No-reference image quality assessment method based on visual clue correlation analysis

Abstract

The application discloses a no-reference image quality evaluation method based on visual clue correlation analysis, and steps are as follows: a local non-equal interaction transpose attention subnetwork is constructed; a visual clue correlation analysis module is constructed; a no-reference image quality evaluation network is generated; all images in a generated training set are linearly coded; the no-reference image quality evaluation network is trained; and quality evaluation is performed on a no-reference image to be evaluated. The local non-equal interaction transpose attention subnetwork can imitate the perception process of a human visual system on an image, and the visual clue correlation analysis module can fuse the correlation between semantic information of different granularity levels. The application improves the prediction precision of distorted image quality, and has the advantages of higher consistency with human eye visual perception and stronger generalization performance when evaluating no-reference image quality.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically to a no-reference image quality assessment method based on visual cue association analysis in the field of image quality assessment technology. This invention can be used in scenarios where the degree of image distortion needs to be evaluated without additional information. Background Technology

[0002] Referenceless image quality assessment (BIQA) refers to a method that accurately and automatically predicts the perceived image quality without requiring a reference image. It is widely used in the design, optimization, and evaluation of image processing, computer vision, and computer graphics applications. Since various types of distortion can be introduced at each stage of these image processing systems, leading to image quality degradation, developing image quality assessment methods that can automatically predict human observer perception is particularly important.

[0003] Currently, no-reference image quality assessment methods are mainly divided into traditional methods based on handcrafted features and methods based on deep learning. Image quality assessment models based on handcrafted features typically consist of a feature extraction unit and a quality regression model. Depending on the feature extraction method, these models can be further divided into methods based on the statistical properties of natural scenes and methods guided by the human visual system. No-reference image quality assessment based on the statistical properties of natural scenes relies on the premise that high-fidelity images adhere to specific statistical properties. However, these statistical properties can change due to image quality degradation. The human visual system is the ultimate receiver of visual signals, and utilizing the perceptual characteristics of the human visual system is crucial in the design of image quality indicators. Currently, the two most prominent methods guided by the human visual system are those based on the free energy principle and those based on visual sensitivity. However, image quality assessment methods based on handcrafted features have some significant drawbacks, including reliance on specialized knowledge, limited generalization ability, and difficulty in adapting to new image types and degradation patterns. These limitations have prompted image quality assessment to shift towards machine learning-based methods to achieve more accurate, flexible, and automated image quality assessment.

[0004] Deep learning-based image quality assessment methods represent a significant leap forward in terms of consistency between subjective and objective assessment methods compared to traditional approaches. Abstract features learned from image classification tasks during pre-training show a high correlation with no-reference image quality assessment tasks. The transfer from image classification to no-reference image quality assessment primarily involves supervised constraints on abstract features by image quality labels, enabling the model to effectively construct a quality-aware feature manifold and achieve strong generalization capabilities. Currently, many no-reference image quality assessment methods have borrowed strategies from pre-trained language models and their variants for downstream tasks, achieving significant success. However, the projection distance between images with different distortions in the manifold space does not represent the quality-aware distance. In other words, the regression relationship between the quality label MOS score and the image itself is affected by the non-uniformity of the quality-aware feature manifold. This semantically aware pre-trained model encourages similar representations for images of the same category, while ignoring variations in perceived image quality.

[0005] Xiamen University disclosed a no-reference image quality assessment method based on a diffusion model in its patent application "A No-Reference Image Quality Assessment Method Based on Diffusion Model" (Application No.: CN202410070380.X; Publication No.: CN 117593296 A). The implementation scheme of this method is as follows: First, an image quality assessment model is created based on a BIQA teacher module, a BIQA student module, a feature fusion module, a noise adaptation module, and an output module. The BIQA teacher module uses a CLIP model as its backbone network, and the BIQA student module uses a Transformer model as its backbone network. Second, a large number of images are acquired, preprocessed, and an image dataset is constructed. Third, the image quality assessment model is trained using the image dataset. Fourth, the trained image quality assessment model is used to perform image quality assessment. The feature fusion module in this method fuses the input distortion features and quality level features at a granular level. A noise adaptation module ensures consistency between the fused coarse-to-fine granular features and the predefined noise level, thereby further enhancing the alignment of the quality level features and ultimately improving the accuracy of no-reference image quality assessment. However, this method still has shortcomings: the interaction process between local blocks in the image quality assessment process is symmetrical, and parts with different attention weights have the same impact on the final prediction result, which does not conform to the perceptual process of the human visual system, leading to inaccurate evaluation results.

