3DGS improved algorithm based on attenuation-growth adaptive strategy

Abstract

The invention discloses a 3DGS improved algorithm based on an attenuation-growth adaptive strategy, and the algorithm comprises the following steps: image initialization: extracting and matching feature points from a multi-view image through COLMAP to recover a 3D point cloud, and initializing a Gaussian set; self-adaptive density control: training the model after a sparsely reconstructed Gaussian ball is obtained, and realizing transition from an initial sparse Gaussian set to a denser set; adaptive frequency control is carried out, frequency parameters act on the size of a Gaussian ball, the Gaussian frequency can be approximately adjusted, and dynamic adjustment can be carried out according to regional information; adaptive strategy adjustment is carried out, a staged adaptive strategy adjustment method is adopted, and cloning-separation and attenuation-growth strategies are systematically applied according to gradient conditions of different stages; and performing Gaussian splash rendering, and updating corresponding training parameters by using Gaussian rendering splash. According to the 3DGS improved algorithm based on the attenuation-growth adaptive strategy, efficient and high-quality three-dimensional scene reconstruction is achieved by dynamically adjusting the number and frequency of Gaussian balls.

Description

Technical Field

[0001] The present invention relates to the technical field of visual reconstruction, and in particular to an improved 3DGS algorithm based on a decay-growth adaptive strategy. Background Art

[0002] 3DGS: 3DGS represents the scene as a series of Gaussian balls with color and opacity attributes. Each Gaussian ball is described by position, color, opacity, and covariance matrix. Through the combination of these Gaussian balls, complex three-dimensional scenes can be rendered efficiently. The rendering process of 3DGS uses the projected 2D Gaussian ball to calculate the color of each pixel. The rendering formula is as follows:

[0003]

[0004] in:

[0005] K: represents the total number of Gaussian balls;

[0006] c k : The color of the kth Gaussian ball;

[0007] α k : The transparency of the kth Gaussian sphere;

[0008] Gaussian distribution value of the 2D projection of the k-th Gaussian ball at position x;

[0009] The cumulative opacity term is the transmittance of all Gaussian spheres before the kth Gaussian sphere.

[0010] By using a series of Gaussian spheres to represent the geometric and texture information in the scene, 3DGS can flexibly adapt to geometric structures of different scales and shapes. However, the traditional 3DGS method still has some limitations when dealing with complex scenes, especially in the number and distribution optimization of Gaussian spheres, which affects the efficiency and quality of reconstruction. Summary of the invention

[0011] The object of the present invention is to provide an improved 3DGS algorithm based on a decay-growth adaptive strategy, which achieves efficient and high-quality three-dimensional scene reconstruction by dynamically adjusting the number and frequency of Gaussian spheres.

[0012] The technical solution adopted by the 3DGS improved algorithm based on the decay-growth adaptive strategy disclosed in the present invention is:

[0013] An improved 3DGS algorithm based on a decay-growth adaptive strategy comprises the following steps:

[0014] Image initialization: obtain image video frames through the shooting device, extract and match feature points from multi-view images through COLMAP to restore the 3D point cloud, and initialize the Gaussian set;

[0015] Adaptive density control: After obtaining the sparsely reconstructed Gaussian sphere, the model can be trained. Since the initial number of Gaussians is small, adaptive density control is needed to adjust the number and density of Gaussians, so as to achieve the transition from the initial sparse Gaussian set to a denser set.

[0016] Adaptive frequency control: By applying the frequency parameter to the size of the Gaussian ball, the frequency of the Gaussian can be adjusted approximately. This can be adjusted dynamically based on regional information without relying on repeated training.

[0017] Adaptive strategy adjustment. During the training process, in order to gradually optimize the number and density distribution of Gaussian balls, a phased adaptive strategy adjustment method is adopted. According to the gradient conditions at different stages, the clone-separation and decay-growth strategies are systematically applied to more effectively control the distribution density of Gaussian balls.

