Underwater enhanced image quality evaluation method and device based on prior guidance, and medium
By introducing the fusion of prior information and convolutional features in the underwater image quality evaluation, the problems of inaccuracy and high labor consumption of quality evaluation in the prior art are solved, and more accurate underwater enhanced image quality evaluation is achieved.
Patent Information
- Application Number
- CN202411342178.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2024-12-27
AI Technical Summary
The existing underwater image quality evaluation method cannot accurately measure the quality of the image after enhancement, especially ignores distortions such as over-enhancement and structural loss, and has poor correlation with human visual perception, resulting in a large number of human and material resources required in a large number of image libraries. Subjective evaluation.
A method of underwater enhanced image quality evaluation based on prior guidance was designed. Two-dimensional depth maps and one-dimensional prior parameters were extracted through the prior estimation module, combined with convolutional features for fusion, and the quality evaluation module was used to predict image quality scores. Imitate physical parameters and subjective perceptual rules during underwater imaging.
It improves the accuracy of underwater enhanced image quality evaluation, can pay more accurately to the image content and distortion, reduces manpower and material investment, and improves the efficiency and accuracy of evaluation.
Smart Images
Figure CN119205697A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a method, device and medium for evaluating underwater enhanced image quality based on prior guidance. Background Art
[0002] The original underwater images acquired by underwater robots or underwater autonomous vehicles have serious degradation, including but not limited to color cast, blur, uneven background light and low contrast, which will lead to the degradation of image quality, not only affecting people's viewing experience, but also seriously affecting various machines' recognition of image content and other tasks. Underwater image enhancement methods are widely used to improve the quality of underwater images. Accurate quality evaluation is of great significance for guiding underwater image enhancement tasks, but due to the lack of corresponding quality evaluation indicators, it is often difficult to accurately measure the quality of the enhancement results.
[0003] Existing underwater enhanced image quality evaluation methods are mainly based on physical models. The public underwater image quality evaluation methods mainly include UCIQE, UIQM, CCF and NUIQ algorithms. Among them, UCIQE extracts the three characteristics of image chromaticity, saturation and contrast, and obtains the quality score of the image through regression calculation; UIQM is based on the three indicators of color, clarity and contrast. The weight coefficient of each indicator is obtained by machine learning regression of training image data. UIQM is similar to UCIQE, but pays more attention to the overall color and clarity of an image; CCF is a weighted fusion of color index, contrast index and fog concentration index. Due to light absorption, using color index to evaluate underwater images will cause color deviation problems. Due to light scattering, using contrast index and fog concentration index to evaluate underwater images will cause blurring problems; NUIQ is a non-parametric underwater image quality measurement method based on the first real underwater image quality evaluation dataset SAUD. It extracts quality perception features from the chromaticity and brightness components of the image respectively, and then predicts by training the SVMrank model. This method has problems of color deviation, unsatisfactory brightness, image blurring and detail loss in the evaluated image.
[0004] In summary, the existing underwater image quality evaluation indicators are based on manual feature extraction and regression. They only focus on the distortion (color cast and blur) originally existing in the underwater image, but ignore the distortions such as over-enhancement and structural loss that may exist in the enhanced image. These distortions are introduced by various image enhancement algorithms and also seriously affect the quality of the underwater enhanced image. Therefore, the existing underwater image quality evaluation indicators have a poor correlation with the human visual perception system and cannot directly reflect the subjective visual quality of the image. Therefore, in the field of underwater image enhancement, researchers still use subjective quality evaluation to measure the visual perception quality of an enhanced image. However, this method consumes a lot of manpower and material resources, and is particularly weak when facing a large number of image libraries. Therefore, it is urgent to design a quality evaluation method for underwater enhanced images to further improve the accuracy of underwater enhanced image quality evaluation. Summary of the invention
[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method, device and medium for underwater enhanced image quality evaluation based on prior guidance, so as to further improve the accuracy of underwater enhanced image quality evaluation.