[0006] Xi'an University of Electronic Science and Technology disclosed a method for evaluating the quality of no-reference images based on hybrid attention in its patent application document "A Method for Evaluating the Quality of No-Reference Images Based on Hybrid Attention" (application number: CN 202310294941.X; application publication number: CN 1163094886 A). The implementation scheme of this method is as follows: First, obtain the training sample set and the test sample set; Second, construct a network model based on hybrid attention for no-reference image quality assessment: This network model includes a saliency feature map generation model and a hybrid attention network. The distorted image is first input into the saliency feature map generation model to obtain the saliency features of the image, and then the original distorted image is fused with its saliency features; the resulting fused features are then passed through the hybrid attention network to achieve channel-level and pixel-level feature weight allocation, and finally, the weighted feature map is used to regress the image quality score; Third, iteratively train the network model based on hybrid attention for no-reference image quality assessment; Fourth, obtain the no-reference image quality assessment results: Use the test sample set from the first step as the input to the trained no-reference image quality assessment network model based on hybrid attention for forward inference to obtain the quality prediction score for each test sample. This method uses an attention weighting method combined with pyramid features to construct a saliency mapping weighted map, taking into account the differences in distortion in different regions of the image, which leads to different impacts of each distorted part on the overall image quality score, thus improving the accuracy of quality score prediction. However, this method still has shortcomings: it relies on information at a single level and is difficult to adapt to various complex distortion types and diverse image content. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing methods by proposing a no-reference image quality assessment method based on visual cue association parsing. This method solves the problem that existing attention interaction methods lack a primary-secondary relationship between tokens, leading to interactions between different regions of the image that do not conform to human perception. Furthermore, these methods largely rely on single-level features and are difficult to adapt to various complex distortion types and diverse image content.

[0008] The approach to achieving the objective of this invention is as follows: This invention constructs a locally non-equivalent transposed attention subnetwork to correct the interactions between local regions of an image into non-equivalent relationships. The distance between the current local content projected onto the metric subspace and other content in the image can represent the proximity between local subgraphs, and the argmax() function is used to find the "closest" parent node to the current local content. To mimic the attention interaction mode of the human visual system when evaluating images, this invention uses a transposed attention subnetwork. In forward attention, the current local content receives similarity information from all other nodes, while in backward attention, information is transmitted to other nodes according to normalized probability values, using the second-dimensional normalized information to transmit information to parts with high attention weights. This solves the problem in existing attention interaction modes where there is no primary or secondary relationship between tokens, leading to interactions between different regions of the image that do not conform to human perception. This invention uses differentiable sampling from group clustering tasks to construct the similarity between coarse-grained and fine-grained local blocks of the image. By using the association parsing module to find the similarity between local contents on image block nodes, the information of important similar local blocks is aggregated onto a single node. This solves the problem that most existing no-reference image quality assessment methods rely on single-level features and are difficult to adapt to various complex distortion types and diverse image content.

[0009] The specific steps to achieve the objective of this invention are as follows:

[0010] Step 1, construct the transposed attention subnetwork with local unequal interactions:

[0011] Construct two transposed attention subnetworks with the same structure but different parameters, and their structures are as follows: first convolutional layer, second convolutional layer, transpose operation layer, third convolutional layer, first normalization layer, fully connected layer, and second normalization layer. Residual connections are made between the first convolutional layer and the first normalization layer, and between the first normalization layer and the second normalization layer.

[0012] Step 2, construct the visual cue association analysis module:

[0013] Two visual cue association parsing modules with the same structure but different parameter settings were built. The structure of each visual cue association parsing module is as follows: first convolutional layer, second convolutional layer, probability estimation layer, and third convolutional layer. The first and third convolutional layers are connected by skip connections.

[0014] Step 3: Connect the first attention sub-network, the first visual cue association parsing module, the second attention sub-network, and the second visual cue association parsing module in series. Connect the first and second visual cue association parsing modules in parallel and then connect them in series with the image quality regression module to form a referenceless image quality evaluation network based on visual cue association parsing.

[0015] Step 4: Perform linear encoding on all images in the generated training set:

[0016] Step 5: Train a no-reference image quality assessment network based on visual cue association parsing:

[0017] Set the training parameters, input the training set into the no-reference image quality assessment network based on visual cue association parsing, and use the Adamw optimizer to iteratively update the network parameters until the loss function converges, thus obtaining the trained no-reference image quality assessment network.

