A reference-free quality evaluation method and system for invisible watermark images

Abstract

The present invention discloses a reference-free quality evaluation method and system for invisible watermark images, which relate to the technical field of image processing. The invisible watermark image is input into a trained reference-free image quality evaluation network to output a quality evaluation result of the invisible watermark image. The training process of the reference-free image quality evaluation network is as follows: style matching based on watermark distortion is used to obtain a style subset; pseudo-labels are fitted to damaged images in the style subset to obtain image quality labels; image quality labels are assigned to a reconstructed invisible watermark image set to establish an invisible watermark image data set with image quality labels; the reference-free image quality evaluation network is trained using the invisible watermark image data set. The reference-free quality evaluation method and system can effectively capture subtle feature information of the invisible watermark image and achieve an accurate prediction effect.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a reference-free quality evaluation method and system for invisible watermarked images. Background Art

[0002] In recent years, with the increasing demand for digital content protection and copyright management, digital watermarking technology has seen widespread development. Watermarking plays a vital role in multimedia applications, including copyright protection, content authentication, and tamper detection. In many scenarios, such as distance education, teleworking, screen sharing, and online advertising, the quality of invisible watermarked images is crucial to the user experience. Therefore, effectively evaluating the quality of invisible watermarked images has become an important research topic.

[0003] However, due to factors such as device performance, information loss during storage or transmission, images with invisible watermarks may produce various types of distortions, such as contrast changes, compression distortion, etc. These distortions not only affect the transparency and robustness of the watermark, but also affect the overall visual experience of users.

[0004] Over the past few decades, research on image quality assessment has primarily focused on natural images and images without watermarks. However, research has found that features extracted from natural images are not generally applicable to watermarked images. In particular, the types and distribution of distortion in invisible watermarked images differ significantly from those in natural images. Consequently, no-reference quality assessment methods designed for natural images cannot be directly applied to invisible watermarked images.

[0005] Currently, most no-reference image quality assessment methods predict quality scores by extracting features from the entire image. However, invisible watermarked images, due to their unique characteristics and distortion types, require more detailed feature extraction and analysis. Furthermore, existing methods typically only consider spatial intensity and distribution when measuring image structure, ignoring the watermark's unique directional information and frequency domain characteristics. Therefore, the measurement of structural information in invisible watermarked images can be further optimized. Summary of the Invention

[0006] Based on the technical problems existing in the background technology, the present invention proposes a reference-free quality evaluation method and system for invisible watermark images, which can effectively capture the subtle feature information of invisible watermark images, reflect the perceptual quality of the image, and achieve accurate prediction results.

[0007] The present invention proposes a no-reference quality assessment method for invisible watermark images, which inputs the invisible watermark image into a trained no-reference image quality assessment network to output a quality assessment result of the invisible watermark image;

[0008] The training process of the no-reference image quality assessment network is as follows:

[0009] Step 1: Style matching of watermark distortion: Calculate the generated invisible watermark image set The difference between each invisible watermark image and the damaged image in the pre-set classic image quality evaluation dataset is calculated, and the calculated distortion types are arranged in ascending order. The damaged images corresponding to the top n distortion types constitute a style subset.

[0010] Step 2: Pseudo-label fitting: A pseudo-label is constructed by combining the scores of each damaged image in the style subset using three full-reference image quality assessment methods: structure, texture, and color. The Pearson product-moment correlation coefficient and Spearman rank correlation coefficient are used as metrics to evaluate the pseudo-label's approximation to the true label, thereby fitting the image quality label.

[0011] Step 3: Construct invisible watermark image dataset: Regenerate a set of invisible watermark image datasets with different embedding capacities and different robustness , and is the generated invisible watermark image set Assign image quality labels and build an invisible watermark image dataset with image quality labels ;

[0012] Step 4: Training of the no-reference image quality assessment network: using the invisible watermark image dataset The no-reference image quality assessment network is trained, and a loss function is constructed to adjust the trainable parameters in the no-reference image quality assessment network.

[0013] Furthermore, in step 2, the pseudo-label The build is as follows:

[0014] ;

[0015] 、 、 、 Optimize by:

[0016] ;

[0017] in, 、 、 Evaluation scores 、 、 The weight of To balance the weights, is the evaluation score obtained by the structure-based full-reference image quality evaluation method, is the evaluation score obtained by the texture-based full-reference image quality evaluation method, is the evaluation score obtained by the color-based full-reference image quality evaluation method, is the Pearson product-moment correlation coefficient, is the Spearman rank correlation coefficient, is the true label.

