Tidal flat three-dimensional rendering method and system based on neural radiation field

Abstract

The invention discloses a tidal flat three-dimensional rendering method and system based on a neural radiation field. The method comprises the following steps: acquiring a plurality of tidal flat environment two-dimensional images; obtaining camera pose information corresponding to each tidal flat environment two-dimensional image, wherein the camera pose information comprises a space coordinate and a visual angle direction; a random Fourier feature function is adopted to encode space coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image, and space coordinate codes are obtained; a spherical harmonic function is adopted to encode the view angle direction in the camera pose information corresponding to each tidal flat two-dimensional image, and view angle direction codes are obtained; the tidal flat three-dimensional reconstruction neural radiation field network is trained by adopting space coordinate codes and view angle directions, the input of the trained tidal flat three-dimensional reconstruction neural radiation field network is a certain view angle direction, and the output of the trained tidal flat three-dimensional reconstruction neural radiation field network is a tidal flat three-dimensional model comprising the color and density of each space point in the view angle direction; and carrying out visual display on the output tidal flat three-dimensional model.

Description

Technical Field

[0001] The present invention belongs to the technical field of tidal flat rendering, and specifically relates to a tidal flat three-dimensional rendering method and system based on a neural radiation field. Background Art

[0002] As an important part of the ecosystem, tidal flats have multiple ecological functions, such as maintaining biodiversity, protecting coastlines, and supporting fishery resources. With the impact of global climate change and human activities, the tidal flat environment is facing severe challenges, and the importance of its protection and management is becoming increasingly prominent. In order to effectively protect these precious ecological resources, it is crucial to monitor and study the tidal flat environment.

[0003] However, the current tidal flat protection technology mainly relies on large-scale and long-term monitoring methods. These technologies usually include the use of satellite images and aerial photography to observe and record the ecological and topographic changes of tidal flats from a macro perspective. The time span of this monitoring method is usually measured in years and the spatial span is measured in kilometers. It can cover a wide area of ​​tidal flats and provide valuable environmental change data for scientific research and policy making. Although this macro monitoring method can reveal long-term and large-scale environmental trends and provide a macro view of tidal flat environmental changes, it has obvious shortcomings in capturing fast-occurring local events and small ecological changes. For example, local pollution events, ecological degradation and sudden changes in specific species are often difficult to reflect in these large-scale monitoring data. In addition, these methods have long data collection and processing cycles and poor real-time performance, making it more difficult to respond quickly and take protective measures in emergency situations. In addition, due to the complexity and variability of tidal flat areas, the limitations of traditional monitoring technologies in refined management and real-time data processing are more prominent. This not only affects the timeliness and effectiveness of protection measures, but also limits the ability of scientific research to deeply understand tidal flat ecosystems. Summary of the invention

[0004] Purpose of the invention: In order to solve the limitations of traditional monitoring technology in refined management and real-time data processing, the present invention proposes a tidal flat three-dimensional rendering method and system based on neural radiation field. The present invention utilizes neural radiation field technology to capture complex lighting and reflection details, thereby generating highly realistic three-dimensional images. Although traditional monitoring technology has shown good results in long-term environmental monitoring and overall trend analysis, the present invention makes up for the gap in local detail observation and rapid response in tidal flat environmental monitoring.

[0005] Technical solution: A tidal flat three-dimensional rendering method based on neural radiation field, comprising the following steps:

[0006] Step 1: Acquire multiple two-dimensional images of tidal flat environment;

[0007] Step 2: Obtain camera pose information corresponding to each tidal flat environment two-dimensional image, wherein the camera pose information includes spatial coordinates and viewing direction;

[0008] Step 3: Use a random Fourier characteristic function to encode the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image to obtain a spatial coordinate code; use a spherical harmonic function to encode the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image to obtain a viewing direction code;

[0009] Step 4: Use spatial coordinate encoding and viewing direction encoding to train the neural radiation field network for tidal flat 3D reconstruction. The trained neural radiation field network for tidal flat 3D reconstruction takes spatial coordinate encoding and viewing direction encoding as input, and uses the color and density of the spatial points corresponding to each spatial coordinate in the viewing direction;

[0010] Step 5: Visualize the color and density of the spatial points output in step 4.

