A three-dimensional gaussian splatting reconstruction method for underwater scenes

Abstract

The application discloses a three-dimensional Gaussian splashing reconstruction method for underwater scenes, and belongs to the technical field of computer vision and three-dimensional reconstruction. The method comprises the following steps: acquiring a monocular video frame sequence of a target underwater scene, a corresponding camera pose sequence and an initial sparse point cloud, initializing a three-dimensional Gaussian point set and learnable physical parameters of an underwater imaging model; in the training process, the reconstruction error is weighted and evaluated based on an opacity weighted multi-view consistency mechanism, the importance score of each Gaussian point is calculated, and a densification operation is performed on the three-dimensional Gaussian point set; a phased freezing strategy is adopted to cooperatively optimize the three-dimensional Gaussian point set and the underwater imaging model parameters; based on the optimized three-dimensional Gaussian point set and the underwater imaging model parameters, rendering and underwater image synthesis are performed on any new view camera pose, and a new view synthesized image is output to represent the reconstruction result. The application improves the geometric compactness, visual fidelity and physical interpretability of the underwater three-dimensional reconstruction result.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and 3D reconstruction technology, and in particular to a 3D Gaussian splash reconstruction method for underwater scenes. Background Technology

[0002] With the development of applications such as underwater robots, marine archaeology, and underwater exploration, there is an urgent need for high-quality 3D reconstruction and novel perspective synthesis technologies for turbid underwater scenes. However, underwater imaging is affected by physical effects such as forward scattering, backscattering, and spectral selective attenuation caused by suspended particles in the water, resulting in problems such as low contrast, color shift, and limited visibility in the acquired images, which pose significant challenges to 3D modeling.

[0003] Existing 3D reconstruction methods mainly include traditional methods based on multi-view geometry and the recently emerging Neural Radiation Field (NeRF) and its extensions. Traditional methods rely on clear texture features and cross-view consistency, making it difficult to obtain reliable matching under low signal-to-noise ratio conditions underwater. While NeRF can achieve high-quality view synthesis, its implicit representation leads to high training overhead and slow convergence speed. Furthermore, it lacks explicit modeling of optical processes in water, making it difficult to guarantee the physical plausibility of the reconstruction results.

[0004] To address the issues of rendering efficiency and explicit representation, the 3D Gaussian Splatting (3DGS) method explicitly models the scene using anisotropic Gaussian ellipsoids and employs differentiable rendering for efficient optimization, achieving good reconstruction results in air. However, when this method is directly applied to underwater scenes, it still faces two key challenges: First, in low-opacity regions, reconstruction residuals caused by backscattering are still misjudged as structural details by the densityization mechanism, triggering unnecessary Gaussian generation and resulting in redundant geometry and floating pseudo-structures. Second, when jointly optimizing water optical parameters (such as attenuation coefficient, scattering coefficient, and background color at infinity), the background light term is prone to absorbing geometric and color errors in the early stages of training, degenerating into non-physical understanding, thus interfering with the learning process of Gaussian geometry.

[0005] Therefore, how to achieve a three-dimensional Gaussian representation reconstruction method that balances geometric compactness, parameter stability, and visual fidelity in an underwater environment remains a technical problem that urgently needs to be solved. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a three-dimensional Gaussian splash reconstruction method for underwater scenarios.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a three-dimensional Gaussian splash reconstruction method for underwater scenes, comprising the following steps:

[0008] Step 1: Obtain a multi-view 2D image sequence of the target underwater scene, camera pose parameters corresponding to the multi-view 2D image sequence, and an initial sparse point cloud. Initialize a 3D Gaussian point set based on the initial sparse point cloud. The Gaussian point set includes geometric parameters and appearance parameters. At the same time, initialize the learnable physical parameters of the underwater imaging model. The physical parameters include at least the attenuation coefficient, backscattering coefficient, and background color at infinity.

