An underwater image compressive sensing method and system based on semantic graph guidance
By employing a semantic graph-guided and graph network-assisted underwater image compression sensing method, which combines underwater physical priors and semantic information for sampling and reconstruction, the problems of underwater image blurring and detail loss are solved, achieving high-quality underwater image reconstruction suitable for underwater exploration and surveying.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MACAO POLYTECHNIC INST
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-05
AI Technical Summary
Existing compressed sensing methods suffer from problems such as blurring, distortion, and loss of detail in underwater image acquisition, especially with insufficient reconstruction performance at extremely low sampling rates, which affects the accuracy of underwater detection and exploration.
We employ a semantic graph-guided differential sampling and graph network-assisted reconstruction strategy, combining underwater physical priors and semantic information for sampling. We utilize graph neural networks to mine the complementarity of features within the same semantic meaning, and use UNet networks for feature fusion and GAN training to improve reconstruction quality.
Achieving high-quality underwater image reconstruction at extremely low sampling rates alleviates blurring and detail loss problems, improves the efficiency and generalization ability of the algorithm, and is suitable for underwater exploration and surveying tasks.
Smart Images
Figure CN122156330A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of underwater image processing and deep learning technology, and in particular to an underwater image compression sensing method and system based on semantic graph guidance. Background Technology
[0002] With the increasing demand for marine resource development, underwater exploration and detection have received widespread attention. Underwater image acquisition mainly relies on remotely operated vehicles (ROVs) or autonomous underwater vehicles (AUVs), but the narrow bandwidth of underwater transmission channels and limited hardware performance lead to problems such as blurriness and distortion in the acquired underwater images. Compressed sensing (CS) theory points out that if a signal is sparsely represented in a specific domain, complete signal reconstruction can be achieved through sampling at a rate far lower than the Nyquist sampling rate, providing a feasible solution for efficient acquisition and transmission of underwater images.
[0003] In recent years, deep learning-based reconstruction (CS) methods have achieved good results in land-based imagery. However, underwater images differ significantly from land-based images: underwater images have less semantic meaning, and features within the same semantic meaning have high similarity; due to the influence of light scattering and absorption, the degree of degradation of different semantic meanings is related to the imaging distance. Existing CS methods have not incorporated these characteristics into their algorithm design, resulting in significant room for improvement in reconstruction performance when applied to underwater images. Especially at the extremely low sampling rates commonly used in underwater scenes (such as 0.01, 0.04, etc.), existing CS algorithms often suffer from severe block artifacts, structural blurring, and even loss of detail, which seriously affects the accuracy of subsequent tasks. The inability to obtain high-quality reconstructed images at extremely low sampling rates is one of the bottleneck problems currently faced in underwater exploration and surveying. Summary of the Invention
[0004] The purpose of this application is to provide a semantic graph-guided underwater image compression sensing method and system to solve the above-mentioned technical problems, aiming to achieve high-quality underwater image sampling and reconstruction.
[0005] In some embodiments of this application, a semantic graph-guided differential sampling and graph network-assisted reconstruction strategy is employed. During the sampling phase, by combining underwater physical priors and semantic information, sampling points are concentrated in semantic regions with high importance and low degradation, preserving key information. During the reconstruction phase, a graph neural network is used to mine the complementarity of features within the same semantic domain, capturing long-distance feature correlations. All features are then fused using a UNet network, and end-to-end network training is performed in a GAN manner to improve the detail accuracy and overall quality of the reconstructed image, achieving high-quality underwater image reconstruction at extremely low sampling rates.
[0006] In some embodiments of this application, the structure of GAN is introduced into the compressed sensing task to supplement details, improve the realism of the image, effectively alleviate the problems of blurring, distortion and loss of details, and at the same time take into account the algorithm efficiency and generalization ability, so as to achieve a high degree of adaptability to the actual application scenarios of underwater detection and exploration.
[0007] In some embodiments of this application, a semantic graph-guided underwater image compression sensing method is provided, including:
[0008] Generate a preprocessed dataset based on the input image and the preprocessing model; A sampling network is constructed, and a sampling matrix is set based on the sampling network and the preprocessed dataset. Underwater images are then acquired based on the sampling matrix. Construct a reconstruction network and output reconstruction results based on the reconstruction network and underwater acquired images; The sampling network includes: Feature analysis model and sampling matrix generation model In some embodiments of this application, generating the preprocessed dataset includes: Establish a preprocessing model; The preprocessing model includes: a semantic segmentation subnetwork and a prior estimation subnetwork; Input the input image into the semantic segmentation network, and output semantic graph groups; The semantic graph group includes semantic graphs of multiple semantics; The input image is fed into the prior estimation subnetwork, which outputs background light B and depth map D. A preprocessed dataset is generated, which includes: a semantic graph group, a background light B, and a depth map D.