[0018] Furthermore, in step 1, the generated invisible watermark image set is calculated The difference between each invisible watermark image and the damaged image in the pre-set classic image quality evaluation dataset is calculated as follows:

[0019] ;

[0020] in, is the difference value, Invisible watermark image set Middle An invisible watermark image, A classic image quality evaluation dataset Middle The first in the group A damaged image, and They are the first The eccentric covariance matrix between the feature map and the features calculated by the layer, is the number of layers of the visual geometry group network, The image index.

[0021] Furthermore, in step 4, the no-reference image quality assessment network uses the ResNet18 framework as the backbone network, and the input of the backbone network is the invisible watermark image dataset. The output of the backbone network is connected to a fully connected layer to output the quality score of the invisible watermark image;

[0022] The loss function The definition is as follows:

[0023] ;

[0024] in, Invisible watermark image dataset Middle The image quality label corresponding to the invisible watermark image, Invisible watermark image dataset Middle The invisible watermark image passes through the quality score output by the no-reference image quality evaluation network, Invisible watermark image dataset The total number of invisible watermark images in .

[0025] A no-reference quality assessment system for invisible watermarked images, which inputs the invisible watermarked image into a trained no-reference image quality assessment network to output a quality assessment result of the invisible watermarked image;

[0026] The training process of the no-reference image quality assessment network is as follows:

[0027] Step 1: Style matching of watermark distortion: Calculate the generated invisible watermark image set The difference between each invisible watermark image and the damaged image in the pre-set classic image quality evaluation dataset is calculated, and the calculated distortion types are arranged in ascending order. The damaged images corresponding to the top n distortion types constitute a style subset.

[0028] Step 2: Pseudo-label fitting: A pseudo-label is constructed by combining the scores of each damaged image in the style subset using three full-reference image quality assessment methods: structure, texture, and color. The Pearson product-moment correlation coefficient and Spearman rank correlation coefficient are used as metrics to evaluate the pseudo-label's approximation to the true label, thereby fitting the image quality label.

[0029] Step 3: Construct invisible watermark image dataset: Regenerate a set of invisible watermark image datasets with different embedding capacities and different robustness , and is the generated invisible watermark image set Assign image quality labels and build an invisible watermark image dataset with image quality labels ;

[0030] Step 4: Training of the no-reference image quality assessment network: using the invisible watermark image dataset The no-reference image quality assessment network is trained, and a loss function is constructed to adjust the trainable parameters in the no-reference image quality assessment network.

[0031] Furthermore, in step 2, the pseudo-label The build is as follows:

[0032] ;

[0033] 、 、 、 Optimize by:

[0034] ;

[0035] in, 、 、 Evaluation scores 、 、 The weight of To balance the weights, is the evaluation score obtained by the structure-based full-reference image quality evaluation method, is the evaluation score obtained by the texture-based full-reference image quality evaluation method, is the evaluation score obtained by the color-based full-reference image quality evaluation method, is the Pearson product-moment correlation coefficient, is the Spearman rank correlation coefficient, is the true label.

[0036] The advantages of the non-reference quality assessment method and system for invisible watermark images provided by the present invention are: the distortion type similar to the watermark distortion is selected by style matching, which fully considers the characteristics of the invisible watermark image. Starting from the structure, texture, and color information, a pseudo-label scoring method based on the full reference image quality assessment method is adopted to design image quality labels for invisible watermark images, avoiding the tedious and laborious manual scoring. Utilizing the classic watermark embedding framework, a relatively sufficient invisible watermark image dataset is established with full consideration of the requirements of embedding content and robustness. The non-reference image quality assessment network is designed using the classic ResNet as the basic framework, which can effectively capture the subtle feature information of the invisible watermark image, reflect the perceptual quality of the image, and achieve accurate prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the process of the present invention;

[0038] Figure 2 It is a logical framework diagram of the present invention;

[0039] Figure 3 Schematic diagram of the quality evaluation performance of the present invention on an invisible watermark image dataset. DETAILED DESCRIPTION

[0040] The technical solutions of the present invention are described in detail below through specific embodiments. Numerous specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0041] like Figures 1 to 3 As shown, the present invention proposes a no-reference quality assessment method for invisible watermark images, which inputs the invisible watermark image into the trained no-reference image quality assessment network to output the quality assessment result of the invisible watermark image;

[0042] The training process of the no-reference image quality assessment network is as follows:

[0043] Step 1: Style matching of watermark distortion: Calculate the generated invisible watermark image set The difference between each invisible watermark image and the damaged image in the pre-set classic image quality evaluation dataset is calculated, and the calculated distortion types are arranged in ascending order. The damaged images corresponding to the top n distortion types constitute a style subset.