[0006] The purpose of the present invention can be achieved by the following technical solutions:
[0007] The present invention provides a method for evaluating underwater enhanced image quality based on priori guidance, comprising the following steps:
[0008] Obtain an underwater image without enhancement and an underwater image after enhancement, input them into an underwater enhanced image quality evaluation network based on prior guidance, and obtain a quality score of the underwater enhanced image;
[0009] The underwater enhanced image quality assessment network based on prior guidance includes a prior estimation module, a prior guidance module and a quality assessment module connected in sequence, wherein the prior estimation module is used to extract prior information based on the unenhanced underwater image, wherein the prior information includes a two-dimensional depth map and a one-dimensional prior parameter, wherein the one-dimensional prior parameter includes an absorption coefficient a c (λ), scattering coefficient b c (λ), the distance between the maximum and minimum objects and the water surface D(x), and the background light
[0010] The prior guidance module is used to combine the prior features extracted based on the prior information with the convolution features extracted based on the enhanced underwater image. e (x) is fused to obtain the fusion feature f e ′(x);
[0011] The quality assessment module is used to evaluate the quality of the underwater image and the fusion feature f e′(x) predicts the quality score of the enhanced underwater image.
[0012] Furthermore, the prior estimation module includes a first sub-network and a second sub-network in parallel, the first sub-network is a UNet network composed of multiple residual dense blocks, used to extract the two-dimensional depth map; the second sub-network is a DenseNet network composed of multiple dense blocks, used to extract the one-dimensional prior parameters.
[0013] Furthermore, the specific structure of the first sub-network is as follows: the first sub-network includes a structurally symmetrical encoder and decoder, the encoder includes multiple cascaded residual dense blocks, and each residual dense block in the encoder is jump-connected with the residual dense block of the same level in the decoder.
[0014] Furthermore, the residual dense block includes a plurality of convolution units connected in sequence and a fusion unit, each convolution unit includes a convolution layer, a normalization layer and a ReLU activation function for feature extraction, the fusion unit includes a cascade and a convolution layer for feature fusion, dense connections are introduced between the convolution units, and residual connections are introduced after the fusion unit.
[0015] Furthermore, the specific structure of the second sub-network is as follows: the second sub-network includes a plurality of dense blocks connected in sequence, and different one-dimensional prior parameters are output through dense blocks of different levels.
[0016] Furthermore, the prior features include the features of the two-dimensional depth map f d (x) and the characteristic sum of all one-dimensional prior parameters f p (x), in the a priori guidance module, first, f p (x) and the convolution feature f e (x) phase cascade, obtain f e,p (x); then, f d (x), f p (x) and f e (x) The global maximum pooling layer is input together for feature dimension reduction, and the channel coefficient λ is calculated through the fully connected layer network. e,p (x); Finally, based on λ e,p (x) and f e,p (x) Output fusion feature f e ′(x), the fusion feature f e The specific expression of ′(x) is as follows: e ′(x)=f e,p (x)*λ e,p (x)+f e,p (x).
[0017] Furthermore, the quality assessment module includes a plurality of dense blocks connected in sequence and a fully connected layer, and the fusion features f extracted by the prior guidance module are respectively introduced after the last two dense blocks. e ′(x).
[0018] Furthermore, the underwater enhanced image quality assessment network based on prior guidance is trained by a loss function, and the expression of the loss function is specifically as follows:
[0019] L1=||Q out -Q label ||
[0020] Among them, L1 is the loss function, Q out is the quality score of the enhanced underwater image predicted by the network, Q label is the subjective quality label MOS of the image.
[0021] The present invention also provides an electronic device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the above method when executing the program.
[0022] The present invention also provides a computer-readable storage medium on which a computer program is stored, and the program implements the above method when executed by a processor.
[0023] Compared with the prior art, the present invention has the following beneficial effects:
[0024] 1. The present invention proposes a method for evaluating the quality of underwater enhanced images based on prior guidance, which is implemented by an underwater enhanced image quality evaluation network based on prior guidance. The network includes a prior estimation module, a prior guidance module and a quality assessment module connected in sequence, wherein the prior estimation module is used to extract prior information based on the unenhanced underwater image, including a two-dimensional depth map and a one-dimensional prior parameter, and the one-dimensional prior parameter includes an absorption coefficient a c (λ), scattering coefficient b c (λ), the distance between the maximum and minimum objects and the water surface D(x) and the background light B c ∞ These prior information jointly determine the degree of color deviation and blur in an underwater image; the prior guidance module is used to fuse the prior features extracted based on the prior information with the convolution features extracted based on the enhanced underwater image to obtain the fused features; the quality assessment module is used to predict the quality score of the enhanced underwater image based on the enhanced underwater image and the fused features. The features of the original image can be introduced through the fused features, which can make the quality assessment network pay more attention to the content and distortion of the image, and further improve the accuracy of the quality assessment of the underwater enhanced image.