[0009] In some embodiments of this application, the setting of the sampling matrix includes: Establish a feature analysis model; The feature analysis model includes: a prior modeling model and a semantic weighting model; The prior modeling model adjusts its parameters based on the background light B and the depth map D output features. and ; The features of the image to be sampled are corrected according to the preset imaging model; The preprocessed dataset semantic weight model generates a weight mask set based on the semantic graph group and the depth graph D; The sampling matrix generation model sets the sampling matrix based on the weight mask set.
[0010] In some embodiments of this application, the preset imaging model includes: ; in, It is an underwater image to be sampled, affected by underwater imaging degradation. It contains no degraded structural or detailed information. It's a depth map. It is background light;
[0011]
[0012] The above model is modified as follows: in, and It is two through and The learned features are used to adjust the parameters.
[0013] In some embodiments of this application, the semantic weight model includes:
[0014] in, Let be the semantic graph of the i-th semantic element in the semantic graph group. The learnable weights are generated by the school based on the depth information for the i-th semantic.
[0015] In some embodiments of this application, the output reconstruction result includes: Construct and reconstruct the network; The reconstruction network includes: a feature matching network and a UNet reconstruction network; The feature matching network includes eight parallel semantic processing branches and one structural branch. Feature matching networks perform long-distance modeling of features corresponding to each semantic in a semantic graph group; Graph network formula:
[0016] Where X is the input feature matrix, A is the adjacency matrix of the features, and W is the learnable weight matrix; Generated by two consecutive graph convolutions:
[0017] Where A is the input feature. The adjacency matrix is given by W, which is a matrix reflecting the correlation between the current feature and other features. The reconstruction results are output based on the network reconstruction and modeling results from UNet.
[0018] In some embodiments of this application, a semantic graph-guided underwater image compression sensing system is provided, comprising: The preprocessing unit is used to generate a preprocessed dataset based on the input image and the preprocessing model; Sampling units are used to establish a sampling network; The sampling network includes a feature analyzer and a matrix generator, wherein the feature analyzer contains a feature analysis model and the matrix generator contains a sampling matrix generation model. The sampling unit is used to set a sampling matrix based on the sampling network and the preprocessed dataset, and to acquire underwater images based on the sampling matrix. Reconstruction units are used to build reconstruction networks; The reconstruction unit is used to output reconstruction results based on the reconstruction network and underwater acquired images.
[0019] In some embodiments of this application, the preprocessing unit is further configured to: Establish a preprocessing model; The preprocessing model includes: a semantic segmentation subnetwork and a prior estimation subnetwork; Input the input image into the semantic segmentation network, and output semantic graph groups; The semantic graph group includes semantic graphs of multiple semantics; The input image is fed into the prior estimation subnetwork, which outputs background light B and depth map D. A preprocessed dataset is generated, which includes: a semantic graph group, a background light B, and a depth map D.
[0020] In some embodiments of this application, the sampling unit is further configured to: Establish a feature analysis model; The feature analysis model includes: a prior modeling model and a semantic weighting model; The prior modeling model adjusts its parameters based on the background light B and the depth map D output features. and ; The features of the image to be sampled are corrected according to the preset imaging model; The preprocessed dataset semantic weight model generates a weight mask set based on the semantic graph group and the depth graph D; The sampling matrix generation model sets the sampling matrix based on the weight mask set; The preset imaging model includes: ; in, It is an underwater image to be sampled, affected by underwater imaging degradation. It contains no degraded structural or detailed information. It's a depth map. It is background light;
[0021]
[0022] The above model is modified as follows: in, and It is two through and The learned features are used to adjust the parameters; The semantic weight model includes:
[0023] in, Let be the semantic graph of the i-th semantic element in the semantic graph group. The learnable weights are generated by the school based on the depth information for the i-th semantic.
[0024] In some embodiments of this application, the reconstruction unit is further configured to: Construct and reconstruct the network; The reconstruction network includes: a feature matching network and a UNet reconstruction network; The feature matching network includes eight parallel semantic processing branches and one structural branch. Feature matching networks perform long-distance modeling of features corresponding to each semantic in a semantic graph group; Graph network formula:
[0025] Where X is the input feature matrix, A is the adjacency matrix of the features, and W is the learnable weight matrix; Generated by two consecutive graph convolutions:
[0026] Where A is the input feature. The adjacency matrix is given by W, which is a matrix reflecting the correlation between the current feature and other features. The reconstruction results are output based on the network reconstruction and modeling results from UNet.