[0018] Step 6: Perform quality assessment on the image to be evaluated that has no reference.

[0019] Using the same method as in steps 4 and 5, the no-reference image to be evaluated is normalized, preprocessed, and linearly encoded sequentially. The linear encoding result is then input into the trained no-reference image quality assessment network to output a quality assessment score.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] First, this invention constructs a transposed attention subnetwork with locally unequal interactions, correcting the interactions between local image regions to be unequal relationships. This overcomes the problem in existing attention interaction methods where the lack of a hierarchy between tokens leads to interactions between different image regions that do not conform to human perception. This makes the image evaluation process in this invention more consistent with the visual masking characteristics of human visual mechanisms, thereby improving the subjective consistency of image quality evaluation results.

[0022] Second, this invention employs differentiable sampling from group clustering tasks to construct a visual cue association parsing module, integrating the correlations between semantic information at different granular levels. This overcomes the limitations of existing no-reference image quality assessment methods, which mostly rely on single-level features and struggle to adapt to various complex distortion types and diverse image content. This gives the invention the advantages of good robustness and generalization ability in different scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0024] The following is in conjunction with the appendix Figure 1 The specific implementation steps of the embodiments of the present invention will be described in further detail below.

[0025] Step 1: Construct a transposed attention subnetwork with local non-equal interactions.

[0026] Two transposed attention subnetworks with the same structure but different parameters and local non-equal interactions are constructed. Their structures are as follows: first convolutional layer, second convolutional layer, transpose operation layer, third convolutional layer, first normalization layer, fully connected layer, and second normalization layer. Residual connections are made between the first convolutional layer and the first normalization layer, and between the first normalization layer and the second normalization layer.

[0027] The number of feature groups (Heads) in the first convolutional layer of the first attention subnetwork is set to 6, the number of input channels in the first to third convolutional layers are set to 64, 128 and 128 respectively, the size of the convolutional kernel is set to 1×1, 3×3 and 1×1 respectively, the stride is set to 1, and the number of nodes in the fully connected layer is set to 128.

[0028] The number of input channels in the first to third convolutional layers of the second attention subnetwork is set to 128, 256 and 256 respectively, the number of nodes in the fully connected layer is set to 64, and the other parameters are set the same as those in the first attention subnetwork.

[0029] Step 2: Construct a visual cue association analysis module.

[0030] Two visual cue association parsing modules with the same structure but different parameter settings were built. The structure of each visual cue association parsing module is as follows: first convolutional layer, second convolutional layer, probability estimation layer, and third convolutional layer. The first and third convolutional layers are connected by skip connections.

[0031] In the first visual cue association parsing module, the number of input and output channels of the first to third convolutional layers are all set to 128, the kernel size is set to 1×1, 3×3 and 1×1 respectively, the stride is set to 1, and the probability estimation layer uses the Gumble-softmax function.

[0032] In the second visual cue association parsing module, the number of input and output channels of the first to third convolutional layers is set to 64, and the other parameters are set in the same way as in the first visual cue association parsing module.

[0033] Step 3: Connect the first attention sub-network, the first visual cue association parsing module, the second attention sub-network, and the second visual cue association parsing module in series. Connect the first and second visual cue association parsing modules in parallel and then connect them in series with the image quality regression module to form a referenceless image quality evaluation network based on visual cue association parsing.

[0034] The image quality regression module is structured as follows: a normalization layer, a self-attention decoding block, and a fully connected layer. The number of nodes in the fully connected layer is set to 16. The self-attention decoding block consists of three cascaded self-attention decoding layers with identical structures and parameter settings. Each self-attention decoding layer consists of three cascaded convolutional layers, where the number of input channels and output channels of each convolutional layer is set to 192.

[0035] Step 4: Generate training and test sets.

[0036] In the embodiments of the present invention, the KonIQ-10k known database refers to the image quality assessment database proposed by Hosu V et al. in “Koniq-10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment[J].IEEE Transactions on Image Processing,2020,29:4041-4056.”, abbreviated as KonIQ-10k known database.

[0037] A sample set is formed by selecting 6000 images with dimensions W×H×C and their score labels, where W, H, and C represent the width, height, and number of channels of the RGB image, respectively, with W≥224, H≥224, and C=3.