[0009] Step 2: Based on the multi-view 2D image sequence and the corresponding camera pose parameters, iteratively optimize and train the 3D Gaussian point set and the learnable physical parameters of the underwater imaging model; during the iterative optimization and training process, when the preset density scheduling conditions are met, calculate the importance score of each Gaussian point based on the opacity-weighted multi-view consistency mechanism, and perform the density operation of Gaussian points according to the importance score, so that structural refinement occurs first in the region with significant visual contribution, thereby suppressing the generation of redundant Gaussian points in the low opacity region;

[0010] Step 3: A phased freezing strategy is adopted to co-optimize the parameters of the Gaussian point set and the underwater imaging model. The parameters of the Gaussian point set and the underwater imaging model are selectively frozen or updated at different training stages to avoid the degradation of the water optical parameters in the early stage of training and reduce their interference with the learning of geometric structure, thereby improving the convergence stability of geometric modeling and physical imaging, and training to obtain the optimized three-dimensional Gaussian point set and underwater imaging model parameters.

[0011] Step 4: After training is completed, based on the optimized Gaussian point set and underwater imaging model parameters, render and synthesize underwater images for any new viewpoint camera pose, and output the synthesized image from the new viewpoint, as well as the 3D Gaussian representation reconstruction result of the underwater scene.

[0012] Furthermore, the iterative optimization training described in step 2 includes: generating a predicted image using differentiable rendering technology in each iteration and comparing it with the corresponding real image to calculate the training loss, and updating the parameters in reverse based on the training loss.

[0013] Furthermore, the density scheduling conditions described in step 2 include: triggering the calculation of the importance score and the density operation within a preset iteration range and at a preset iteration interval during the training process.

[0014] Furthermore, the opacity-weighted multi-view consistency mechanism described in step 2 is as follows: under multiple training views, the pixel-by-pixel reconstruction error and the rendering opacity of the corresponding pixel are weighted, a high-error region is constructed based on the weighted error, and the number of high-error pixels covered by each Gaussian point under multiple views is counted as the basis for the density decision of the Gaussian point.

[0015] Furthermore, the high error region is obtained by normalizing and thresholding the opacity-weighted error, where the threshold is a preset value.

[0016] Furthermore, the phased freezing strategy described in step 3 includes: optimizing only the Gaussian point set and an auxiliary learnable background color parameter during the initial training phase; then fixing the three-dimensional Gaussian point set, using the auxiliary learnable background color parameter as the initial value of the infinity background color parameter, freezing the infinity background color parameter, and optimizing the attenuation coefficient and backscattering coefficient separately; next, freezing all underwater imaging model parameters, comprehensively optimizing the Gaussian point set to adapt to the calibrated optical model; and finally, using an alternating optimization method to jointly optimize the three-dimensional Gaussian point set and the underwater imaging model parameters.

[0017] Furthermore, the loss function used in the training process includes a weighted combination of pixel reconstruction error loss and structural similarity loss.

[0018] Furthermore, the camera pose and the initial sparse point cloud can be obtained through the motion recovery structure method.

[0019] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described thereon.

[0020] The present invention also provides an electronic device, including a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the method.

[0021] Compared with the prior art, the present invention has at least the following beneficial effects:

[0022] (1) By introducing an opacity-aware multi-view consistency density control mechanism, the generation and splitting of Gaussian points are focused on the region that contributes significantly to the final image, effectively suppressing redundant Gaussian points and floating pseudo-structures caused by error accumulation in the low opacity background, improving the geometric compactness of the model and reducing redundant Gaussian points in the low opacity region.

[0023] (2) By using a phased freezing and collaborative optimization strategy, the risk of water background light terms being driven by noise and degraded in the early training stage is reduced, and the stability and interpretability of water physical parameter estimation such as attenuation and scattering are improved.

[0024] (3) The overall method does not require additional hardware input such as depth sensors. It can realize underwater 3D Gaussian representation reconstruction by relying only on monocular video and the estimated camera pose, which has good practicality and promotion value. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the overall process of the three-dimensional Gaussian splash reconstruction method for underwater scenes provided in the embodiments of the present invention;

[0027] Figure 2 This is a schematic diagram of the phased freezing and collaborative optimization training strategy provided in an embodiment of the present invention;

[0028] Figure 3 These are schematic diagrams comparing the results of the novel perspective synthesis of the method of the present invention and the comparative method in four underwater scenarios. Among them, (a) is a comparison diagram in the Curasao scenario, (b) is a comparison diagram in the Japanese Gardens-RedSea scenario, (c) is a comparison diagram in the Panama scenario, and (d) is a comparison diagram in the IUI3-RedSea scenario. Detailed Implementation

[0029] To enable those skilled in the art to better understand the technical solutions in the embodiments of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of this application.

