A method for underwater scene reconstruction and water medium separation based on 3D Gaussian model

Abstract

The present invention discloses a method for underwater scene reconstruction and water medium separation based on a 3D Gaussian model, which relates to the field of computer graphics technology. The method comprises the following steps: constructing a Gaussian object; selecting a fixed number of Gaussians with random mean values ​​to simulate a Gaussian water medium and initializing the Gaussian water medium; rendering the Gaussian object and the Gaussian water medium to obtain a 2D image of the Gaussian object and a 2D image of the Gaussian water medium, and outputting a medium transmission map; fusing the 2D image of the Gaussian object, the 2D image of the Gaussian water medium, and the medium transmission map to obtain a 2D image of the underwater scene; substituting the 2D image of the underwater scene and the enhanced underwater image into the loss function of a background reconstruction algorithm for training and optimization to obtain a trained Gaussian object; rendering the trained Gaussian object along a custom directional viewing angle, and outputting a 2D image of the underwater scene separated from the water medium. The present invention meets the high efficiency requirements of practical application scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of computer graphics, and in particular to a method for underwater scene reconstruction and water medium separation based on a 3D Gaussian model. Background Art

[0002] In the field of underwater 3D reconstruction, traditional technical means such as photometric stereo, structured light and stereo vision, although they have been widely used in fields such as underwater equipment maintenance, environmental monitoring and archaeological excavation, are still subject to many limitations. On the one hand, these methods often rely on specific hardware equipment, which not only leads to high costs but also places relatively strict requirements on the use environment. On the other hand, due to the influence of the water medium, these technologies often lead to problems such as color distortion, brightness reduction and image blur in the reconstruction results. In the past two years, underwater scene reconstruction technology based on neural radiance field (NeRF) has been able to use deep learning to draw scenes and achieve separation of water media while rendering high-quality views. However, this technology does not capture scene details accurately enough and has low rendering efficiency, making it difficult to meet the needs of actual application scenarios.

[0003] Against this backdrop, 3D Gaussian modeling, a newly emerging technology in computer vision, has become an ideal solution for underwater scene reconstruction due to its directness, flexibility, and efficiency, along with the rapidity of Gaussian splash rendering. However, if 3D Gaussian modeling is directly applied to underwater environments, the model struggles to eliminate interference from the water medium and cannot accurately capture the true nature of the scene, potentially misrendering the water as solid floating objects. Furthermore, the reconstruction process lacks specialized strategies for reconstructing blurred background areas. Summary of the Invention

[0004] In order to solve the above-mentioned problems in underwater scene reconstruction and water medium separation, the present invention proposes an underwater scene reconstruction and water medium separation method based on a 3D Gaussian model, which solves the problem that the existing 3D Gaussian technology fails to accurately model the water medium in underwater scene reconstruction and lacks an effective reconstruction strategy for background blurred areas.

[0005] To achieve the above objectives, the present invention provides a method for underwater scene reconstruction and water medium separation based on a 3D Gaussian model, comprising:

[0006] Constructing an underwater scene dataset, obtaining sparse point cloud data of the underwater scene, constructing a Gaussian object based on the sparse point cloud data, and initializing the Gaussian object;

[0007] Select a fixed number of Gaussians with random means to simulate Gaussian water medium and initialize the Gaussian water medium;

[0008] Using a differentiable 3D Gaussian sputtering rendering method, the Gaussian object and the Gaussian water medium are rendered respectively to obtain a 2D image of the Gaussian object and a 2D image of the Gaussian water medium, and output a medium transmission map;

[0009] fusing the 2D image of the Gaussian object, the 2D image of the Gaussian water medium, and the medium transmission map using an image formation model to obtain a 2D image of the underwater scene;

[0010] The underwater scene 2D image and the enhanced underwater image are substituted into the loss function of the background reconstruction algorithm for training optimization to obtain a trained Gaussian object, and the trained Gaussian object is rendered along a custom directional perspective to output a 2D image of the underwater scene separated by the water medium.

[0011] Preferably, constructing the underwater scene dataset includes:

[0012] A monocular camera is used to capture underwater scenes. The camera moves around the target scene to capture continuous information of the scene, and each frame of the video is enhanced based on the color channel compensation method to obtain an enhanced underwater image sequence.