[0011] Furthermore, the colmap algorithm is used to obtain the camera pose information corresponding to each tidal flat environment 2D image.

[0012] Furthermore, the random Fourier characteristic function is used to encode the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image to obtain the spatial coordinate encoding, which specifically includes:

[0013] According to the following formula, the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image are encoded to obtain the spatial coordinate encoding:

[0014]

[0015] Where x represents the spatial coordinate X(x,y,z), w represents a random matrix sampled from a standard normal distribution, b represents an offset vector sampled from a uniform distribution [0,2π), D represents the dimension of the spatial coordinate, and M represents the number of Fourier features.

[0016] Furthermore, the use of spherical harmonic functions to encode the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image to obtain the viewing direction encoding specifically includes:

[0017] According to the following formula, the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image is encoded to obtain the viewing direction encoding:

[0018]

[0019]

[0020] in, is the coefficient, j is the parameter used to describe the order of the spherical harmonic function, m is the parameter used to describe the degree of the function at a given order j, m takes any integer value between -j and j, m = -j,...,-1,0,1,...,j.

[0021] Furthermore, the tidal flat three-dimensional reconstruction neural radiation field network includes: a fully connected layer linear(60, 256), a first fully connected layer linear(256, 256), a second fully connected layer linear(256, 256), a first fully connected layer linear(256, 128), a third fully connected layer linear(256, 256), a fourth fully connected layer linear(256, 256), a fifth fully connected layer linear(256, 256), a second fully connected layer linear(256, 128), a fully connected layer linear(128, 16) and a fully connected layer linear(60, 128); wherein the fully connected layer linear(60, 256), the first fully connected layer linear(256, 256), the first fully connected layer linear(256, 128), the third fully connected layer linear(256, 256), the fourth fully connected layer linear(256, 256), the fifth fully connected layer linear(256, 256), the second fully connected layer linear(256, 128), the fully connected layer linear(128, 16) and the fully connected layer linear(60, 128). ReLU activation function layers are provided between r(256, 256), the second fully connected layer linear(256, 256), the first fully connected layer linear(256, 128), the third fully connected layer linear(256, 256), the fourth fully connected layer linear(256, 256), the fifth fully connected layer linear(256, 256), the second fully connected layer linear(256, 128), and the fully connected layer linear(128, 16), and the input of the fully connected layer linear(60, 128) is connected to the input of the fully connected layer linear(60, 256), and the output of the fully connected layer linear(60, 128) is connected to the first fully connected layer linear(256, 128);

[0022] The inputs of the fully connected layer linear(60, 128) and the fully connected layer linear(60, 256) are both spatial coordinate encodings; the output of the fully connected layer linear(60, 128) is connected to the Sigmoid activation function layer;

[0023] The neural radiation field network for tidal flat 3D reconstruction also includes a fully connected layer linear(31,64), a fully connected layer linear(64,64), and a fully connected layer linear(64,3), and a ReLU activation function layer is arranged between the fully connected layer linear(31,64), the fully connected layer linear(64,64), and the fully connected layer linear(64,3); the output of the fully connected layer linear(64,3) is connected to a Sigmoid activation function layer;

[0024] The input of the fully connected layer linear(31,64) is the view direction encoding and the output of the fully connected layer linear(60,128).

[0025] Furthermore, the use of spatial coordinate encoding and viewing direction to train the neural radiation field network for tidal flat three-dimensional reconstruction specifically includes:

[0026] Generate ray r according to the view direction encoding, use the current tidal flat 3D reconstruction neural radiation field network to obtain the color c and density σ of all three-dimensional spaces on ray r, and obtain the estimated color on ray r according to the color c and density σ of all three-dimensional spaces on ray r

[0027]

[0028] Where N represents the number of spatial points on the ray, T i represents the transmittance at the i-th spatial point. The transmittance refers to the probability that the light from the light source to the spatial point is not absorbed. σ i represents the density of the i-th spatial point, δ i represents the distance from the i-th spatial point to the i+1-th spatial point, c i Represents the color at the i-th spatial point;

[0029] The training of the neural radiation field network for tidal flat 3D reconstruction is guided by the following loss function:

[0030]

[0031] in, represents the loss function, represents the set of rays, C(r) represents the true color on ray r, ||·|| 2 represents the two-norm.