[0038] Each image in the sample set was cropped into a sub-image with dimensions of 224×224×3. Each cropped sub-image was then randomly flipped and pixel normalized. The mean of the normalized sub-images in the sample set was set to mean = [0.485, 0.456, 0.406], and the standard deviation was set to std = [0.229, 0.224, 0.225]. The R, G, and B channels of the sub-images were processed using the following formulas:

[0039]

[0040] Where R, G, and B represent the pixel values ​​in the red, green, and blue channels of each sub-image in the sample set, respectively, and max represents the maximum possible pixel value in the sub-image. Norm G Norm B Norm These represent the pixel values ​​in the red, green, and blue channels of the sub-image after normalization.

[0041] Obtain the quality score label of the original image corresponding to each sub-image, and form a training set by combining all sub-images and their corresponding quality score labels.

[0042] Take another 1000 images and their corresponding quality score labels from the KonIQ-10k public database and generate a test set using the same steps as generating the training set.

[0043] Step 5: Perform linear encoding on all images in the training and test sets.

[0044] Each image in the training and test sets is divided into non-overlapping image blocks, resulting in 256 local blocks of size 14×14×3. Linear embedding encoding is used to map all local blocks of an image into feature vectors. All feature vectors are combined into an input sequence. A CLS is added at the beginning of the input sequence as a learnable token, where CLS is a random vector with the same dimension as the feature vectors.

[0045] Step 6: Train a no-reference image quality assessment network based on visual cue association parsing.

[0046] Set the training parameters, input the encoded training set into the no-reference image quality assessment network based on visual cue association parsing, and use the Adamw optimizer to iteratively update the network parameters until the loss function converges, thus obtaining the trained no-reference image quality assessment network.

[0047] The loss function is as follows:

[0048]

[0049] Where i represents the image index in the training set. y represents the quality score obtained after the i-th image is processed by the no-reference image quality assessment network. i Let represent the quality score label of the i-th image, and ∥·∥1 represent the absolute value loss function.

[0050] The specific settings for the training parameters are as follows: the small positive number is set to epsilon = 1e-8, and the learning rate is set to learn_ratio = 2 × 10. -5 The width and height of the subgraph are set to w = 224 and h = 224, the width and height of the local block are set to P = 14, and the number of training epochs is set to 30.

[0051] Step 7: Perform a quality assessment on the image to be evaluated that has no reference.

[0052] The linearly encoded test set is input into the trained no-reference image quality assessment network, which outputs a quality assessment score.

[0053] The effects of this invention will be further illustrated below with simulation experiments:

[0054] 1. Simulation experimental conditions:

[0055] The simulation experiment hardware platform of this invention is: CPU is Intel(R) Core(TM) i9-7900X CPU 3.30GHz, graphics card is NVIDIA Geforce RTX 3090 GPU.

[0056] The simulation experiment software platform of this invention is: the operating system version is Ubuntu 20.04.6LTS, the CUDA version is 11.3, and the PyTorch deep learning framework version is 11.8.

[0057] The input images used in the simulation experiments of this invention are derived from the well-known image quality assessment databases TID2013, KonIQ-10k, and LIVEC.

[0058] The TID2013 database includes 25 reference images and 3,000 distorted images.

[0059] The KonIQ-10k database contains 10,073 distorted images.

[0060] The LIVEC database contains 1,162 distorted images.

[0061] 2. Simulation content and result analysis:

[0062] The simulation experiment of this invention uses this invention and two existing technologies (LPF and MetaIQA) to perform no-reference image quality assessment on distorted images in three known image quality assessment databases: TID2013, KonIQ-10k, and LIVEC.

[0063] The evaluation index is obtained by calculating the consistency between the quality prediction value of the no-reference image and the image label. The evaluation index is used to measure the no-reference image evaluation effect of the present invention and two prior art methods (LPF method and MetaIQA method) on the distorted images in the above three known image quality evaluation databases.

[0064] The two existing technologies used in the simulation experiment are:

[0065] The LPF method refers to the no-reference image quality assessment method proposed by Huang Q et al. in "A Lightweight Parallel Framework for Blind Image Quality Assessment[J].arXiv preprint arXiv:2402.12043,2024.", abbreviated as LPF method.

[0066] MetaIQA refers to the no-reference image quality assessment method proposed by Zhu H et al. in "MetaIQA: Deep meta-learning for no-reference image quality assessment[C] / / Proceedings of the IEEE / CVF conferenceon computer vision and pattern recognition.2020:14143-14152.", abbreviated as MetaIQA method.