[0027] Compared with existing technologies, the underwater image compression sensing method and system based on semantic graph guidance proposed in this application have the following advantages: This study employs a semantic graph-guided differential sampling and graph network-assisted reconstruction strategy. In the sampling phase, by combining underwater physical priors and semantic information, sampling points are concentrated in semantic regions with high importance and low degradation, preserving key information. In the reconstruction phase, graph neural networks are used to mine the complementarity of features within the same semantic domain, capturing long-distance feature correlations. All features are then fused using a UNet network and trained end-to-end using a GAN approach, improving the detail accuracy and overall quality of the reconstructed image and achieving high-quality underwater image reconstruction at extremely low sampling rates.
[0028] Introducing the GAN structure into compressed sensing tasks to supplement details enhances image realism and effectively alleviates problems such as blurring, distortion, and loss of detail. It also balances algorithm efficiency and generalization ability, achieving a high degree of adaptability to practical application scenarios of underwater detection and exploration. Attached Figure Description
[0029] Figure 1 This is a flowchart illustrating a semantic graph-guided underwater image compression sensing method in a preferred embodiment of this application. Detailed Implementation
[0030] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.
[0031] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.
[0032] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.
[0033] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.
[0034] like Figure 1 As shown, a preferred embodiment of this application provides a semantic graph-guided underwater image compression sensing method, comprising: S101: Generate a preprocessed dataset based on the input image and the preprocessing model; S102: Construct a sampling network, set a sampling matrix based on the sampling network and the preprocessed dataset, and acquire underwater images based on the sampling matrix; S103: Construct a reconstruction network and output reconstruction results based on the reconstruction network and underwater acquired images; The sampling network includes: Feature analysis model and sampling matrix generation model.
[0035] Specifically, generating a preprocessed dataset includes: Establish a preprocessing model; The preprocessing model includes: a semantic segmentation subnetwork and a prior estimation subnetwork; Input the input image into the semantic segmentation network, and output semantic graph groups; A semantic graph group contains semantic graphs with multiple semantic meanings; The input image is fed into the prior estimation subnetwork, which outputs background light B and depth map D. Generate a preprocessed dataset, which includes: a semantic graph set, background light B, and a depth map D.
[0036] Specifically, the semantic segmentation network in this application is preferably the segmentation network corresponding to the SUIM dataset. In actual deployment, it can also be replaced with any other semantic segmentation network; the prior estimation subnetwork is a two-stage UNet that acquires background light and depth maps respectively, and in actual deployment, it can also be replaced with other underwater physical prior estimation networks.
[0037] Specifically, a SUIM dataset is established, which includes multiple sets of labeled samples of underwater image-semantic map pairings. All input images and corresponding semantic maps of the SUIM dataset are adjusted to an input size of 256×256, and input images are generated based on the adjustment results.
[0038] Specifically, the input image is fed into the preprocessing model to obtain semantic map groups, background light and depth maps. Using this information and the original image (i.e. the image to be sampled), the network parameters of the sampling network and the reconstruction network are trained through joint optimization of multiple loss functions.
[0039] Specifically, underwater image acquisition data is obtained according to the acquisition sub-strategy.
[0040] Understandably, in the above embodiments, the semantic graph-guided differentiated sampling and graph network-assisted reconstruction strategy, during the sampling phase, combines underwater physical priors and semantic information to concentrate sampling points in semantic regions with high importance and low degradation, thus preserving key information. During the reconstruction phase, graph neural networks are used to mine the complementarity of features within the same semantic domain, capturing long-distance feature correlations. This achieves high-quality underwater image reconstruction at extremely low sampling rates.
[0041] In a preferred embodiment of this application, a sampling matrix is set, including: Establish a feature analysis model; Feature analysis models include: prior modeling models and semantic weighting models; The prior modeling model adjusts its parameters based on the background light B and the depth map D output features. and ; The features of the image to be sampled are corrected according to the preset imaging model; The preprocessed dataset semantic weight model generates a weight mask set based on the semantic graph group and the depth graph D; The sampling matrix generation model sets the sampling matrix based on the weight mask set.