[0030] The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0031] like Figure 1As shown, this embodiment of the invention provides a three-dimensional Gaussian splash reconstruction method for underwater scenes. The overall process includes: Step S1: Based on a multi-view two-dimensional image sequence of the target underwater scene, the camera pose parameters are determined using existing multi-view geometric reconstruction methods, and an initial sparse point cloud is reconstructed. A three-dimensional Gaussian point set is initialized based on the initial sparse point cloud, the Gaussian point set containing geometric parameters and appearance parameters. Simultaneously, the learnable physical parameters of the underwater imaging model are initialized, the physical parameters including attenuation coefficient, backscattering coefficient, and background color at infinity. Based on the multi-view two-dimensional image sequence and the corresponding camera pose parameters, the three-dimensional Gaussian point set and the learnable physical parameters of the underwater imaging model are iteratively optimized and trained.

[0032] Step S2: During the iterative optimization training process, when the preset density scheduling conditions are met, the importance score of each Gaussian point is calculated based on the opacity-weighted multi-view consistency mechanism, and the density operation of the Gaussian point is performed according to the importance score, so that the structural refinement occurs first in the region with significant visual contribution, thereby suppressing the generation of redundant Gaussian points in the low opacity region.

[0033] Step S3: In the iterative optimization process, a phased freezing strategy is adopted to co-optimize the parameters of the Gaussian point set and the underwater imaging model. The parameters of the Gaussian point set and the underwater imaging model are selectively frozen or updated at different training stages to avoid the degradation of water optical parameters in the early stage of training and reduce their interference with geometric structure learning, thereby improving the convergence stability of geometric modeling and physical imaging, and training to obtain the optimized three-dimensional Gaussian point set and underwater imaging model parameters.

[0034] Step S4: After the iterative optimization training is completed, based on the optimized Gaussian point set and underwater imaging model parameters, render and synthesize underwater images for any new viewpoint camera pose, and output the synthesized image from the new viewpoint, as well as the 3D Gaussian representation reconstruction result of the underwater scene.

[0035] In a preferred embodiment, step S1 is implemented in the following manner:

[0036] Acquire multi-view 2D image sequences of the target underwater scene (in ), and the camera pose sequence corresponding to each image. (in ), and obtain the initial sparse point cloud. Using the initial point cloud Initialize a three-dimensional Gaussian point set Each Gaussian point Includes geometric parameters (center position) Covariance matrix ) and appearance parameters (third-order spherical harmonic color coefficient) Opacity ).

[0037] Simultaneously, three learnable physical parameters of the underwater imaging model are initialized:

[0038] Attenuation coefficient , used to describe the exponential decay rate of light in water under different color channels;

[0039] Backscattering coefficient , used to describe the intensity of backscattering caused by a suspended medium as depth increases;

[0040] Infinity background color , representing the stable background color of the water body as the propagation distance approaches infinity.

[0041] In the early stages of training, It can be learned by an auxiliary background color parameter Initialization is performed to improve the accuracy of water body parameter initialization; the specific usage strategy for this parameter will be detailed in step S3.

[0042] Among them, the three learnable physical parameters mentioned above are all global three-channel vectors, which can be optimized through backpropagation during the training process.

[0043] In this step, the multi-view two-dimensional image sequence is the original image captured by the underwater camera, which includes color shift and contrast attenuation caused by the turbid medium; the camera pose and sparse point cloud are obtained by existing multi-view geometric reconstruction methods, such as reconstruction methods based on motion recovery structures; the specific acquisition method does not constitute a limitation of this invention.

[0044] As a preferred embodiment, the iterative optimization training in step S1 specifically involves: during the training phase, constructing a training loop; in each iteration, performing differentiable rendering of the 3D Gaussian point set based on the camera pose parameters to obtain a predicted image, a predicted depth map, and a predicted opacity map; comparing the predicted image with the corresponding training image to calculate the training loss, which includes at least a weighted sum of pixel reconstruction error and structural similarity loss terms between the predicted image and the corresponding training image, and optionally other image losses; at the end of each iteration, backpropagating the training loss to update the Gaussian point set parameters and underwater imaging model parameters, wherein the Gaussian point set parameters and underwater imaging model parameters are subject to rules for freezing and updating based on the iteration count; performing adaptive density control on the Gaussian point set at a specific iteration count, including densityization and pruning; and ending the iterative optimization process when the preset number of iterations is reached.