[0067] In the simulation experiment, the three known image quality assessment databases used are:

[0068] The TID2013 public database refers to the image quality assessment database proposed by Ponomarenko N et al. in "Color Image Database TID2013: Peculiarities and Preliminary Results [C]. In European Workshop on Visual Information Processing. IEEE, 2013: 106-111.", abbreviated as TID2013 public database.

[0069] The KonIQ-10k public database refers to the image quality assessment database proposed by Hosu V et al. in "Koniq-10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment[J].IEEE Transactions on Image Processing,2020,29:4041-4056.", abbreviated as KonIQ-10k public database.

[0070] The LIVEC database refers to the image quality assessment database proposed by Ghadiyaram D et al. in “Massive Online Crowdsourced Study of Subjective and Objective Picture Quality[J].IEEE Transactions on Image Processing,2015,25(1):372-387.”, abbreviated as LIVEC database.

[0071] To evaluate the effectiveness of this invention and two other prior art methods in evaluating the quality of no-reference images, this simulation experiment uses two indicators: Spearman Rank-Order Correlation Coefficient (SROCC) and Pearson Linear Correlation Coefficient (PLCC), to objectively assess the effectiveness of this invention and the two prior art methods in evaluating the quality of no-reference images.

[0072] Spearman Rank Order Correlation Coefficient (SROCC)

[0073] Spearman correlation determines the strength and direction of the monotonic relationship between two variables, measuring the monotonicity of the algorithm's predictions. The expression is:

[0074]

[0075] Where, r xi r represents the subjective quality assessment result of the i-th image to be tested. yi This represents the objective quality evaluation result, (r xi -r yi ) 2 This indicates the magnitude of the difference between the two, calculated using a sorted difference set.

[0076] Pearson Linear Correlation Coefficient (PLCC)

[0077] x i and y i These represent the subjective quality assessment score and the objective score of the i-th tested image, respectively. Their expressions are:

[0078]

[0079] Where n is the total number of images, and These are the average of the subjective evaluation scores given by the human eye to the database and the evaluation scores automatically calculated by the objective evaluation algorithm. The linear correlation coefficient describes the correlation between the algorithm's evaluation value and the subjective human score, and also measures the accuracy of the algorithm's predictions.

[0080] The simulation experiment of this invention uses the method of this invention and two existing technologies to evaluate images in three different known databases, and calculates two consistency indices for the evaluation results of each method. The calculation results are shown in Table 1.

[0081] Table 1. Comparison of Evaluation Results of the Three Methods

[0082]

[0083] As shown in Table 1, the Spearman rank correlation coefficient (SROCC) and Pearson linear correlation coefficient (PLCC) of the evaluation results of the present invention on three known image quality assessment databases are both higher than those of the two existing technologies, proving that the present invention has a better effect on no-reference image quality assessment.

Claims

1. A method for evaluating the quality of no-reference images based on visual cue association analysis, characterized in that, The quality assessment of no-reference images is performed by constructing a no-reference image quality assessment network based on visual cue association parsing, which includes a transposed attention subnetwork with local non-equivalent interactions and a visual cue association parsing module. The steps of this no-reference image assessment method are as follows: Step 1, construct the transposed attention subnetwork with local unequal interactions: Construct two transposed attention subnetworks with the same structure but different parameters, each with a local non-equal interaction. The structure of each subnetwork is as follows: first convolutional layer, second convolutional layer, transpose operation layer, third convolutional layer, first normalization layer, fully connected layer, and second normalization layer. Residual connections are made between the first convolutional layer and the first normalization layer, and between the first normalization layer and the second normalization layer. Step 2, construct the visual cue association analysis module: Two visual cue association parsing modules with the same structure but different parameter settings were built. The structure of each visual cue association parsing module is as follows: first convolutional layer, second convolutional layer, probability estimation layer, and third convolutional layer. The first and third convolutional layers are connected by skip connections. Step 3: Connect the first attention sub-network, the first visual cue association parsing module, the second attention sub-network, and the second visual cue association parsing module in series. Connect the first and second visual cue association parsing modules in parallel and then connect them in series with the image quality regression module to form a referenceless image quality evaluation network based on visual cue association parsing. Step 4: Perform linear encoding on all images in the generated training set: Step 5: Train a no-reference image quality assessment network based on visual cue association parsing: Set the training parameters, input the training set into the no-reference image quality assessment network based on visual cue association parsing, and use the Adamw optimizer to iteratively update the network parameters until the loss function converges, thus obtaining the trained no-reference image quality assessment network. Step 6: Perform quality assessment on the image to be evaluated that has no reference. Using the same method as in step 4, the no-reference image to be evaluated is normalized, preprocessed, and linearly encoded sequentially. The linear encoding result is then input into the trained no-reference image quality assessment network to output a quality assessment score.

2. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The transposed attention subnetwork with different parameters in step 1 refers to setting the number of feature groups (Heads) of the first convolutional layer in the first attention subnetwork to 6, the number of input channels of the first to third convolutional layers to 64, 128 and 128 respectively, the size of the convolutional kernel to 1×1, 3×3 and 1×1 respectively, the stride to 1 for all of them, and the number of nodes in the fully connected layer to 128; and setting the number of input channels of the first to third convolutional layers in the second attention subnetwork to 128, 256 and 256 respectively, the number of nodes in the fully connected layer to 64, and the remaining parameter settings to be the same as those of the first attention subnetwork.

3. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The visual cue association parsing modules with different parameter settings mentioned in step 2 refer to the following: in the first visual cue association parsing module, the number of input and output channels of the first to third convolutional layers are all set to 128, the kernel size is set to 1×1, 3×3 and 1×1 respectively, the stride is set to 1, and the probability estimation layer uses the Gumble-softmax function; in the second visual cue association parsing module, the number of input and output channels of the first to third convolutional layers are all set to 64, and the remaining parameter settings are the same as those of the first visual cue association parsing module.

4. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The structure of the image quality regression module in step 3 is as follows: normalization layer, self-attention decoding block, and fully connected layer. The number of nodes in the fully connected layer is set to 16. The self-attention decoding block is composed of three cascaded self-attention decoding layers with the same structure and parameter settings. Each self-attention decoding layer is composed of three cascaded convolutional layers. The number of input channels and the number of output channels of the convolutional layers are all set to 192.

5. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The normalization process mentioned in step 4 refers to: setting the mean of the normalized image in the sample set to mean = [0.485, 0.456, 0.406] and the standard deviation to std = [0.229, 0.224, 0.225], and processing the R, G, and B channels of the image using the following formulas respectively: Where R, G, and B represent the pixel values ​​of each image in the sample set in the red, green, and blue channels, respectively, and max represents the maximum possible pixel value in the image. Norm G Norm B Norm These represent the pixel values ​​in the red, green, and blue channels of the image after normalization.

6. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The steps for generating the training set in step 4 are as follows: The first step is to form a sample set by taking L images of dimensions W×H×C and their score labels, where L≥1000, W, H, and C represent the width, height, and number of channels of the RGB image, respectively, W≥224, H≥224, and C=3; The second step is to crop each image in the sample set into a w×h×c sub-image, and then perform random flipping and pixel normalization on each cropped sub-image in turn. Here, w, h, and c represent the width, height, and dimension of the input image of the no-reference image quality assessment network based on visual cue association parsing, respectively. The third step is to add a quality score label to each processed sub-image; this score is the average of the evaluation scores of the distortion level of the same sub-image by M evaluators. The fourth step is to combine all the processed sub-images with their corresponding quality score labels to form a training set.

7. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The linear encoding mentioned in step 4 refers to: dividing each image in the training set into non-overlapping image blocks to obtain N local blocks of size P×P×3; mapping all local blocks of an image to feature vectors through linear embedding encoding; forming an input sequence with all feature vectors; and adding a CLS as a learnable token at the beginning of the input sequence. Here, N represents the total number of non-overlapping image blocks, P and P represent the width and height of each local block, respectively, P is a common factor of w and h, and CLS is a random vector with the same dimension as the feature vectors.

8. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The specific settings for the training parameters in step 5 are as follows: the small positive number is set to epsilon = 1e-8, and the learning rate is set to learn_ratio = 2 × 10. -5 The width and height of the subgraph are set to w = 224 and h = 224, the width and height of the local block are set to P = 14, and the number of training epochs is set to 30.

9. The method for evaluating the quality of no-reference images based on visual cue association analysis according to claim 1, characterized in that, The loss function mentioned in step 5 is as follows: Where i represents the image index in the training set. y represents the quality score obtained after the i-th image is processed by the no-reference image quality assessment network. i Let represent the quality score label of the i-th image, and ∥·∥1 represent the absolute value loss function.

Images