[0032] The present invention discloses a tidal flat three-dimensional rendering system based on neural radiation field, comprising:

[0033] A tidal flat environment two-dimensional image acquisition module is used to acquire multiple tidal flat environment two-dimensional images;

[0034] A camera pose information acquisition module, used to acquire the camera pose information corresponding to each tidal flat environment two-dimensional image, wherein the camera pose information includes space coordinates and viewing direction;

[0035] The color and density output module of the spatial point is used to obtain the color and density of the spatial point corresponding to each spatial coordinate in the input viewing direction based on the tidal flat three-dimensional reconstruction neural radiation field network;

[0036] A visualization module is used to visualize the color and density of the spatial points output by the color and density output module of the spatial points;

[0037] The neural radiation field network based on tidal flat three-dimensional reconstruction is trained according to the following steps:

[0038] The spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image are encoded by using a random Fourier characteristic function to obtain a spatial coordinate encoding; the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image is encoded by using a spherical harmonic function to obtain a viewing direction encoding;

[0039] Spatial coordinate encoding and viewing direction encoding are used to train the neural radiation field network for tidal flat 3D reconstruction. The trained neural radiation field network for tidal flat 3D reconstruction takes spatial coordinate encoding and viewing direction encoding as input, and uses the color and density of the spatial points corresponding to each spatial coordinate in the viewing direction.

[0040] Furthermore, the random Fourier characteristic function is used to encode the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image to obtain the spatial coordinate encoding, which specifically includes:

[0041] According to the following formula, the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image are encoded to obtain the spatial coordinate encoding:

[0042]

[0043] Where x represents the spatial coordinate X(x,y,z), w represents a random matrix sampled from a standard normal distribution, b represents an offset vector sampled from a uniform distribution [0,2π), D represents the dimension of the spatial coordinate, and M represents the number of Fourier features;

[0044] The method of using spherical harmonic functions to encode the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image to obtain the viewing direction encoding specifically includes:

[0045] According to the following formula, the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image is encoded to obtain the viewing direction encoding:

[0046]

[0047] in, is the coefficient, j is the parameter used to describe the order of the spherical harmonic function, m is the parameter used to describe the degree of the function at a given order j, m takes any integer value between -j and j, m = -j,...,-1,0,1,...,j.

[0048] Furthermore, the tidal flat three-dimensional reconstruction neural radiation field network includes: a fully connected layer linear(60, 256), a first fully connected layer linear(256, 256), a second fully connected layer linear(256, 256), a first fully connected layer linear(256, 128), a third fully connected layer linear(256, 256), a fourth fully connected layer linear(256, 256), a fifth fully connected layer linear(256, 256), a second fully connected layer linear(256, 128), a fully connected layer linear(128, 16) and a fully connected layer linear(60, 128); wherein the fully connected layer linear(60, 256), the first fully connected layer linear(256, 256), the first fully connected layer linear(256, 128), the third fully connected layer linear(256, 256), the fourth fully connected layer linear(256, 256), the fifth fully connected layer linear(256, 256), the second fully connected layer linear(256, 128), the fully connected layer linear(128, 16) and the fully connected layer linear(60, 128). ReLU activation function layers are provided between r(256, 256), the second fully connected layer linear(256, 256), the first fully connected layer linear(256, 128), the third fully connected layer linear(256, 256), the fourth fully connected layer linear(256, 256), the fifth fully connected layer linear(256, 256), the second fully connected layer linear(256, 128), and the fully connected layer linear(128, 16), and the input of the fully connected layer linear(60, 128) is connected to the input of the fully connected layer linear(60, 256), and the output of the fully connected layer linear(60, 128) is connected to the first fully connected layer linear(256, 128);