[0044] Different robustness and capacity requirements will result in different watermark features embedded in the image. Therefore, it is necessary to match all possible watermark distortion features in the distortion types of the pre-set classic image quality assessment dataset.

[0045] Specifically:

[0046] First, a large number of invisible watermark images are generated with different settings (i.e., different capacity, different robustness) to form an invisible watermark image set. ,in, For the An invisible watermark image, , Invisible watermark image set The number of images in

[0047] Then use the following formula to calculate the invisible watermark image set Each invisible watermark image in the image is compared with the pre-set classic image quality evaluation dataset The difference between the damaged images in :

[0048] ;

[0049] in, is the difference value, Invisible watermark image set Middle An invisible watermark image, A classic image quality evaluation dataset Middle The first in the group A damaged image, and They are the first The eccentric covariance matrix (Gram matrix) between the feature map and the feature feature calculated by the layer, is the number of layers of the visual geometry group network, is the image index, i.e. the invisible watermark image set Compared with classic image quality assessment dataset The image content in is the same, except The middle one is a watermarked image, but the watermark is not visible. These are damaged images with different distortion types, without watermarks.

[0050] In addition, ImageNet is the name of a computer vision system recognition project. It is currently the largest image recognition database in the world. It was established by computer scientists at Stanford University in the United States to simulate the human recognition system.

[0051] After calculating all the style differences, the top n groups with the minimum values ​​are selected as the most matching style subsets, thereby matching all possible watermark distortion features in the distortion types of the pre-set classic image quality evaluation dataset.

[0052] For example, as shown in Table 1, when the classic image quality assessment dataset selects the image quality assessment dataset TID2013 (IQA dataset TID2013), there are 24 types of distortion in the IQA dataset TID2013. The invisible watermark image is compared with these 24 types of distortion to calculate the style matching similarity. The 9 most similar types are selected, namely additive Gaussian noise, mask noise, high-frequency noise, JPEG compression, unbiased pattern noise, comfort noise, image color quantization dithering, chromatic aberration, and sparse sampling and reconstruction. The results are shown in the table:

[0053] Table 1 shows the distortion types in the subset of the IQA dataset TID2013 and their corresponding relationships with the actual situation and HVS.

[0054]

[0055] According to Table 1, distortion types similar to the watermark distortion style are selected from the IQA dataset TID2013, and the possible distortions caused by the watermark are comprehensively considered.

[0056] Step 2: Pseudo-label fitting: A pseudo-label is constructed by combining the scores of each damaged image in the style subset using three full-reference image quality assessment methods: structure, texture, and color. The Pearson product-moment correlation coefficient and Spearman rank correlation coefficient are used as metrics to evaluate the pseudo-label's approximation to the true label, thereby fitting the image quality label.

[0057] Considering that watermark distortion mainly involves the human eye's perception of the structure, texture, and color of the image, this embodiment selects three classic full-reference image quality evaluation methods from the aspects of structure, texture, and color to construct a pseudo-label mechanism, obtain a pseudo-label fitting formula, and use the Pearson linear correlation coefficient (PLCC) and Spearman rank correlation coefficient (SROCC) indicators to design a pseudo-label score. The higher the PLCC and SROCC values, the better the pseudo-label fitting formula and the more accurate the score.

[0058] Specifically (a1) to (a3)

[0059] (a1) Three full-reference image quality assessment methods are selected to evaluate the structure, texture, and color of invisible watermark images: feature similarity index (FSIMc), gradient magnitude similarity deviation (GMSD), and full-reference image quality assessment (VCGS) based on visual saliency based on color appearance and gradient similarity. FSIMc is used to evaluate structure, GMSD is used to evaluate texture, and VCGS is used to evaluate color.

[0060] (a2) The overall pseudo-labeling mechanism is obtained based on the joint score of the three evaluation methods (FSIMc, GMSD, VCGS), as shown in the following formula:

[0061]

[0062] in, is the joint score, 、 、 Evaluation scores 、 、 The weight of To balance the weights, is the evaluation score obtained by the structure-based full reference image quality evaluation method, that is, the evaluation score corresponding to FSIMc, is the evaluation score obtained by the texture-based full-reference image quality evaluation method, that is, the evaluation score corresponding to GMSD, is the evaluation score obtained by the color-based full-reference image quality evaluation method, that is, the evaluation score corresponding to VCGS.