[0042] Specifically, the background light B and depth map D are input into the prior modeling model, and feature adjustment parameters are generated through learning from the background light and depth map. and Based on the underwater image imaging model, the features of the sampled image are adjusted and enhanced to eliminate blurring and color shift caused by underwater degradation.
[0043] Specifically, the preset imaging models include: ; in, It is an underwater image to be sampled, affected by underwater imaging degradation. It contains no degraded structural or detailed information. It's a depth map. It is background light;
[0044]
[0045] The above model is modified as follows: in, and It is two through and The learned features are used to adjust the parameters.
[0046] Specifically, D(x) and B(x) are the priors that have already been obtained.
[0047] Specifically, the adjustment parameters generated by the preprocessing model are used to adjust the input features, thereby obtaining structural and detailed information without degradation.
[0048] Specifically, semantic weighting models include:
[0049] in, Let be the semantic graph of the i-th semantic element in the semantic graph group. The learnable weights are generated by the school based on the depth information for the i-th semantic.
[0050] Specifically, the semantic weight model analyzes the relationship between each semantic term and depth based on neural networks, generating a corresponding weight mask for each semantic term in the semantic graph group, thus generating a weight mask set. According to the undistorted formula in the above imaging model, underwater images follow the characteristic that the farther the imaging distance, the more blurred they become. Therefore, the weight mask reflects the relationship between each semantic term and structure in the current input, that is, the degree of attention that should be received during sampling, so that the sampling focuses on semantic regions with low degradation and high importance.
[0051] Specifically, after the weights of the input features are adjusted, the sampling matrix generation model sorts them according to the priority of each feature, selects the required number of sampling points, generates a sampling matrix, completes the sampling of the input image, and obtains the sampling result.
[0052] It is understandable that in the above embodiments, by combining underwater physical priors and semantic information, sampling points are concentrated in semantic regions with high importance and low degradation, thus preserving key information and achieving differentiated sampling.
[0053] In a preferred embodiment of this application, the output reconstruction result includes: Construct and reconstruct the network; The reconstruction network includes: a feature matching network and a UNet reconstruction network; The feature matching network includes eight parallel semantic processing branches and one structural branch. Feature matching networks perform long-distance modeling of features corresponding to each semantic in a semantic graph group; Graph network formula:
[0054] Where X is the input feature matrix, A is the adjacency matrix of the features, and W is the learnable weight matrix; Generated by two consecutive graph convolutions:
[0055] Where A is the input feature. The adjacency matrix is given by W, which is a matrix reflecting the correlation between the current feature and other features. Output the reconstruction results based on the UNet network reconstruction and modeling results. Specifically, W is a learnable weight matrix that represents the strength of the correlation between each position in the current feature matrix X and other positions. By performing two consecutive graph convolutions, feature extraction is made more comprehensive.
[0056] Specifically, by establishing a reconstruction network, the similarity of features under the same semantic meaning in underwater images is fully utilized to achieve feature complementarity and improve reconstruction quality. Each semantic meaning is processed by the weight mask obtained from the semantic graph to process the features under the corresponding semantic meaning. The graph neural network is used to perform long-distance modeling of the features under the current semantic meaning, fully explore the long-distance feature dependencies, and achieve complementarity.
[0057] Specifically, by reconstructing the feature aggregation of the network, the sampled features are used to supplement the unsampled features. All subsequent features are then fused through the UNet network, and the network is trained end-to-end in a GAN manner to improve the detail accuracy and overall quality of the reconstructed image.
[0058] It is understandable that in the above embodiments, the introduction of GAN structure into the compressed sensing task to supplement details improves the realism of the image, effectively alleviates the problems of blurring, distortion and loss of details, and at the same time takes into account the algorithm efficiency and generalization ability, so as to achieve a high degree of adaptability to the actual application scenarios of underwater detection and exploration.
[0059] Based on any of the above preferred embodiments, another preferred embodiment of the semantic graph-guided underwater image compressed sensing method provides a semantic graph-guided underwater image compressed sensing system, comprising: The preprocessing unit is used to generate a preprocessed dataset based on the input image and the preprocessing model; Sampling units are used to establish a sampling network; The sampling network includes a feature analyzer and a matrix generator. The feature analyzer contains a feature analysis model, and the matrix generator contains a sampling matrix generation model. The sampling unit is used to set the sampling matrix according to the sampling network and the preprocessed dataset, and to acquire underwater images according to the sampling matrix; Reconstruction units are used to build reconstruction networks; The reconstruction unit is used to output reconstruction results based on the reconstruction network and underwater acquired images.