[0049] The inputs of the fully connected layer linear(60, 128) and the fully connected layer linear(60, 256) are both spatial coordinate encodings; the output of the fully connected layer linear(60, 128) is connected to the Sigmoid activation function layer;

[0050] The neural radiation field network for tidal flat 3D reconstruction also includes a fully connected layer linear(31,64), a fully connected layer linear(64,64), and a fully connected layer linear(64,3), and a ReLU activation function layer is arranged between the fully connected layer linear(31,64), the fully connected layer linear(64,64), and the fully connected layer linear(64,3); the output of the fully connected layer linear(64,3) is connected to a Sigmoid activation function layer;

[0051] The input of the fully connected layer linear(31,64) is the view direction encoding and the output of the fully connected layer linear(60,128).

[0052] Furthermore, the spatial coordinate encoding and viewing direction are used to train the neural radiation field network for tidal flat 3D reconstruction, including:

[0053] Generate ray r according to the view direction encoding, use the current tidal flat 3D reconstruction neural radiation field network to obtain the color c and density σ of all three-dimensional spaces on ray r, and obtain the estimated color on ray r according to the color c and density σ of all three-dimensional spaces on ray r

[0054]

[0055] Where N represents the number of spatial points on the ray, T i represents the transmittance at the i-th spatial point. The transmittance refers to the probability that the light from the light source to the spatial point is not absorbed. σ i represents the density of the i-th spatial point, δ i represents the distance from the i-th spatial point to the i+1-th spatial point, c i Represents the color at the i-th spatial point;

[0056] The training of the neural radiation field network for tidal flat 3D reconstruction is guided by the following loss function:

[0057]

[0058] in, represents the loss function, represents the ray set, C(r) represents the true color on ray r, ‖·‖ 2 represents the two-norm.

[0059] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0060] (1) The present invention uses Neural Radiance Field (NeRF) technology to reconstruct a 3D environment by training a neural network. Neural Radiance Field (NeRF) technology can capture complex lighting and reflection details to generate highly realistic 3D images.

[0061] (2) The present invention efficiently encodes spatial coordinates and viewing angle information through random Fourier characteristic functions and spherical harmonic functions. This encoding method significantly improves the data processing speed and the training efficiency of the neural network, making the reconstruction process of the three-dimensional environment not only more accurate but also faster. The application of this encoding technology not only optimizes the processing of data input, but also improves the model's sensitivity to spatial and viewing angle changes, allowing the neural network model to more accurately predict the color and density of each three-dimensional space point. Therefore, the three-dimensional image finally generated is more visually realistic, with stronger detail expression, and can better meet the needs of high-quality rendering.

[0062] (3) The present invention also proposes the development of a matching web interactive interface, so that the smooth three-dimensional model display and interactive functions implemented on the web client provide users with an immersive visual experience, and support complex user interactions, such as perspective conversion, zooming and custom rendering settings, providing users with a high degree of interactivity; these functions greatly improve the convenience and experience of users operating three-dimensional models in a web environment;

[0063] (4) The smooth 3D model display and interaction functions implemented by the present invention on the Web client can not only be loaded and displayed in a browser without any plug-ins, but also support users to perform perspective conversion, zooming and other interactions through simple mouse operations, making viewing and analyzing 3D scenes more intuitive and convenient. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flowchart of the steps of the tidal flat three-dimensional rendering method based on neural radiation field proposed by the present invention;

[0065] Figure 2 This is a schematic diagram of the NeRF network for 3D reconstruction of tidal flats proposed in the present invention;

[0066] Figure 3 The present invention is a schematic diagram of viewing angle conversion and area enlargement display by mouse operation on a Web client. DETAILED DESCRIPTION

[0067] The technical solution of the present invention is now further described in conjunction with the accompanying drawings.

[0068] Embodiment 1:

[0069] like Figure 1 As shown, this embodiment discloses a tidal flat three-dimensional rendering method based on a neural radiation field, which mainly includes the following steps:

[0070] Step 1: Use a drone equipped with a high-definition camera to take all-round photos of the tidal flat environment and obtain several two-dimensional images of the tidal flat environment.