[0013] The enhanced underwater image sequence is detected and matched using the motion structure recovery (SfM) method to reconstruct sparse point cloud data of the underwater scene, and the camera parameters are simultaneously extracted to obtain the underwater scene dataset.

[0014] Preferably, the color channel compensation method includes:

[0015] The average brightness values ​​of the blue, green and red color channels in the image are compared and analyzed. When the average brightness value of the blue channel is the highest, the blue channel will be used to compensate the green channel, and the green channel will be used to compensate the red channel; if the average brightness value of the green channel is the highest, the green channel will be used to compensate the blue and red channels; if the average brightness value of the red channel is the highest, the green channel will be used to compensate the blue channel, and the red channel will be used to compensate the green channel.

[0016] Preferably, initializing the Gaussian object includes:

[0017] The position coordinates of any sparse point cloud are set to the Gaussian mean of the Gaussian object, and the K nearest neighbor algorithm is used to obtain the average distance D = {d1, d2, ..., d n};

[0018] Set the Gaussian scaling parameters, rotation parameters, Gaussian opacity and Gaussian spherical harmonic coefficients respectively to obtain the initialized Gaussian object.

[0019] Preferably, performing the Gaussian water medium initialization includes:

[0020] The random numbers in the preset interval are used as the Gaussian mean of the Gaussian water medium, and the K-nearest neighbor algorithm is used to obtain the average distance between each Gaussian and the three nearest Gaussians.

[0021] The scaling parameters, rotation parameters, opacity, training interval and spherical harmonic coefficients of the Gaussian are set respectively to obtain the initialized Gaussian water medium.

[0022] Preferably, rendering the Gaussian object comprises:

[0023] According to the camera intrinsic parameters, camera extrinsic parameters and perspective projection function, the Gaussian projection of the Gaussian object is performed onto the image plane, and the 2D mean, 2D covariance matrix and volume density contribution of the Gaussian corresponding to the Gaussian object are obtained. The color value of the Gaussian corresponding point of the Gaussian object is rendered using the volume density formula to obtain a 2D image of the Gaussian object;

[0024] Rendering the Gaussian water medium includes:

[0025] According to the camera intrinsic parameters, camera extrinsic parameters and perspective projection function, the Gaussian of the Gaussian water medium is projected onto the image plane to obtain the 2D mean, 2D covariance matrix and volume density contribution of the Gaussian corresponding to the Gaussian water medium at a point on the image plane. The color value of the Gaussian corresponding point of the Gaussian water medium is rendered using the volume density formula to obtain a 2D image of the Gaussian water medium.

[0026] Preferably, obtaining the 2D image of the underwater scene includes:

[0027] Inputting the 2D image of the Gaussian object, the 2D image of the water medium, and the medium transmission map into the image formation model for fusion to generate the 2D image of the underwater scene;

[0028] The 2D image of the Gaussian object is used as the unabsorbed radiation value, and the 2D image of the water medium is used as the forward scattered component.

[0029] Preferably, obtaining the trained Gaussian object includes:

[0030] The background area of ​​the underwater scene 2D image and the enhanced underwater image is estimated using a background reconstruction algorithm, and the reconstruction of the background area is guided by enhancing the loss term of the background area to obtain the trained Gaussian object; wherein the entire training process undergoes 30,000 cycles of iteration.

[0031] Compared with the prior art, the present invention has the following advantages and technical effects:

[0032] (1) The present invention does not rely on specific equipment. It only requires ordinary underwater imaging equipment to continuously capture video of the target underwater scene. Through data training, it can efficiently and accurately reconstruct the underwater scene from the underwater video or image sequence captured by a monocular camera, and output a 2D image that is not affected by the water medium. In addition, the present invention has the ability to output a 2D image sequence of underwater scenes without the influence of the water medium at a rate of 130 frames per second, fully meeting the efficiency requirements of practical application scenarios.

[0033] (2) In the field of underwater 3D reconstruction, the present invention has a wider range of applications, particularly in key areas such as underwater equipment maintenance, underwater environmental monitoring, and underwater archaeological excavation, providing a more efficient and accurate solution. Furthermore, the present invention contributes an innovative data generation method for underwater environmental exploration.