[0063] (a3) Use the Pearson product-moment correlation coefficient (PLCC) and the Spearman rank correlation coefficient (SROCC) as indicators to evaluate the performance of pseudo-labels in approximating true labels. 、 、 、 Optimize by:

[0064] ;

[0065] in and The calculation formula is:

[0066] ;

[0067] ;

[0068] in, represents the true label, express and The covariance between represents the standard deviation, Is an invisible watermark image set Middle The ranking difference between the invisible watermarked images in the pseudo-label and the real label, is the number of damaged images in the style subset, is the Pearson product-moment correlation coefficient, is the Spearman rank correlation coefficient.

[0069] Three quality scores are calculated for the damaged images in the style subset using three full-reference image quality evaluation methods, and the joint score is obtained using the pseudo-label mechanism formula in (a2) Then the pseudo labels are evaluated by the Pearson linear correlation coefficient and Spearman rank correlation coefficient of (a3) ​​to determine 、 、 、 The specific value of makes the pseudo label closer to the real label, and sets the joint score in the fitting process. The value range is 0-5, guiding 、 、 、 Generate accurate values, and then get the determined pseudo labels based on the pseudo label mechanism formula, and use the determined pseudo labels as the image quality labels to be fitted , thereby calculating the image quality from three aspects: structure, texture and color.

[0070] To evaluate the approximate performance of pseudo labels, this embodiment compares the proposed scoring mechanism with eight full-reference image quality assessment methods (peak signal-to-noise ratio (PSNR), structural similarity (SSIM), feature similarity (FSIMc), edge similarity for screen content image quality assessment (ESIM), image quality assessment based on gradient similarity (GSM), gradient magnitude similarity deviation (GMSD), the most obvious distortion: full-reference image quality assessment and the role of strategy (MAD), and full-reference image quality assessment based on visual saliency of color appearance and gradient similarity (VCGS)). As can be seen from the PLCC and SROCC values ​​in Table 2, the pseudo label fitting method proposed in this embodiment achieves the best performance:

[0071] Table 2 shows the performance comparison of the classic full-reference image quality assessment (FR-IQA) method on a subset of the IQA dataset TID2013.

[0072]

[0073] In Table 2, Ours refers to the performance parameters corresponding to the no-reference quality evaluation method of this embodiment.

[0074] Step 3: regenerate a set of invisible watermark images with different embedding capacities and robustness, assign image quality labels to the generated invisible watermark images, and establish an invisible watermark image dataset with image quality labels;

[0075] Currently, there is a lack of databases containing invisible watermarked images and quality labels, and manual scoring is time-consuming and labor-intensive, resulting in a lack of reference-free quality assessment networks specifically for invisible watermarked images. This lack of databases has severely restricted the research and development of invisible watermarked image quality assessment methods.

[0076] Therefore, the purpose of establishing a database in this embodiment is to train a reference-free image quality assessment network that can evaluate the quality of invisible watermark images. Therefore, the images in the database should include different robustness and embedding capabilities so that the trained reference-free image quality assessment network has good performance, can efficiently extract the features of invisible watermark images and obtain accurate quality scores.

[0077] The establishment of invisible watermark image dataset includes the following steps:

[0078] (b1) Use pre-trained watermark models with different robustness and embedding capabilities to generate invisible watermark image sets ,in, for Middle An invisible watermark image, , , Invisible watermark image set The number of images in Invisible watermark image set The number of images in , specifically set as follows:

[0079] Robustness: Dropout, Cropout, Gaussian noise, Gaussian blur, JPEG compression, Median filter, Salt and Pepper noise;

[0080] Embedding capacity: 30, 60, 90, 120 bits;

[0081] It should be noted that the invisible watermark image set Set of images with invisible watermarks The images in are invisible watermark images generated with the same robustness and embedding ability, except that the invisible watermark image set The invisible watermark images in the dataset are relatively small (for example, 30 images), so as to facilitate style matching with the damaged images in the classic image quality assessment dataset, thereby obtaining the closest n types of distortion. Based on the style matching, the style subset containing n types of distortion is fitted with pseudo labels, thereby obtaining n types of image quality labels. Then, the invisible watermark image set is used to match the damaged images in the dataset. There are many invisible watermark images in the image set (for example, 10,000 images), and then Each image in the image is assigned a corresponding label according to the generated n types of image quality labels, thereby constructing an invisible watermark image dataset for training a reference-free image quality assessment network. , where the invisible watermark image set and invisible watermark image sets The images in the same.