[0025] 2. The first sub-network of the present invention is constructed based on the U-Net structure, with relatively small computational complexity and strong generalization ability. It can extract and combine information from different receptive fields, better integrate global information and local information, and more accurately estimate the two-dimensional depth map. Residual connections and dense connections are introduced in the residual dense blocks of the first sub-network at the same time, where residual connections can solve the degradation problem caused by the increase in network depth, and dense connections can realize feature reuse and improve training efficiency. The second sub-network outputs different one-dimensional prior parameters through dense blocks of different levels, which can further improve the guiding role of one-dimensional prior parameters in subsequent quality evaluation.
[0026] 3. The prior guidance module of the present invention introduces underwater imaging priors into the quality assessment module through alternating guidance of two-dimensional depth map features and one-dimensional prior parameter features. This unique combination imitates the way in which prior parameters are combined in the physical imaging process of the original underwater image, which is more conducive to the network imitating subjective perception rules for quality assessment, thereby improving the accuracy of the quality evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 The overall structure diagram of the PGUIQA network based on prior-guided underwater enhanced image quality assessment is shown in Figure 2.
[0028] Among them, I R (x) and I(x) are the unenhanced underwater image and the enhanced underwater image, respectively. Sub-RDUNet and Sub-DN are the first sub-network and the second sub-network, respectively. d(x) is the two-dimensional depth map, λ n (x) is a one-dimensional prior parameter, a c (λ) is the absorption coefficient, b c (λ) is the scattering coefficient, D(x) is the distance between the largest and smallest objects and the water surface, is the background light, f d (x) is the feature of the two-dimensional depth map, f p (x) is the characteristic sum of all one-dimensional prior parameters;
[0029] Figure 2 The overall structure diagram of the PGUIQA network based on prior-guided underwater enhanced image quality assessment;
[0030] Figure 3 Schematic diagram of the structure of the residual dense block;
[0031] Figure 4 Schematic diagram of the structure of the prior guidance module. DETAILED DESCRIPTION
[0032] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0033] Example:
[0034] Underwater images follow the underwater physical imaging model. Various water qualities, different color deviations, and different blur levels can all be reflected by different values of the coefficients in the model. All parameters that appear in the underwater physical imaging model, including the distance from the water surface to the object (vertical distance), the distance from the camera to the object (horizontal distance), background light, absorption coefficient, and scattering coefficient, jointly determine the degree of color deviation and blur in an underwater image. This embodiment provides a method for underwater enhanced image quality assessment (Prior-guided Underwater Enhanced Image Quality Assessment, PGUIQA) based on prior guidance. The method uses underwater imaging prior information such as depth maps and other parameters that appear in underwater physical imaging models to perform underwater image quality assessment tasks. The specific steps are as follows: Obtain an unenhanced underwater image I R (x) and the enhanced underwater image I(x) are input into the underwater enhanced image quality evaluation network based on prior guidance to obtain the quality score of the underwater enhanced image.
[0035] like Figure 1 As shown in FIG. 1 , the underwater enhanced image quality assessment network based on prior guidance includes a prior estimation module (i.e. Figure 1 The underwater prior estimation module in the Figure 1 The prior estimation module is used to extract prior information based on the unenhanced underwater image. The prior estimation module includes a first sub-network Sub-RDUNet and a second sub-network Sub-DN in parallel. The first sub-network is used to extract a two-dimensional depth map d(x), that is, the distance from the object to the camera. The second sub-network is used to extract a one-dimensional prior parameter, and the one-dimensional prior parameter includes an absorption coefficient a c (λ), scattering coefficient b c (λ), the distance between the maximum and minimum objects and the water surface D(x), and the background light The prior guidance module is used to combine the prior features extracted based on the prior information with the convolutional features extracted based on the enhanced underwater image. e (x) is fused to obtain the fusion feature f e ′(x), the prior features include the features of the two-dimensional depth map f d (x) and the characteristic sum of all one-dimensional prior parameters f p(x). The quality assessment module is used to evaluate the quality of the underwater image and the fusion feature f e ′(x) predicts the quality score of the enhanced underwater image. By introducing prior information in the evaluation process, the above network can pay more attention to the content and distortion of the image, thereby more accurately evaluating the quality of the enhanced underwater image.