[0034] (3) Referring to the image formation model, the present invention uses two Gaussian components of water medium and object to express the underwater scene, which is not only more consistent with the physical process, but also can separate the water medium and the object, thereby obtaining a clear rendering result of the scene that is not affected by the water medium.

[0035] (4) The present invention provides a background light estimation method and a corresponding background light loss function to solve the problem of insufficient reconstruction of blurred areas in underwater scenes, which helps to further improve the realism and accuracy of rendering effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0037] Figure 1 This is a flow chart of a method for underwater scene reconstruction and water medium separation based on a 3D Gaussian model according to an embodiment of the present invention;

[0038] Figure 2 Schematic diagram of an image formation model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0040] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] The present invention proposes a method for underwater scene reconstruction and water medium separation based on 3D Gaussian model. Figure 1-Figure 2 ,include:

[0042] Constructing an underwater scene dataset, obtaining sparse point cloud data of the underwater scene, constructing a Gaussian object based on the sparse point cloud data, and initializing the Gaussian object;

[0043] Select a fixed number of Gaussians with random means to simulate Gaussian water medium and initialize the Gaussian water medium;

[0044] Using a differentiable 3D Gaussian sputtering rendering method, the Gaussian object and the Gaussian water medium are rendered respectively to obtain a 2D image of the Gaussian object and a 2D image of the Gaussian water medium, and output a medium transmission map;

[0045] fusing the 2D image of the Gaussian object, the 2D image of the Gaussian water medium, and the medium transmission map using an image formation model to obtain a 2D image of the underwater scene;

[0046] The underwater scene 2D image and the enhanced underwater image are substituted into the loss function of the background reconstruction algorithm for training optimization to obtain a trained Gaussian object, and the trained Gaussian object is rendered along a custom directional perspective to output a 2D image of the underwater scene separated by the water medium.

[0047] This embodiment does not rely on specific equipment; it simply requires conventional underwater imaging equipment to continuously capture a target underwater scene. Through data training, it can efficiently and accurately reconstruct underwater scenes from underwater video or image sequences captured by a monocular camera, and output 2D images unaffected by the water medium. Furthermore, the present invention is capable of outputting 2D image sequences of underwater scenes unaffected by the water medium at a rate of 130 frames per second, fully meeting the efficiency requirements of practical applications.

[0048] Furthermore, an underwater scene dataset is constructed, including:

[0049] A monocular camera is used to capture underwater scenes. The camera moves around the target scene to capture continuous information of the scene, and each frame of the video is enhanced based on the color channel compensation method to obtain an enhanced underwater image sequence.

[0050] The enhanced underwater image sequence is detected and matched using the motion structure recovery (SfM) method to reconstruct sparse point cloud data of the underwater scene, and the camera parameters are simultaneously extracted to obtain the underwater scene dataset.

[0051] Specifically, a monocular camera G is used to capture video of an underwater scene S, slowly circling the scene's target to collect continuous information about the scene. For each image, the average value of each of the three color channels is calculated. Depending on the average value of each color channel, a corresponding color channel compensation method is employed. Specifically, when the blue channel's average value is the highest, the blue channel is used to compensate the green channel, while the green channel simultaneously compensates the red channel. If the green channel's average value is the highest, the green channel compensates both the blue and red channels. When the red channel's average value is the highest, the red channel is used to compensate the green channel, while the green channel compensates the blue channel. This process yields a color-corrected and enhanced image sequence.

[0052] The color channel compensation method is an image enhancement technology that uses one of the three color channels, red, green, and blue, to compensate for another wavelength adjacent color channel. By compensation channel and compensated channel Calculation yields:

[0053]

[0054] in, and are the mean of the compensation channel and the mean of the compensated channel respectively.

[0055] The enhanced image sequence is input into SfM detection and matching to obtain the sparse point cloud P = {p1, p2, ..., p n}、Camera internal parameters External Reference The position coordinates of the sparse point cloud P are V = {v1,v2,…,v n},

[0056] Furthermore, the Gaussian object is initialized, including:

[0057] The position coordinates of any sparse point cloud are set to the Gaussian mean of the Gaussian object, and the K nearest neighbor algorithm is used to obtain the average distance D = {d1, d2, ..., d n};

[0058] Set the Gaussian scaling parameters, rotation parameters, Gaussian opacity and Gaussian spherical harmonic coefficients respectively to obtain the initialized Gaussian object.

