divided into two sub-stages. In the first deep feature extraction stage, i.e., the third stage, a Patch Merging module and 18 consecutive Swin Transformer Block modules are combined to extract the deep features of this stage. In the second deep feature extraction stage, i.e., the fourth stage, a Patch Merging module and 2 consecutive Swin Transformer Block modules are combined to extract the deep features of this stage. The dual attention prediction and adaptive weighting network is used to perform weighted calculation of image quality evaluation based on channel feature information and spatial feature information extracted from the obtained feature map; The weighted calculation of image quality evaluation based on the channel feature information and spatial feature information extracted from the obtained feature map includes: Combined with the first image quality assessment result , and the second image quality evaluation result , the weighted calculation of image quality evaluation is performed using the following formula: ； in, and represents the weighting coefficient, which can be determined according to the adaptive parameters that can be learned during the training process, represents the weighted calculation result of the image quality evaluation obtained; The evaluation accuracy was measured by Pearson linear correlation coefficient (PLCC) and Spearman rank correlation coefficient (SROCC).

2. The method according to claim 1, characterized in that: The dual attention prediction and adaptive weighting network includes a channel attention branch network and a spatial attention branch network, wherein: The channel attention branch network includes a channel attention module for extracting and processing channel feature information of a feature map, and a fully connected layer for performing image quality assessment in combination with the obtained channel feature information; The spatial attention branch network includes a spatial attention module for extracting and processing spatial feature information of a feature map, and a fully connected layer for combining the obtained spatial feature information to perform image quality assessment.

3. The method according to claim 2, characterized in that Based on the channel feature information extracted from the obtained feature map, image quality assessment is performed, including: Convolution is performed on the feature maps of the four stages of multi-scale extracted by the feature extraction network to obtain a first feature map with a uniform number of corresponding channels; Each of the first feature maps is input into the channel attention branch network, and the channel attention branch network extracts the channel attention feature map from the first feature map, and performs image quality assessment based on the extracted channel attention feature map.

4. The method according to claim 3, characterized in that: The channel attention feature map extracted by the channel attention branch network is expressed as : ； in, represents the input feature vector, Represents the four stages of the feature extraction network based on Swin Transformer, Indicates the average pooling operation on the input feature vector. Indicates the maximum pooling operation on the input feature vector. , represents the fully connected layer, Represents the sigmoid nonlinear activation function; The first image quality assessment result predicted by the channel attention branch network is expressed as : ； in, represents the fully connected layer, Represents the dot product operation.

5. The method according to claim 4, characterized in that Based on the spatial feature information extracted from the obtained feature map, image quality evaluation is performed, including: The sizes of the feature maps of the last three stages extracted by the feature extraction network are adjusted respectively to obtain a second feature map whose image size is consistent with the feature map of the first stage; The second feature maps and the feature maps of the first stage are spliced ​​together, and the obtained new feature maps are input into the spatial attention branch network, and the spatial attention branch network extracts the spatial attention feature maps from the new feature maps, and performs image quality evaluation based on the extracted spatial attention feature maps.

6. The method according to claim 5, characterized in that The spatial attention feature map extracted by the spatial attention branch network is expressed as : ； in, It indicates that the number of output channels is 1, and a convolution operation with a convolution kernel of 7×7 and a stride of 1 is used. The second image quality assessment result predicted by the spatial attention branch network is expressed as : 。 7. An image quality evaluation system, characterized in that: The system comprises a data set construction module, a model construction module, a data preprocessing module and an image quality scoring module, wherein: the data set construction module is used to construct a training image data set, and the training image data set contains real distorted images; The model building module is used to build an initial network model for combining image channels and spatial feature information to perform image quality weighted evaluation calculation; The data preprocessing module is used to perform preprocessing operations on each image included in the training image data set according to a preset preprocessing method to obtain corresponding preprocessed training data; The model building module is further used to input the pre-processed training data into the initial network model for training to obtain a trained target network model; The image quality scoring module is used to perform the preprocessing operation on the real distorted image to be evaluated, and input the obtained preprocessed data to be evaluated into the target network model for processing to obtain a corresponding image quality score.

8. A readable storage medium, characterized in that the readable storage medium includes an image quality assessment method program, and when the image quality assessment method program is executed by a processor, the steps of the method as described in any one of claims 1 to 6 are implemented.

Images