[0060] In a preferred embodiment of this application, the preprocessing unit is further configured to: Establish a preprocessing model; The preprocessing model includes: a semantic segmentation subnetwork and a prior estimation subnetwork; Input the input image into the semantic segmentation network, and output semantic graph groups; A semantic graph group contains semantic graphs with multiple semantic meanings; The input image is fed into the prior estimation subnetwork, which outputs background light B and depth map D. Generate a preprocessed dataset, which includes: a semantic graph set, background light B, and a depth map D.
[0061] In a preferred embodiment of this application, the sampling unit is further used for: Establish a feature analysis model; Feature analysis models include: prior modeling models and semantic weighting models; The prior modeling model adjusts its parameters based on the background light B and the depth map D output features. and ; The features of the image to be sampled are corrected according to the preset imaging model; The preprocessed dataset semantic weight model generates a weight mask set based on the semantic graph group and the depth graph D; The sampling matrix generation model sets the sampling matrix based on the weight mask set; The preset imaging model includes: ; in, It is an underwater image to be sampled, affected by underwater imaging degradation. It contains no degraded structural or detailed information. It's a depth map. It is background light;
[0062]
[0063] The above model is modified as follows: in, and It is two through and The learned features are used to adjust the parameters; Semantic weighting models include:
[0064] in, Let be the semantic graph of the i-th semantic element in the semantic graph group. The learnable weights are generated by the school based on the depth information for the i-th semantic.
[0065] In a preferred embodiment of this application, the reconstruction unit is further configured to: Construct and reconstruct the network; The reconstruction networks include: feature matching networks and UNet reconstruction networks; The feature matching network includes eight parallel semantic processing branches and one structural branch. Feature matching networks perform long-distance modeling of features corresponding to each semantic in a semantic graph group; Graph network formula:
[0066] Where X is the input feature matrix, A is the adjacency matrix of the features, and W is the learnable weight matrix; Generated by two consecutive graph convolutions:
[0067] Where A is the input feature. The adjacency matrix is given by W, which is a matrix reflecting the correlation between the current feature and other features. The reconstruction results are output based on the network reconstruction and modeling results from UNet.
[0068] Based on the first concept of this application, a semantic graph-guided differential sampling and graph network-assisted reconstruction strategy is employed. In the sampling phase, by combining underwater physical priors and semantic information, sampling points are concentrated in semantic regions with high importance and low degradation, preserving key information. In the reconstruction phase, a graph neural network is used to mine the complementarity of features within the same semantic meaning, capturing long-distance feature correlations. All features are then fused using a UNet network and trained end-to-end using a GAN approach, improving the detail accuracy and overall quality of the reconstructed image, achieving high-quality underwater image reconstruction at extremely low sampling rates.
[0069] According to the second concept of this application, the structure of GAN is introduced into the compressed sensing task to supplement details, improve the realism of the image, effectively alleviate the problems of blurring, distortion and loss of details, and at the same time take into account the algorithm efficiency and generalization ability, so as to achieve a high degree of adaptability to the actual application scenarios of underwater detection and exploration.
[0070] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application.
Claims
1. A semantic graph-guided underwater image compressed sensing method, characterized in that, include: Generate a preprocessed dataset based on the input image and the preprocessing model; A sampling network is constructed, and a sampling matrix is set based on the sampling network and the preprocessed dataset. Underwater images are then acquired based on the sampling matrix. Construct a reconstruction network and output reconstruction results based on the reconstruction network and underwater acquired images; The sampling network includes: Feature analysis model and sampling matrix generation model.
2. The underwater image compression sensing method based on semantic graph guidance as described in claim 1, characterized in that, The generation of the preprocessed dataset includes: Establish a preprocessing model; The preprocessing model includes: a semantic segmentation subnetwork and a prior estimation subnetwork; Input the input image into the semantic segmentation network, and output semantic graph groups; The semantic graph group includes semantic graphs of multiple semantics; The input image is fed into the prior estimation subnetwork, which outputs background light B and depth map D. A preprocessed dataset is generated, which includes: a semantic graph group, a background light B, and a depth map D.
3. The underwater image compression sensing method based on semantic graph guidance as described in claim 2, characterized in that, The defined sampling matrix includes: Establish a feature analysis model; The feature analysis model includes: a prior modeling model and a semantic weighting model; The prior modeling model adjusts its parameters based on the background light B and the depth map D output features. and ; The features of the image to be sampled are corrected according to the preset imaging model; The preprocessed dataset semantic weight model generates a weight mask set based on the semantic graph group and the depth graph D; The sampling matrix generation model sets the sampling matrix based on the weight mask set.