[0059] Specifically, the position coordinates of the sparse point cloud P are set to the Gaussian coordinates of the Gaussian object mean

[0060] The K-nearest neighbor algorithm is used to obtain the average distance D = {d1, d2, ..., d n}; In this embodiment, the scaling parameter of Gauss is set to The average distance D = {d1, d2, ..., d n The square of the i-th element in}; the Gaussian rotation parameter is set to Gaussian opacity Uniformly set to 1; Gaussian spherical harmonic coefficient f i obj Set to 0;

[0061] The initialized Gaussian object is represented as

[0062] Furthermore, Gaussian water medium initialization is performed, including:

[0063] The random numbers in the preset interval are used as the Gaussian mean of the Gaussian water medium, and the K-nearest neighbor algorithm is used to obtain the average distance between each Gaussian and the three nearest Gaussians.

[0064] The scaling parameters, rotation parameters, opacity, training interval and spherical harmonic coefficients of the Gaussian are set respectively to obtain the initialized Gaussian water medium.

[0065] Specifically, the Gaussian water medium

[0066] In this embodiment, the number of Gaussians in the Gaussian water medium is fixed to 100. A random number between [-20, 20] is used as the Gaussian mean. Use the K nearest neighbor algorithm to find the average distance between each Gaussian and the three nearest Gaussians. The scaling parameter of the Gaussian is set to is the average distance The square of the jth element in ; the Gaussian rotation parameter is set to the matrix Gaussian opacity Set to 0.1; limited to the interval (0,0.1] throughout the training process; Gaussian spherical harmonic coefficients Set to 0.

[0067] The initialized Gaussian water medium is expressed as

[0068] Furthermore, the Gaussian object is rendered, including:

[0069] According to the camera intrinsic parameters, camera extrinsic parameters and perspective projection function, the Gaussian projection of the Gaussian object is performed onto the image plane, and the 2D mean, 2D covariance matrix and volume density contribution of the Gaussian corresponding to the Gaussian object are obtained. The color value of the Gaussian corresponding point of the Gaussian object is rendered using the volume density formula to obtain a 2D image of the Gaussian object;

[0070] The perspective projection function π(x,E,K) can transform the three-dimensional coordinate point Transform to the two-dimensional coordinate point of the image The specific process is:

[0071]

[0072] in, are the homogeneous coordinates of the two-dimensional coordinate point, are the homogeneous coordinates of the three-dimensional coordinate point, is the camera internal parameter, is the camera extrinsic parameter.

[0073] Specifically, according to the camera internal parameters Camera extrinsics And the perspective projection function π(x; E, K), which transforms the Gaussian of the Gaussian object into Projected onto the image plane, Gaussian The corresponding 2D mean 2D covariance matrix A point on the image plane The volume density contribution is in, is the mean of the Gaussian object, is the 2D covariance matrix, J is the Jacobi affine approximation matrix of E, is the 3D covariance matrix, E T is the transpose of E, J T is the transpose of J, T represents the matrix transpose, is the 2D mean of the Gaussian object.

[0074] Refer to the volume density formula to render the color value of the Gaussian corresponding point of the Gaussian object:

[0075]

[0076] Among them, q is the direction from the camera focus to x, Is f i obj Encoded spherical harmonics, N is the number of Gaussian objects, is the point of the i-th Gaussian object on the image plane The volume density contribution of is the point on the image plane of the i-1 Gaussian object before calculation The remaining volume density value of the pixel after the volume density contribution of ; i is the i-th Gaussian object, k is the k-th Gaussian object, and the range of k is 1 to i-1.

[0077] Furthermore, the Gaussian water medium is rendered, including:

[0078] According to the camera intrinsic parameters, camera extrinsic parameters and perspective projection function, the Gaussian of the Gaussian water medium is projected onto the image plane to obtain the 2D mean, 2D covariance matrix and volume density contribution of the Gaussian corresponding to the Gaussian water medium at a point on the image plane. The color value of the Gaussian corresponding point of the Gaussian water medium is rendered using the volume density formula to obtain a 2D image of the Gaussian water medium.