[0036] The detailed structure of the network is as follows Figure 2 As shown. The prior information is divided into two categories according to the dimension, one is two-dimensional parameters (i.e., depth map), and the other is one-dimensional parameters (maximum and minimum vertical distances, background light, absorption parameters, scattering parameters). Accordingly, the present invention designs different networks to estimate these two types of parameters respectively. For two-dimensional parameters, the present invention designs a first sub-network, which is a UNet network composed of multiple residual dense blocks. The specific structure is as follows: the first sub-network includes a symmetrically structured encoder and decoder, the encoder includes three cascaded residual dense blocks (RDBs) and three down-samplings, the decoder includes three cascaded residual dense blocks (RDBs) and three up-samplings, and each residual dense block in the encoder is jump-connected to the residual dense blocks of the same level in the decoder, so that the network can extract and combine information from different receptive fields, better fuse global information and local information, and more accurately estimate the depth map in the prior parameters. As shown Figure 3 As shown in the figure, the residual dense block includes multiple convolution units connected in sequence and a fusion unit. Each convolution unit includes a convolution layer, a normalization layer and a ReLU activation function for feature extraction. The fusion unit includes a cascade and a convolution layer for feature fusion. Dense connections are introduced between each convolution unit, and residual connections are introduced after the fusion unit. The residual dense block inherits the characteristics of the residual network (ResNet) and the dense network (DenseNet), and has the advantages of fewer parameters, feature reuse, easier training, and preventing gradient disappearance during training.
[0037] For the one-dimensional parameters, the present invention designs a second subnetwork, which is a DenseNet-121 network composed of four dense blocks. In particular, different one-dimensional prior parameters are output through dense blocks at different levels, and the specific settings are as follows: after the third dense block, the absorption coefficient a is estimated through the fully connected layer c (λ) and scattering coefficient b c (λ), after the fourth dense block, the distance D(x) between the maximum and minimum objects and the water surface and the background light are estimated through the fully connected layer This setting is because the overall performance of outputting parameters through dense blocks in different layers is better than the performance of outputting all parameters in dense blocks in the same layer.
[0038] The prior guidance module is used to extract prior features from prior information to assist in subsequent image quality assessment. According to the underwater physical imaging model, there are:
[0039]
[0040] Among them, I c (x) represents the overall signal of point x captured by the camera, Expressed as the initial energy at point x, Item and All items are calculations between one-dimensional prior parameters, which can be converted to features using f p (x) indicates that Among the terms, β c (λ) represents the attenuation coefficient (related to the wavelength λ), d(x) represents the two-dimensional depth map, After multiplying the term by the term in parentheses, it can be expressed as f d (x)*f p (x). Finally, the entire model can be approximately expressed as follows when converted to the feature level:
[0041] f raw (x) = f d (x)*f p (x)+f p (x)
[0042] Among them, f raq (x) represents all the prior features that can be extracted from the original image and added to the quality assessment module, f d (x) represents the features of the 2D depth map, f p (x) represents the sum of features of all one-dimensional prior parameters.
[0043] The structure of the prior guidance module is as follows Figure 4 As shown, first, f p (x) and the convolution feature f e (x) phase cascade, obtain f e,p (x); then, f d (x), f p (x) and f e (x) The global maximum pooling layer is used to reduce the feature dimension and then sent to a series of fully connected layer networks to calculate the channel coefficient λ. e,p (x); Finally, based on λ e,p (x) and f e,p (x) Output fusion feature f e ′(x), fusion feature f e The specific expression of ′(x) is as follows:
[0044] f e ′(x)=f e,p (x)*λ e,p (x)+f e,p (x).