[0071] Step 2: Use the colmap algorithm to calculate the acquired tidal flat environment two-dimensional images and calculate the camera pose information corresponding to each tidal flat two-dimensional image. The camera pose information includes the spatial coordinates X (x, y, z) and the corresponding viewing direction.

[0072] Step 3: Use the random Fourier characteristic function to encode the spatial coordinates X(x, y, z) in the camera pose information corresponding to each tidal flat 2D image. The specific process is as follows:

[0073]

[0074] In the formula, x represents the spatial coordinate X(x,y,z), w represents a random matrix sampled from a standard normal distribution, b represents an offset vector sampled from a uniform distribution [0,2π), D represents the dimension of the spatial coordinate, and M represents the number of Fourier features, that is, the dimension of the encoded high-dimensional feature space. This number determines how many cosine features are generated.

[0075] Step 4: Use spherical harmonics to calculate the viewing direction in the camera pose information corresponding to each tidal flat 2D image Encoding, the specific process is as follows:

[0076]

[0077]

[0078] in, The parameter j is a coefficient, and describes the "order" of the spherical harmonic function, representing the number of ripples of the function on the sphere. The value is a non-negative integer, j = 0, 1, 2, ..., J. In this embodiment, J is 3. The parameter m describes the "degree" of the function under a given order j, representing the change of the function in the azimuth angle. m takes any integer value between -j and j, m = -j, ..., -1, 0, 1, ..., j.

[0079] Step 5: Build the NeRF network for tidal flat 3D reconstruction, such as Figure 2As shown, it includes: multiple fully connected layers and activation functions; wherein the activation functions mainly use ReLU and Sigmoid. Specifically, it includes: fully connected layer linear(60,256), the first fully connected layer linear(256,256), the second fully connected layer linear(256,256), the first fully connected layer linear(256,128), the third fully connected layer linear(256,256), the fourth fully connected layer linear(256,256), the fifth fully connected layer linear(256,256), the second fully connected layer linear(256,128), the fully connected layer linear(128,16) and the fully connected layer linear(60,128); wherein the fully connected layer linear(60,256), the first fully connected layer linear(256,256) 6), ReLU activation function layers are provided between the second fully connected layer linear(256, 256), the first fully connected layer linear(256, 128), the third fully connected layer linear(256, 256), the fourth fully connected layer linear(256, 256), the fifth fully connected layer linear(256, 256), the second fully connected layer linear(256, 128), and the fully connected layer linear(128, 16), and the input of the fully connected layer linear(60, 128) is connected to the input of the fully connected layer linear(60, 256), and the output of the fully connected layer linear(60, 128) is connected to the first fully connected layer linear(256, 128);

[0080] The inputs of the fully connected layer linear(60, 128) and the fully connected layer linear(60, 256) are both spatial coordinate encodings; the output of the fully connected layer linear(60, 128) is connected to the Sigmoid activation function layer;

[0081] The NeRF network for tidal flat 3D reconstruction also includes a fully connected layer linear(31,64), a fully connected layer linear(64,64), and a fully connected layer linear(64,3), and a ReLU activation function layer is arranged between the fully connected layer linear(31,64), the fully connected layer linear(64,64), and the fully connected layer linear(64,3); the output of the fully connected layer linear(64,3) is connected to a Sigmoid activation function layer;

[0082] The input of the fully connected layer linear(31,64) is the view direction encoding and the output of the fully connected layer linear(60,128).