[0045] In a preferred embodiment, step S2 is implemented in the following manner:

[0046] During the iterative optimization process, when the preset density scheduling conditions are met (e.g., the number of iterations is between 500 and 15000 and every 100 iterations), the following operations are performed to calculate the importance score of each Gaussian point:

[0047] Step S2.1: Sample multiple training views.

[0048] Randomly select without replacement from the training camera set One perspective, recorded as These represent the first through Kth viewpoints, respectively.

[0049] Step S2.2: Construct an opacity-weighted high-error region mask.

[0050] For each perspective First, the Gaussian point set is rendered to obtain the predicted image. and the corresponding opacity map ,in A larger value indicates that the pixel is less transparent, and a smaller value indicates that it is more transparent, which is consistent with the real image. Calculate the pixel-by-pixel L1 reconstruction error:

[0051] ;

[0052] The pixel-by-pixel L1 reconstruction error is multiplied by the opacity map pixel by pixel to obtain the opacity-weighted error:

[0053] ;

[0054] Normalize the opacity weighted error to The interval is used to obtain the normalized error plot:

[0055] ;

[0056] in, To prevent division by zero by small constants, the normalized error plot is weighted by opacity. The high-error mask is obtained by normalization, and therefore corresponds to pixel regions with high opacity and large reconstruction errors. Subsequently, the normalized error map is thresholded to generate the high-error mask.

[0057] ;

[0058] in For indicator functions, This is a preset threshold; when the condition within the parentheses is met... ,otherwise .

[0059] Step S2.3: Count the number of high error pixels covered by each Gaussian point.

[0060] For Gauss point From the perspective The two-dimensional projection region below is denoted as Through the rasterization process, statistics are collected within this region that satisfy... The number of pixels, denoted as .in, Indicates perspective Lower Gauss point The number of pixels selected by a high-error mask within the two-dimensional projection area, which has been constructed in the previous steps using opacity-weighted error.

[0061] Step S2.4: Calculate cross-perspective importance scores and make decisions dense.

[0062] The Gaussian point is obtained by averaging over all sampled viewpoints. Importance score:

[0063] ;

[0064] like If the condition is met, then a cloning or splitting operation is performed on the Gaussian point; otherwise, it remains unchanged, where... This is the density threshold parameter; in one embodiment, It can be set to 5. This step uses opacity... The pixel-by-pixel L1 reconstruction error is weighted, and the weighted error is normalized and thresholded to construct a high-error mask. This makes the density operation preferentially occur in visually significant (high opacity) areas with large reconstruction errors, effectively suppressing the generation of redundant Gaussian points in low-opaque areas (such as the background).

[0065] In a preferred embodiment, step S3 is implemented in the following manner:

[0066] like Figure 2 As shown, this invention employs a four-stage training strategy to strictly control the timing of parameter updates for each module:

[0067] Phase 1: Geometric Warm-up Phase (Iterations 1 to 10000)

[0068] Optimize only Gaussian point set All parameters (center position, covariance, color, opacity) and learnable background color parameter No underwater imaging modules are enabled, including the infinity background color, which has a clear physical meaning within the modules. Instead, it introduces an auxiliary learnable background color parameter. and in Phase Two as The initial value.

[0069] In each iteration of the iterative optimization, for all training images, the current 3D Gaussian point set is rendered differently using the corresponding camera pose parameters to obtain the current Gaussian point set. The media-free clean image rendered from the corresponding viewpoint and opacity images And synthesize the predicted image using the background synthesis formula. For the predicted image , and the real input image Calculate the training loss. The loss function is defined as:

[0070] ;

[0071] SSIM stands for Structural Similarity, which measures the consistency between the predicted image and the real image in terms of brightness, contrast, and structural information. As a hyperparameter, in one embodiment During this stage, The background color is used as a learnable element in the synthesis and parameter optimization. During this phase, a densityization operation is performed every 100 iterations (according to S2), and the opacity of the Gaussian point is reset to a small positive value (e.g., 0.001) at the 3000th, 6000th, and 9000th iterations.