[0079] Specifically, the Gaussian water medium Projected onto the image plane, Gaussian The corresponding 2D mean 2D covariance matrix A point on the image plane The volume density contribution is Where J is the Jacobi affine approximation matrix of E, is the 3D covariance matrix, E T is the transpose of E, J T is the transpose of J, is the 2D mean of the Gaussian water medium.

[0080] Refer to the volume density formula to render the color value of the Gaussian corresponding point of the Gaussian water medium Therefore Encoded spherical harmonics, is the jth Gaussian water medium at a point on the image plane The volume density contribution of is the j-1 Gaussian water medium at a point on the image plane before calculation The remaining volume density value of the pixel after the volume density contribution; j is the jth Gaussian water medium, l is the lth Gaussian water medium, and the range of l is 1 to j.

[0081] For the medium transmission diagram, the accumulated residual opacity at a point x on the image plane is Wherein, M is the number of Gaussians in the aqueous medium, which is 100 in this embodiment.

[0082] Furthermore, obtaining a 2D image of an underwater scene includes:

[0083] Inputting the 2D image of the Gaussian object, the 2D image of the water medium, and the medium transmission map into the image formation model for fusion to generate the 2D image of the underwater scene;

[0084] The 2D image of the Gaussian object is used as the unabsorbed radiation value, and the 2D image of the water medium is used as the forward scattered component.

[0085] Specifically, the object 2D image I obj As the unabsorbed radiation value, the water medium 2D image I med As the forward scattering component and the medium transmission map T med The input images form a model and are fused to generate a 2D image C of the underwater scene.

[0086] The image formation model is a physical model that simulates underwater image formation. The underwater image is represented by I c (x) = A(1-t(x)) + J c (x)t(x),c∈{r,g,b}, where A is the ambient light, t is the medium transmission map, and J is the unabsorbed radiation value. A(1-t(x)) represents the underwater forward scattering component, J c (x)t(x) represents the direct reflection component that reaches the camera after being absorbed by the medium. The two parts together constitute the underwater image I, c color channel, {r, g, b} are red, green, and blue color channels.

[0087] Furthermore, the trained Gaussian objects are obtained, including:

[0088] The background area of ​​the underwater scene 2D image and the enhanced underwater image is estimated using a background reconstruction algorithm, and the reconstruction of the background area is guided by enhancing the loss term of the background area to obtain the trained Gaussian object; wherein the entire training process undergoes 30,000 cycles of iteration.

[0089] Specifically, the 2D image of the underwater scene is compared to the real image to calculate the gradient of the learned parameters. Simultaneously, a background reconstruction algorithm is used to estimate the background area, and the loss term of the background area is enhanced to guide the reconstruction of the background area. The entire training process goes through 30,000 cycles of iteration.

[0090] The background reconstruction algorithm is used to estimate the background area and evaluate the similarity between the underwater scene 2D image C and the real image in the background object area. Each enhanced image I′ is converted into a grayscale image sequence I gray =RGB2Gray(I′), each grayscale image I gray Subdivided into multiple square local regions R with a side length of 32 pixels i,j =Subdivide(Igray ,32,(i,j)), where R i,j represents the local area of ​​row i and column j, Subdivide represents the subdivision operation, and 32 is the side length of the square. Calculate the variance of these local areas The local area with variance lower than the set threshold of 0.001 is identified as the background area to form a background mask Apply the background mask to I′ to extract the background light block B block,i,j =M bg,i,j I′ i,j , where I′ i,j The background light of the enhanced image is obtained by calculating the average value of all background light blocks. Among them, N is the number of background light blocks. Finally, the mean value of the background light of all enhanced images is taken as the background light value of the enhanced image sequence.

[0091] The loss function for background reconstruction is defined as Among them, ∈ is a regular term with a value of 0.001, C o is the oth pixel of the 2D image of the underwater scene, I o ′ is the i-th pixel in the enhanced image, To enhance the background light value of the image sequence, I′ is the enhanced image, n is the number of pixels, o is the oth pixel, is the loss function for background reconstruction.