[0083] Step 6: Use spatial coordinate encoding and viewing direction encoding to train the tidal flat 3D reconstruction neural radiation field network to obtain the trained tidal flat 3D reconstruction NeRF network. The specific operations include:

[0084] Generate ray r according to the encoding result of the viewing direction, use the current tidal flat 3D reconstruction neural radiation field network to obtain the color c and density σ of all three-dimensional spaces on ray r, and estimate the color in the sight direction according to the color c and density σ of all three-dimensional spaces on ray r, that is, the estimated color on ray r

[0085]

[0086] in, represents the estimated color on ray r, N represents the number of spatial points on the ray, T i represents the transmittance at the i-th spatial point, which indicates the probability that the light from the light source to this spatial point is not absorbed, σ i represents the density of the i-th spatial point, δ i represents the distance from the i-th spatial point to the i+1-th spatial point, c i Represents the color at the i-th spatial point.

[0087] The training of the NeRF network for tidal flat 3D reconstruction is guided by the following loss function:

[0088]

[0089] in, represents the loss function, represents the ray set, i.e., all rays sampled during the training process, C(r) represents the true color on ray r, ||·|| 2 It represents the L2 norm, i.e. the Euclidean distance.

[0090] This loss function is used to guide the training of the tidal flat 3D reconstruction NeRF network, making the images generated by the tidal flat 3D reconstruction NeRF network closer to the real images.

[0091] After training, the tidal flat environment can be recorded by the tidal flat 3D reconstruction NeRF network.

[0092] Step 7: Input the spatial coordinate encoding and the viewing direction encoding into the trained tidal flat 3D reconstruction NeRF network. The tidal flat 3D reconstruction NeRF network can generate a tidal flat 3D model including the color c and density σ of the 3D space point, and further obtain a realistic tidal flat scene image through the rendering formula, thereby achieving the purpose of observing the tidal flat from different perspectives. Specifically, Figure 3As shown, the tidal flat 3D model is exported to the 3D rendering engine, processed in the 3D rendering engine, and finally displayed on the Web client after integration. Then, scene construction and page development are carried out to ensure the smooth display and interaction of the 3D view on the web page, and support users to switch perspectives and zoom in on areas through mouse operations.

Claims

1. A tidal flat three-dimensional rendering method based on neural radiation field, characterized by: The following steps are involved: Step 1: Acquire multiple two-dimensional images of tidal flat environment; Step 2: Obtain camera pose information corresponding to each tidal flat environment two-dimensional image, wherein the camera pose information includes spatial coordinates and viewing direction; Step 3: Use a random Fourier characteristic function to encode the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image to obtain a spatial coordinate code; use a spherical harmonic function to encode the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image to obtain a viewing direction code; Step 4: Use spatial coordinate encoding and viewing direction encoding to train the neural radiation field network for tidal flat 3D reconstruction. The trained neural radiation field network for tidal flat 3D reconstruction takes spatial coordinate encoding and viewing direction encoding as input, and uses the color and density of the spatial points corresponding to each spatial coordinate in the viewing direction; Step 5: Visualize the color and density of the spatial points output in step 4.

2. The method for three-dimensional rendering of tidal flats based on neural radiation fields according to claim 1, characterized in that: The colmap algorithm is used to obtain the camera pose information corresponding to each tidal flat environment two-dimensional image.

3. The method for three-dimensional rendering of tidal flats based on neural radiation fields according to claim 1, characterized in that: The method of using a random Fourier characteristic function to encode the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image to obtain the spatial coordinate encoding specifically includes: According to the following formula, the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image are encoded to obtain the spatial coordinate encoding: Where x represents the spatial coordinate X(x,y,z), w represents a random matrix sampled from a standard normal distribution, b represents an offset vector sampled from a uniform distribution [0,2π), D represents the dimension of the spatial coordinate, and M represents the number of Fourier features.

4. The method for three-dimensional rendering of tidal flats based on neural radiation fields according to claim 1, characterized in that: The method of using spherical harmonic functions to encode the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image to obtain the viewing direction encoding specifically includes: According to the following formula, the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image is encoded to obtain the viewing direction encoding: in, is the coefficient, j is the parameter used to describe the order of the spherical harmonic function, m is the parameter used to describe the degree of the function at a given order j, m takes any integer value between -j and j, m = -j,...,-1,0,1,...,j.