[0082] (b2) Use the pseudo-label fitting method in step 2 to make invisible watermark image set Assign an image quality label to each image in , the image and the corresponding image quality label are combined to form a data pair, and the invisible watermark image set The images in the image list are used to construct corresponding data pairs, and all data pairs are combined to form an invisible watermark image dataset. , yes The set of image quality labels corresponding to each image in .

[0083] Step 4: Evaluate network training: using invisible watermark image dataset The no-reference image quality assessment network is trained, and a loss function is constructed to adjust the trainable parameters in the no-reference image quality assessment network.

[0084] Use the established invisible watermark image dataset A no-reference image quality assessment network is trained with the images and corresponding image quality labels in the dataset. The trained no-reference image quality assessment network is required to be able to evaluate the quality of the invisible watermarked image without the original reference image. That is, the more similar it is to the original image, the higher the score should be.

[0085] The no-reference image quality evaluation network uses the classic ResNet framework as a no-reference quality evaluation model The backbone network of the . Specific settings: It consists of 4 "Res-Blocks (residual blocks)", the convolution kernel size is 3×3, the step size is 1; the receiving size is Watermark image As input, is a real number, Invisible watermark image dataset The height and width of the middle image, the last fully connected layer FC is adjusted to a single channel, and a quality score is output ,represent The quality score of

[0086] Setting the loss function for the no-reference image quality assessment network , requiring the quality score corresponding to each invisible watermark image Should be associated with the image quality label Tightly aligned, as follows:

[0087] ;

[0088] in, Invisible watermark image dataset Middle The image quality label corresponding to the invisible watermark image, Invisible watermark image dataset Middle The invisible watermark image passes through the quality score output by the no-reference image quality evaluation network, Invisible watermark image dataset The total number of invisible watermark images in .

[0089] As shown in Table 3, it shows the performance comparison of the method proposed in this embodiment in the established invisible watermark image database;

[0090] Table 3 shows the results of the no-reference image quality assessment network in this embodiment on the invisible watermark image dataset.

[0091]

[0092] From Table 3, we can see that the PLCC and SROCC of the proposed method reach 0.9435 and 0.8869, respectively, showing excellent performance;

[0093] According to steps one to four, this embodiment matches the watermark distortion with the distortion type of the classic image quality assessment dataset, and selects the distortion type that is closest to the watermark distortion style as the style subset. Then, based on the style subset, the invisible watermark image features are considered from the aspects of structure, texture, and color, and three classic full-reference image quality assessment methods are used to fit a more accurate image quality label as a pseudo-label scoring method. The classic watermark algorithm then generates a large number of invisible watermark images with different robustness and different embedding capabilities, and uses the pseudo-label method to assign a label to each image, thereby establishing an invisible watermark image dataset with quality labels. ; Finally, using this dataset , train a no-reference image quality evaluation network based on Resnet18 for no-reference quality evaluation of invisible watermark images, and obtain the following Figure 3 The figure shows the performance of the quality evaluation of the present embodiment in the invisible watermark image dataset. The score of the present embodiment is close to the pseudo-label score, which can effectively reflect the visual quality of the image.

[0094] Therefore, this embodiment uses style matching to select distortion types similar to the watermark distortion, fully considering the characteristics of the invisible watermark image. Starting from the structure, texture, and color information, a pseudo-label scoring method based on the full reference image quality evaluation method is used to design image quality labels for the invisible watermark image, avoiding the tedious and laborious manual scoring. Using the classic watermark embedding framework, fully considering the requirements of embedding content and robustness, a relatively comprehensive invisible watermark image dataset is established. The classic ResNet is used as the basic framework to design a reference-free image quality evaluation network, which can effectively capture the subtle feature information of the invisible watermark image, reflect the perceptual quality of the image, and achieve accurate prediction results.