[0045] Considering that the DenseNet-121 model requires fewer parameters and presents better performance, the quality assessment module is still built based on this model. The quality assessment module consists of four dense blocks (DB) connected in sequence and a fully connected layer (FC). The fusion features f extracted by the prior guidance module are inserted after the third and fourth dense blocks. e ′(x). The fusion feature introduces prior information into the quality assessment. Although this prior information is extracted from the real underwater image without enhancement, it is still helpful for the quality assessment of the corresponding enhanced image. On the one hand, the enhanced image itself is enhanced from the real underwater image, so learning the parameter information in the original image can help the network judge the rationality of the enhancement result; on the other hand, when people judge the quality of underwater image enhancement, they usually compare it with the original image. Therefore, learning the parameters in the original image is in line with people's subjective prediction habits. Combining the above two reasons, learning the prior features of the original image is more conducive to the network imitating people's subjective perception rules, thereby improving the accuracy of quality evaluation.
[0046] The underwater enhanced image quality evaluation network based on prior guidance is trained through the loss function. The expression of the loss function L1 is as follows:
[0047] L1=||Q out -Q label ||
[0048] Among them, Q out is the quality score of the enhanced underwater image predicted by the network, Q label is the subjective quality label MOS of the image.
[0049] To verify the effectiveness of the above method, this example first synthesized an underwater image dataset, which contains images of 10 water quality types (I, IA, IB, II, III, 1C, 3C, 5C, 7C, 9C, where I is the clearest water quality and 9C is the most turbid). For each image of each water quality, there is a corresponding fixed prior parameter label (absorption coefficient a c (λ), scattering coefficient b c (λ)) is used for training in deep learning. This dataset provides 1,000 color images of the atmosphere with different contents and their depth maps d(x). During training, the appropriate range of prior information is randomly generated for 10 types of water quality, including D(x), d(x), A large number of random training samples and labels can be generated. As an artificially synthesized dataset, it is only used for training the underwater prior estimation network.
[0050] After the training is completed, the method of the present invention is compared with a variety of existing methods on the SAUD dataset. The SAUD dataset is the only publicly available real underwater image quality evaluation dataset, in which all original images are collected in reality. The dataset contains a total of 100 original images, which are enhanced by 10 underwater image enhancement algorithms, and have a total of 1000 enhanced images. All images are labeled with corresponding quality scores. The existing methods compared include UIQM algorithm, UCIQE algorithm, NUIQ algorithm, CCF algorithm, B-FEN algorithm, ResNet-50 network and DenseNet-121 network. These quality evaluation methods are trained and tested separately on each dataset, that is, trained on the training set in the SAUD dataset and tested on the test set in the SAUD dataset. Among them, the physics-based method will refit the coefficients of each feature in the training set by machine learning. According to the commonly used evaluation annotations for image quality evaluation, the PLCC, SROCC, KROCC, RMSE, and PWRC indicators between the network calculation results and the label MOS scores are compared, and the results are shown in Table 1.
[0051] Table 1 Performance of various quality evaluation methods on the SAUD dataset
[0052]
[0053] Table 1 shows the shortcomings of traditional physical algorithms and the advantages of deep learning methods. Physics-based underwater image quality evaluation methods (such as UCIQE, UIQM, CCF and NUIQ) are not as good as deep learning-based methods (such as ResNet-50, DenseNet-121, B-FEN and the method PGUIQA proposed in this invention). This result reflects that deep learning as a tool has great development prospects in the field of underwater image quality evaluation. As the latest publicly available physics-based underwater image evaluation method, NUIQ has all performances that exceed those of other physics-based algorithms. Compared with NUIQ, the method PGUIQA of this invention has a good improvement in all indicators. The indicators of the NUIQ algorithm on the SAUD dataset are 0.6710 (PLCC), 0.6735 (SROCC), 0.4211 (KROCC) and 18.12 (RMSE); while the indicators of the PGUIQA algorithm on the SAUD dataset are 0.8356 (PLCC), 0.8347 (SROCC), 0.6479 (KROCC) and 11.25 (RMSE). At the same time, PGUIQA also leads other deep learning-based methods in various indicators, which proves the superiority of this algorithm in performance.