5. The method for three-dimensional rendering of tidal flats based on neural radiation fields according to claim 1, characterized in that: The neural radiation field network for tidal flat three-dimensional reconstruction includes: a fully connected layer linear (60, 256), a first fully connected layer linear (256, 256), a second fully connected layer linear (256, 256), a first fully connected layer linear (256, 128), a third fully connected layer linear (256, 256), a fourth fully connected layer linear (256, 256), a fifth fully connected layer linear (256, 256), a second fully connected layer linear (256, 128), a fully connected layer linear (128, 16) and a fully connected layer linear (60, 128); wherein the fully connected layer linear (60, 256), the first fully connected layer linear (2 ReLU activation function layers are provided between the second fully connected layer linear(56,256), the second fully connected layer linear(256,256), the first fully connected layer linear(256,128), the third fully connected layer linear(256,256), the fourth fully connected layer linear(256,256), the fifth fully connected layer linear(256,256), the second fully connected layer linear(256,128), and the fully connected layer linear(128,16), and the input of the fully connected layer linear(60,128) is connected to the input of the fully connected layer linear(60,256), and the output of the fully connected layer linear(60,128) is connected to the first fully connected layer linear(256,128); The inputs of the fully connected layer linear(60, 128) and the fully connected layer linear(60, 256) are both spatial coordinate encodings; the output of the fully connected layer linear(60, 128) is connected to the Sigmoid activation function layer; The neural radiation field network for tidal flat 3D reconstruction also includes a fully connected layer linear(31,64), a fully connected layer linear(64,64), and a fully connected layer linear(64,3), and a ReLU activation function layer is arranged between the fully connected layer linear(31,64), the fully connected layer linear(64,64), and the fully connected layer linear(64,3); the output of the fully connected layer linear(64,3) is connected to a Sigmoid activation function layer; The input of the fully connected layer linear(31,64) is the view direction encoding and the output of the fully connected layer linear(60,128).

6. The method for three-dimensional rendering of tidal flats based on neural radiation fields according to claim 1, characterized in that: The method of using spatial coordinate coding and viewing direction coding to train the neural radiation field network for tidal flat three-dimensional reconstruction specifically includes: Generate ray r according to the view direction encoding, use the current tidal flat 3D reconstruction neural radiation field network to obtain the color c and density σ of all three-dimensional spaces on ray r, and obtain the estimated color on ray r according to the color c and density σ of all three-dimensional spaces on ray r Where N represents the number of spatial points on the ray, T i represents the transmittance at the i-th spatial point. The transmittance refers to the probability that the light from the light source to the spatial point is not absorbed. σ i represents the density of the i-th spatial point, δ i represents the distance from the ith spatial point to the i+1th spatial point, c i Represents the color at the i-th spatial point; The training of the neural radiation field network for tidal flat 3D reconstruction is guided by the following loss function: in, represents the loss function, represents the set of rays, C(r) represents the true color on ray r, ||·|| 2 represents the two-norm.

7. A tidal flat three-dimensional rendering system based on neural radiation field, characterized by: include: A tidal flat environment two-dimensional image acquisition module is used to acquire multiple tidal flat environment two-dimensional images; A camera pose information acquisition module, used to acquire the camera pose information corresponding to each tidal flat environment two-dimensional image, wherein the camera pose information includes space coordinates and viewing direction; The color and density output module of the spatial point is used to reconstruct the neural radiation field network based on the tidal flat three-dimensionally, and obtain the color and density of the spatial point corresponding to each spatial coordinate in the input viewing direction; A visualization module is used to visualize the color and density of the spatial points output by the color and density output module of the spatial points; The neural radiation field network based on tidal flat three-dimensional reconstruction is trained according to the following steps: The spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image are encoded by using a random Fourier characteristic function to obtain a spatial coordinate encoding; the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image is encoded by using a spherical harmonic function to obtain a viewing direction encoding; Spatial coordinate encoding and viewing direction encoding are used to train the neural radiation field network for tidal flat 3D reconstruction. The trained neural radiation field network for tidal flat 3D reconstruction takes spatial coordinate encoding and viewing direction encoding as input, and uses the color and density of the spatial points corresponding to each spatial coordinate in the viewing direction.