[0092] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for underwater scene reconstruction and water medium separation based on a 3D Gaussian model, characterized in that: include: Constructing an underwater scene dataset, obtaining sparse point cloud data of the underwater scene, constructing a Gaussian object based on the sparse point cloud data, and initializing the Gaussian object; Select a fixed number of Gaussians with random means to simulate Gaussian water medium and initialize the Gaussian water medium; Using a differentiable 3D Gaussian sputtering rendering method, the Gaussian object and the Gaussian water medium are rendered respectively to obtain a 2D image of the Gaussian object and a 2D image of the Gaussian water medium, and output a medium transmission map of the water medium; fusing the 2D image of the Gaussian object, the 2D image of the Gaussian water medium, and the medium transmission map using an image formation model to obtain a 2D image of the underwater scene; The underwater scene 2D image and the enhanced underwater image are substituted into the loss function of the background reconstruction algorithm for training optimization to obtain a trained Gaussian object, and the trained Gaussian object is rendered along a custom directional perspective to output a 2D image of the underwater scene separated by the water medium.

2. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 1 is characterized in that: Constructing the underwater scene dataset includes: A monocular camera is used to capture underwater scenes. The camera moves around the target scene to capture continuous information of the scene, and each frame of the video is enhanced based on the color channel compensation method to obtain an enhanced underwater image sequence. The enhanced underwater image sequence is detected and matched using the motion structure recovery (SfM) method to reconstruct sparse point cloud data of the underwater scene, and the camera parameters are simultaneously extracted to obtain the underwater scene dataset.

3. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 2 is characterized in that: The color channel compensation method comprises: The average brightness values ​​of the blue, green and red color channels in the image are compared and analyzed. When the average brightness value of the blue channel is the highest, the blue channel will be used to compensate the green channel, and the green channel will be used to compensate the red channel; if the average brightness value of the green channel is the highest, the green channel will be used to compensate the blue and red channels; if the average brightness value of the red channel is the highest, the green channel will be used to compensate the blue channel, and the red channel will be used to compensate the green channel.

4. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 1 is characterized in that: Initializing the Gaussian object includes: The position coordinates of any sparse point cloud are set to the Gaussian mean of the Gaussian object, and the K nearest neighbor algorithm is used to obtain the average distance D = {d1, d2, ..., d n }; Set the Gaussian scaling parameters, rotation parameters, Gaussian opacity and Gaussian spherical harmonic coefficients respectively to obtain the initialized Gaussian object.

5. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 1 is characterized in that: Performing the Gaussian water medium initialization includes: The random numbers in the preset interval are used as the Gaussian mean of the Gaussian water medium, and the K-nearest neighbor algorithm is used to obtain the average distance between each Gaussian and the three nearest Gaussians. The scaling parameters, rotation parameters, opacity, training interval and spherical harmonic coefficients of the Gaussian are set respectively to obtain the initialized Gaussian water medium.

6. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 1 is characterized in that: Rendering the Gaussian object includes: According to the camera intrinsic parameters, camera extrinsic parameters and perspective projection function, the Gaussian projection of the Gaussian object is performed onto the image plane, and the 2D mean, 2D covariance matrix and volume density contribution of the Gaussian corresponding to the Gaussian object are obtained. The color value of the Gaussian corresponding point of the Gaussian object is rendered using the volume density formula to obtain a 2D image of the Gaussian object; Rendering the Gaussian water medium includes: According to the camera intrinsic parameters, camera extrinsic parameters and perspective projection function, the Gaussian of the Gaussian water medium is projected onto the image plane to obtain the 2D mean, 2D covariance matrix and volume density contribution of the Gaussian corresponding to the Gaussian water medium at a point on the image plane. The color value of the Gaussian corresponding point of the Gaussian water medium is rendered using the volume density formula to obtain a 2D image of the Gaussian water medium.

7. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 1 is characterized in that: Obtaining the 2D image of the underwater scene, including: Inputting the 2D image of the Gaussian object, the 2D image of the water medium, and the medium transmission map into the image formation model for fusion to generate the 2D image of the underwater scene; The 2D image of the Gaussian object is used as the unabsorbed radiation value, and the 2D image of the water medium is used as the forward scattered component.

8. The underwater scene reconstruction and water medium separation method based on 3D Gaussian model according to claim 1 is characterized in that: Obtaining the trained Gaussian object includes: The background area of ​​the underwater scene 2D image and the enhanced underwater image is estimated using a background reconstruction algorithm, and the reconstruction of the background area is guided by enhancing the loss term of the background area to obtain the trained Gaussian object; wherein the entire training process undergoes 30,000 cycles of iteration.

Images