[0054] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0055] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.
Claims
1. A method for evaluating underwater enhanced image quality based on prior guidance, characterized in that: The following steps are involved: Obtain an underwater image without enhancement and an underwater image after enhancement, input them into an underwater enhanced image quality evaluation network based on prior guidance, and obtain a quality score of the underwater enhanced image; The underwater enhanced image quality assessment network based on prior guidance includes a prior estimation module, a prior guidance module and a quality assessment module connected in sequence, wherein the prior estimation module is used to extract prior information based on the unenhanced underwater image, wherein the prior information includes a two-dimensional depth map and a one-dimensional prior parameter, wherein the one-dimensional prior parameter includes an absorption coefficient a c (λ), scattering coefficient b c (λ), the distance between the maximum and minimum objects and the water surface D(x), and the background light The prior guidance module is used to combine the prior features extracted based on the prior information with the convolution features extracted based on the enhanced underwater image. e (x) is fused to obtain the fusion feature f e ′ (x); The quality assessment module is used to evaluate the quality of the underwater image and the fusion feature f e ′ (x) Predict the quality score of the enhanced underwater image.
2. The underwater enhanced image quality assessment method based on prior guidance according to claim 1 is characterized in that: The prior estimation module includes a first sub-network and a second sub-network in parallel, wherein the first sub-network is a UNet network composed of multiple residual dense blocks, and is used to extract the two-dimensional depth map; the second sub-network is a DenseNet network composed of multiple dense blocks, and is used to extract the one-dimensional prior parameters.
3. The method for evaluating underwater enhanced image quality based on prior guidance according to claim 2, characterized in that: The specific structure of the first sub-network is as follows: the first sub-network includes a structurally symmetrical encoder and decoder, the encoder includes a plurality of cascaded residual dense blocks, and each residual dense block in the encoder is jump-connected with a residual dense block of the same level in the decoder.
4. The method for evaluating underwater enhanced image quality based on prior guidance according to claim 2, characterized in that: The residual dense block includes a plurality of convolution units connected in sequence and a fusion unit, each convolution unit includes a convolution layer, a normalization layer and a ReLU activation function for feature extraction, the fusion unit includes a cascade and a convolution layer for feature fusion, a dense connection is introduced between the convolution units, and a residual connection is introduced after the fusion unit.
5. The method for evaluating underwater enhanced image quality based on prior guidance according to claim 2, characterized in that: The specific structure of the second sub-network is as follows: the second sub-network includes a plurality of dense blocks connected in sequence, and different one-dimensional prior parameters are output through dense blocks of different levels.
6. The method for evaluating underwater enhanced image quality based on prior guidance according to claim 1, characterized in that: The prior features include the features of the two-dimensional depth map f d (x) and the characteristic sum of all one-dimensional prior parameters f p (x), in the a priori guidance module, first, f p (x) and the convolution feature f e (x) phase cascade, obtain f e,p (x); then, f d (x), f p (x) and f e (x) The global maximum pooling layer is input together to reduce the feature dimension, and the channel coefficient λ is calculated through the fully connected layer network. e,p (x); Finally, based on λ e,p (x) and f e,p (x) Output fusion feature f′ e (x), the fusion feature f e ′ The expression of (x) is as follows: e ′(x)=f e,p (x)*λ e,p (x)+f e,p (x).
7. The method for evaluating underwater enhanced image quality based on prior guidance according to claim 1, characterized in that: The quality assessment module includes a plurality of dense blocks connected in sequence and a fully connected layer, and the fusion features f extracted by the prior guidance module are respectively introduced after the last two dense blocks. e ′(x).
8. The method for evaluating underwater enhanced image quality based on prior guidance according to claim 1, characterized in that: The underwater enhanced image quality assessment network based on prior guidance is trained by a loss function, and the expression of the loss function is as follows: L1=||Q out -Q label || Among them, L1 is the loss function, Q out is the quality score of the enhanced underwater image predicted by the network, Q label is the subjective quality label MOS of the image.
9. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.