8. The tidal flat three-dimensional rendering system based on neural radiation field according to claim 7, characterized in that: The method of using a random Fourier characteristic function to encode the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image to obtain the spatial coordinate encoding specifically includes: According to the following formula, the spatial coordinates in the camera pose information corresponding to each tidal flat environment two-dimensional image are encoded to obtain the spatial coordinate encoding: Where x represents the spatial coordinate X(x,y,z), w represents a random matrix sampled from a standard normal distribution, b represents an offset vector sampled from a uniform distribution [0,2π), D represents the dimension of the spatial coordinate, and M represents the number of Fourier features; The method of using spherical harmonic functions to encode the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image to obtain the viewing direction encoding specifically includes: According to the following formula, the viewing direction in the camera pose information corresponding to each tidal flat two-dimensional image is encoded to obtain the viewing direction encoding: in, is the coefficient, j is the parameter used to describe the order of the spherical harmonic function, m is the parameter used to describe the degree of the function at a given order j, m takes any integer value between -j and j, m = -j,...,-1,0,1,...,j.

9. The tidal flat three-dimensional rendering system based on neural radiation field according to claim 7, characterized in that: The neural radiation field network for tidal flat three-dimensional reconstruction includes: a fully connected layer linear (60, 256), a first fully connected layer linear (256, 256), a second fully connected layer linear (256, 256), a first fully connected layer linear (256, 128), a third fully connected layer linear (256, 256), a fourth fully connected layer linear (256, 256), a fifth fully connected layer linear (256, 256), a second fully connected layer linear (256, 128), a fully connected layer linear (128, 16) and a fully connected layer linear (60, 128); wherein the fully connected layer linear (60, 256), the first fully connected layer linear (2 ReLU activation function layers are provided between the second fully connected layer linear(56,256), the second fully connected layer linear(256,256), the first fully connected layer linear(256,128), the third fully connected layer linear(256,256), the fourth fully connected layer linear(256,256), the fifth fully connected layer linear(256,256), the second fully connected layer linear(256,128), and the fully connected layer linear(128,16), and the input of the fully connected layer linear(60,128) is connected to the input of the fully connected layer linear(60,256), and the output of the fully connected layer linear(60,128) is connected to the first fully connected layer linear(256,128); The inputs of the fully connected layer linear(60, 128) and the fully connected layer linear(60, 256) are both spatial coordinate encodings; the output of the fully connected layer linear(60, 128) is connected to the Sigmoid activation function layer; The neural radiation field network for tidal flat 3D reconstruction also includes a fully connected layer linear(31,64), a fully connected layer linear(64,64), and a fully connected layer linear(64,3), and a ReLU activation function layer is arranged between the fully connected layer linear(31,64), the fully connected layer linear(64,64), and the fully connected layer linear(64,3); the output of the fully connected layer linear(64,3) is connected to a Sigmoid activation function layer; The input of the fully connected layer linear(31,64) is the view direction encoding and the output of the fully connected layer linear(60,128).

10. The tidal flat three-dimensional rendering system based on neural radiation field according to claim 7, characterized in that: The neural radiation field network for tidal flat 3D reconstruction is trained using spatial coordinate encoding and viewing direction, including: Generate ray r according to the view direction encoding, use the current tidal flat 3D reconstruction neural radiation field network to obtain the color c and density σ of all three-dimensional spaces on ray r, and obtain the estimated color on ray r according to the color c and density σ of all three-dimensional spaces on ray r Where N represents the number of spatial points on the ray, T i represents the transmittance at the i-th spatial point. The transmittance refers to the probability that the light from the light source to the spatial point is not absorbed. σ i represents the density of the i-th spatial point, δ i represents the distance from the ith spatial point to the i+1th spatial point, c i Represents the color at the i-th spatial point; The training of the neural radiation field network for tidal flat 3D reconstruction is guided by the following loss function: in, represents the loss function, represents the ray set, C(r) represents the true color on ray r, ‖·‖ 2 represents the two-norm.

Images