[0072] Phase Two: Optical Model Initialization Phase (Iterations 10001 to 11000)

[0073] The auxiliary learnable background color parameters obtained at the end of Phase 1 As the background color of infinity The initial values ​​are set, and the Gaussian point set is frozen. Background color at infinity Only optimize the attenuation parameters With backscattering parameters At this point, the system synthesizes an underwater image:

[0074] ;

[0075] in, This indicates element-wise multiplication. This is a depth map rendered from a Gaussian point set. Indicates the current Gaussian point set A clean, media-free image rendered from the corresponding viewpoint; and and This means that the three-channel vector is broadcast to each pixel channel by channel and then multiplied element-wise with the depth map. At this point, an underwater imaging physics model is used for image synthesis, and its form differs from the background synthesis formula in Stage 1.

[0076] The loss function remains the same. However, backpropagation only updates... and .

[0077] Phase 3: Geometric Adaptation Phase (Iterations 11001 to 13000)

[0078] Freeze all underwater imaging modules (i.e.) , , (All fixed), comprehensively optimize Gaussian point set This stage continues to perform densityization and pruning operations, and resets the opacity at the 12000th iteration to adapt the geometry to the calibrated underwater optical model; wherein, the pruning operation refers to removing Gaussian points with opacity below a preset threshold (e.g., 0.005).

[0079] Phase 4: Joint Optimization Phase (starting from the 13001st iteration)

[0080] An alternating optimization strategy is adopted: In every 100 consecutive iterations, for the first 99 iterations: the underwater imaging module is frozen, the Gaussian point set is optimized, and density and opacity are reset; for the 100th iteration: the Gaussian point set is frozen, and joint optimization is performed. , and .

[0081] In the above-mentioned alternating optimization process, all parameter updates are sequentially advanced under the same main iteration counter, without introducing additional inner loops or skip iterations, thereby ensuring that the densityization and opacity reset operations are performed according to the predetermined training rhythm.

[0082] In a preferred embodiment, step S4 specifically includes:

[0083] After training is complete, load the optimized Gaussian point set. With underwater parameters For any new camera pose from a new perspective The system performs the following operations:

[0084] (1) Render the Gaussian point set to obtain the base image. and the corresponding depth map ;

[0085] (2) Synthesize underwater images based on the base image, depth map, and optimized underwater parameters:

[0086] .

[0087] (3) Output the synthesized underwater image This serves as a new perspective on underwater reconstruction results.

[0088] As a preferred embodiment, the present invention achieves a balance between geometric compactness, background physical plausibility, and rendering fidelity in underwater scenes through the above-described process. To facilitate understanding of the essential effects of the present invention, specific examples will be provided below to demonstrate the novel perspective synthesis effects of the method on different underwater scene data.

[0089] This embodiment selects four underwater scenes as evaluation objects, denoted as Scene 1 to Scene 4, which correspond to the Curasao, Japanese Gardens-RedSea, Panama, and IUI3-RedSea scenes in the SeathruNeRF dataset, respectively. The comparison method is the existing technology SeaSplat (hereinafter referred to as the comparison method). To ensure the fairness of the comparison, the method of this invention and the comparison method are trained and tested under the same data partitioning and the same camera parameter settings, and evaluated on the same set of new perspectives.

[0090] like Figure 3 As shown, the results of novel perspective synthesis of the method of the present invention and the comparison method in four underwater scenes are compared. From left to right, they are the real image, the reconstructed image of the comparison method, and the reconstructed image of the method of the present invention; wherein, Figure 3 In the diagram, (a)-(d) correspond to the visualization results examples for scenarios one through four, respectively. Figure 3 As can be seen, compared with the comparison method, the present invention can effectively suppress three-dimensional Gaussian floating artifacts in low opacity background areas.

[0091] To quantitatively evaluate the synthesis effect of the new perspective, this embodiment calculates Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Perceptual Similarity (LPIPS) as evaluation metrics on the test perspective set. Higher PSNR and SSIM values ​​indicate better synthesis quality, while lower LPIPS values ​​indicate a more perceptually realistic image. All metrics are averaged across the test perspective set to reflect the overall performance of the new perspective synthesis. Simultaneously, to assess the geometric compactness and resource overhead of the model, this embodiment counts the number of 3D Gaussian points used to represent the scene after training convergence. A higher number of Gaussian points generally implies a larger parameter scale and higher storage / rendering overhead. The quantitative comparison results of the proposed method and the comparative method on four underwater scenes are shown in Table 1.