[0095] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A reference-free quality assessment method for invisible watermarked images, characterized in that: Inputting the invisible watermark image into the trained reference-free image quality assessment network to output the quality assessment result of the invisible watermark image; The training process of the no-reference image quality assessment network is as follows: Step 1: Style matching of watermark distortion: Compute the generated invisible watermark image set The difference between each invisible watermark image and the damaged image in the pre-set classic image quality evaluation dataset is calculated, and the calculated distortion types are arranged in ascending order. The damaged images corresponding to the first n distortion types constitute a style subset, where the calculation formula of the difference value is as follows: ； in, is the difference value, Invisible watermark image set Middle An invisible watermark image, A classic image quality assessment dataset Middle The first A damaged image, and They are the first The eccentric covariance matrix between the feature map and the features calculated by the layer, is the number of layers of the visual geometry group network, Index for images; Step 2: Pseudo-label fitting: Use the three full-reference image quality evaluation methods of structure, texture, and color to construct a pseudo-label based on the joint score of each damaged image in the style subset. Use the Pearson product-moment correlation coefficient and the Spearman rank correlation coefficient as indicators to evaluate the performance of the pseudo-label in approximating the true label, thereby fitting the image quality label. The build is as follows: ； , , , Optimize by: ； in, , , Evaluation scores , , The weight of To balance the weight, is the evaluation score obtained by the structure-based full-reference image quality evaluation method, is the evaluation score obtained by the texture-based full-reference image quality evaluation method, is the evaluation score obtained by the color-based full-reference image quality evaluation method, is the Pearson product-moment correlation coefficient, is the Spearman rank correlation coefficient, is the true label; Step 3: Construct invisible watermark image dataset: Regenerate a set of invisible watermark image datasets with different embedding capacities and different robustness , and is the generated invisible watermark image set Assign image quality labels and build an invisible watermark image dataset with image quality labels ; Step 4: Training of the No-reference Image Quality Assessment Network: Using the Invisible Watermark Image Dataset The no-reference image quality assessment network is trained, and a loss function is constructed to adjust the trainable parameters in the no-reference image quality assessment network.

2. The reference-free quality assessment method for invisible watermarked images according to claim 1, characterized in that: In step 4, the no-reference image quality assessment network uses the ResNet18 framework as the backbone network, and the input of the backbone network is the invisible watermark image dataset The output of the backbone network is connected to a fully connected layer to output the quality score of the invisible watermark image. The loss function The definition is as follows: ； in, Invisible watermark image dataset Middle The image quality label corresponding to the invisible watermark image, Invisible watermark image dataset Middle The invisible watermark image is outputted by the quality score of the no-reference image quality evaluation network. Invisible watermark image dataset The total number of invisible watermark images in .

3. A reference-free quality assessment system for invisible watermarked images, characterized in that: Inputting the invisible watermark image into the trained reference-free image quality assessment network to output the quality assessment result of the invisible watermark image; The training process of the no-reference image quality assessment network is as follows: Step 1: Style matching of watermark distortion: Compute the generated invisible watermark image set The difference between each invisible watermark image and the damaged image in the pre-set classic image quality evaluation dataset is calculated, and the calculated distortion types are arranged in ascending order. The damaged images corresponding to the first n distortion types constitute a style subset, where the calculation formula of the difference value is as follows: ； in, is the difference value, Invisible watermark image set Middle An invisible watermark image, A classic image quality assessment dataset Middle The first A damaged image, and They are the first The eccentric covariance matrix between the feature map and the features calculated by the layer, is the number of layers of the visual geometry group network, Index for images; Step 2: Pseudo-label fitting: Use the three full-reference image quality evaluation methods of structure, texture, and color to construct a pseudo-label based on the joint score of each damaged image in the style subset. Use the Pearson product-moment correlation coefficient and the Spearman rank correlation coefficient as indicators to evaluate the performance of the pseudo-label in approximating the true label, thereby fitting the image quality label. The build is as follows: ； , , , Optimize by: ； in, , , Evaluation scores , , The weight of To balance the weight, is the evaluation score obtained by the structure-based full-reference image quality evaluation method, is the evaluation score obtained by the texture-based full-reference image quality evaluation method, is the evaluation score obtained by the color-based full-reference image quality evaluation method, is the Pearson product-moment correlation coefficient, is the Spearman rank correlation coefficient, is the true label; Step 3: Construct invisible watermark image dataset: Regenerate a set of invisible watermark image datasets with different embedding capacities and different robustness , and is the generated invisible watermark image set Assign image quality labels and build an invisible watermark image dataset with image quality labels ; Step 4: Training of the No-reference Image Quality Assessment Network: Using the Invisible Watermark Image Dataset The no-reference image quality assessment network is trained, and a loss function is constructed to adjust the trainable parameters in the no-reference image quality assessment network.

Images