4. The underwater image compression sensing method based on semantic graph guidance as described in claim 3, characterized in that, The preset imaging model includes: ; in, It is an underwater image to be sampled, affected by underwater imaging degradation. It contains no degraded structural or detailed information. It's a depth map. It is background light; The above model is modified as follows: in, and It is two through and The learned features are used to adjust the parameters.
5. The underwater image compression sensing method based on semantic graph guidance as described in claim 4, characterized in that, The semantic weight model includes: in, Let be the semantic graph of the i-th semantic element in the semantic graph group. The learnable weights are generated by the school based on the depth information for the i-th semantic.
6. The underwater image compression sensing method based on semantic graph guidance as described in claim 3, characterized in that, The output reconstruction result includes: Construct and reconstruct the network; The reconstruction network includes: a feature matching network and a UNet reconstruction network; The feature matching network includes eight parallel semantic processing branches and one structural branch. Feature matching networks perform long-distance modeling of features corresponding to each semantic in a semantic graph group; Graph network formula: Where X is the input feature matrix, A is the adjacency matrix of the features, and W is the learnable weight matrix; Generated by two consecutive graph convolutions: Where A is the input feature. The adjacency matrix is given by W, which is a matrix reflecting the correlation between the current feature and other features. The reconstruction results are output based on the network reconstruction and modeling results from UNet.
7. A semantic graph-guided underwater image compressed sensing system, employing the semantic graph-guided underwater image compressed sensing method according to any one of claims 1-6, characterized in that, include: The preprocessing unit is used to generate a preprocessed dataset based on the input image and the preprocessing model; Sampling units are used to establish a sampling network; The sampling network includes a feature analyzer and a matrix generator, wherein the feature analyzer contains a feature analysis model and the matrix generator contains a sampling matrix generation model. The sampling unit is used to set a sampling matrix based on the sampling network and the preprocessed dataset, and to acquire underwater images based on the sampling matrix. Reconstruction units are used to build reconstruction networks; The reconstruction unit is used to output reconstruction results based on the reconstruction network and underwater acquired images.
8. The underwater image compression sensing system based on semantic graph guidance as described in claim 7, characterized in that, The preprocessing unit is also used for: Establish a preprocessing model; The preprocessing model includes: a semantic segmentation subnetwork and a prior estimation subnetwork; Input the input image into the semantic segmentation network, and output semantic graph groups; The semantic graph group includes semantic graphs of multiple semantics; The input image is fed into the prior estimation subnetwork, which outputs background light B and depth map D. A preprocessed dataset is generated, which includes: a semantic graph group, a background light B, and a depth map D.
9. The underwater image compression sensing system based on semantic graph guidance as described in claim 8, characterized in that, The sampling unit is also used for: Establish a feature analysis model; The feature analysis model includes: a prior modeling model and a semantic weighting model; The prior modeling model adjusts its parameters based on the background light B and the depth map D output features. and ; The features of the image to be sampled are corrected according to the preset imaging model; The preprocessed dataset semantic weight model generates a weight mask set based on the semantic graph group and the depth graph D; The sampling matrix generation model sets the sampling matrix based on the weight mask set; The preset imaging model includes: ; in, It is an underwater image to be sampled, affected by underwater imaging degradation. It contains no degraded structural or detailed information. It's a depth map. It is background light; The above model is modified as follows: in, and It is two through and The learned features are used to adjust the parameters; The semantic weight model includes: in, Let be the semantic graph of the i-th semantic element in the semantic graph group. The learnable weights are generated by the school based on the depth information for the i-th semantic.
10. The underwater image compression sensing system based on semantic graph guidance as described in claim 9, characterized in that, The reconstruction unit is also used for: Construct and reconstruct the network; The reconstruction network includes: a feature matching network and a UNet reconstruction network; The feature matching network includes eight parallel semantic processing branches and one structural branch. Feature matching networks perform long-distance modeling of features corresponding to each semantic in a semantic graph group; Graph network formula: Where X is the input feature matrix, A is the adjacency matrix of the features, and W is the learnable weight matrix; Generated by two consecutive graph convolutions: Where A is the input feature. The adjacency matrix is given by W, which is a matrix reflecting the correlation between the current feature and other features. The reconstruction results are output based on the network reconstruction and modeling results from UNet.