[0092] Table 1: Quantitative comparison results of the present invention and the comparative method in four underwater scenarios (average from test perspectives)

[0093] As shown in Table 1, compared with the comparison method, the present invention achieves higher PSNR and SSIM in Scene 1, Scene 2, and Scene 4, indicating that it can improve the synthesis quality of new perspectives in terms of distortion and structural similarity. Simultaneously, the present invention significantly reduces the number of 3D Gaussian points in all four scenes, demonstrating that the proposed opacity-aware density control mechanism can effectively suppress the generation of redundant Gaussian points in low-opacity regions, thereby obtaining a more compact 3D representation and reducing resource overhead. On the other hand, LPIPS, as a perceptual similarity index, did not show a synchronous decrease in this embodiment, indicating that the improvement at the perceptual level is not entirely consistent with PSNR / SSIM. Nevertheless, the present invention achieves a significant reduction in the number of Gaussian points while maintaining or improving PSNR / SSIM, still demonstrating its advantages in geometric compactness and engineering efficiency.

[0094] This invention also provides an electronic device, including a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the method.

[0095] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described thereon.

[0096] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reconstructing three-dimensional Gaussian splashes in underwater scenes, characterized in that, include: The system acquires a multi-view two-dimensional image sequence of the target underwater scene, camera pose parameters corresponding to the multi-view two-dimensional image sequence, and an initial sparse point cloud. It then initializes a three-dimensional Gaussian point set based on the initial sparse point cloud. The three-dimensional Gaussian point set includes geometric parameters and appearance parameters. Initialize the learnable physical parameters of the underwater imaging model, wherein the learnable physical parameters include at least the attenuation coefficient, the backscattering coefficient, and the background color at infinity; Based on the multi-view two-dimensional image sequence and the corresponding camera pose parameters, the learnable physical parameters of the three-dimensional Gaussian point set and the underwater imaging model are iteratively optimized and trained. During the iterative optimization training process, when the preset density scheduling conditions are met, the importance score of each Gaussian point is calculated based on the opacity-weighted multi-view consistency mechanism, and the density operation is performed on the three-dimensional Gaussian point set according to the importance score. The opacity-weighted multi-view consistency mechanism is specifically as follows: under multiple training views, the pixel-by-pixel reconstruction error is multiplied with the rendering opacity of the corresponding pixel, a high-error region is constructed based on the multiplied error, and the number of high-error pixels covered by each Gaussian point under multiple views is counted as the importance score of the Gaussian point or the basis for calculating the importance score. A phased freezing strategy is adopted to collaboratively optimize the learnable physical parameters of the 3D Gaussian point set and the underwater imaging model. The parameters of the 3D Gaussian point set and the learnable physical parameters are selectively frozen or updated at different training stages to obtain the optimized 3D Gaussian point set and underwater imaging model parameters. After training, based on the optimized 3D Gaussian point set and underwater imaging model parameters, rendering and underwater image synthesis are performed on any new perspective camera pose, outputting the new perspective synthesized image and the 3D Gaussian representation reconstruction result of the underwater scene.

2. The method according to claim 1, characterized in that, The density scheduling conditions include: triggering the calculation of the importance score and the density operation within a preset iteration range and at a preset iteration interval during the training process.

3. The method according to claim 1, characterized in that, The iterative optimization training includes: generating a predicted image using differentiable rendering technology in each iteration and comparing it with the corresponding real image to calculate the training loss, and updating the parameters in reverse based on the training loss.

4. The method according to claim 1, characterized in that, The high-error region is obtained by normalizing and thresholding the multiplied error.

5. The method according to claim 1, characterized in that, The phased freezing strategy includes: During the initial training phase, only the three-dimensional Gaussian point set and an auxiliary learnable background color parameter are optimized. Then, the three-dimensional Gaussian point set is fixed, the auxiliary learnable background color parameter is used as the initial value of the background color at infinity, and the background color at infinity is frozen. The attenuation coefficient and the backscattering coefficient are optimized separately. Next, all underwater imaging model parameters were frozen, and the three-dimensional Gaussian point set was comprehensively optimized. Finally, an alternating optimization method is used to jointly optimize the three-dimensional Gaussian point set and the underwater imaging model parameters.

6. The method according to claim 1, characterized in that, The loss function used in the training process includes a weighted combination of pixel reconstruction error loss and structural similarity loss.

7. The method according to claim 1, characterized in that, The camera pose and the initial sparse point cloud are obtained using existing multi-view geometric reconstruction methods.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to implement the method as claimed in any one of claims 1 to 7.

Images