[0018] Gaussian splash rendering, uses Gaussian rendering splash to update the corresponding training parameters.

[0019] As a preferred solution, the specific steps of image initialization are as follows:

[0020] Image import: obtain multi-view image data to be tested;

[0021] Feature matching: Use SI FT feature extraction algorithm to match the extracted feature points;

[0022] Sparse reconstruction: After feature matching, the position of 3D points is calculated by triangulation, and the camera parameters and 3D point positions are optimized in combination with the Bundle Adjustment algorithm to obtain a sparse point cloud.

[0023] Gaussian ball parameter initialization: obtain a sparse point set and initialize the Gaussian ball parameter set

[0024]

[0025] in:

[0026] μ i : represents the center position of the Gaussian sphere, that is, the position of the i-th 3D point extracted from the sparse point cloud;

[0027] sh i : represents the color of the Gaussian sphere, expressed using spherical harmonic sparse expression;

[0028] r i: represents the rotation matrix of the Gaussian ball, which is used to control the direction of the Gaussian in three-dimensional space;

[0029] s i : represents the size parameter of the Gaussian ball, which is used to control the size or range of the Gaussian;

[0030] α i : Indicates the opacity of the Gaussian sphere, which is used to control the effect of the Gaussian on the image during the rendering process;

[0031] β i ,γ i : Indicates the parameters that control the frequency adjustment of the Gaussian sphere;

[0032] N: represents the number of Gaussian balls;

[0033] Note: β i ,γ i are untrained parameters, and the rest are training parameters.

[0034] As a preferred solution, the adaptive density control determines whether cloning or segmentation operations are required through the following gradient and scale conditions:

[0035] Gradient condition: If the gradient vector modulus of a Gaussian sphere ∥g i ∥Exceeds the specified threshold τ d , then the area needs more Gaussian spheres to improve detail coverage;

[0036] ∥g i ∥≥τ d

[0037] Scale condition: If the maximum scale of a Gaussian ball max(s i ) is less than, equal to, or greater than the product of the scene extent and the scale factor;

[0038] max(s i )≤p d ·E

[0039] max(s i )>p d ·E

[0040] Based on the gradient and scale conditions, adaptive density control defines the following cloning and segmentation conditions:

[0041] Cloning conditions:

[0042] mask i =(∥g i ∥≥τ d )∧(max(s i )≤p d ·E)

[0043] Separation conditions:

[0044] mask i =(∥g i ∥≥τ g )∧(max(s i )>p d ·E)

[0045] in:

[0046] g i :The loss function for u i The gradient vector of

[0047] τ d : Density adjustment threshold;

[0048] s i : size vector;

[0049] E: scene range;

[0050] p d : Densification ratio;

[0051] mask i : A Gaussian sphere that satisfies the conditions.

[0052] As a preferred solution, the loss function is:

[0053]

[0054] This loss function combines L1 loss and SSIM loss, and controls the relative balance between the two through weight parameters;

[0055] in:

[0056] λ: The ratio used to balance the L1 and DSSIM losses.

[0057] represents L1 loss, which calculates the absolute value of the pixel difference between the predicted image and the real image;

[0058] Represents the structural similarity index, which is used to measure the similarity of image structures.

[0059] As a preferred solution, the specific steps of the adaptive frequency control are:

[0060] β i =f(β i )*γ i

[0061] r i =(r g or r d )

[0062] γ i =γ i *r i

[0063] s i =s i *β i

[0064] f(x) = ReLU(2.0*sigmoid(x))

[0065] in:

[0066] β i : Frequency control parameter, used to adjust the frequency of the Gaussian ball, initialized to 0;

[0067] r i : Adjustment factor, select growth according to regional information (r g ) or attenuation (r d );

[0068] γ i : Adaptive density coefficient, initialized to 1; 0;

[0069] f(x): frequency control function, implemented using a combination of ReLU and sigmoid functions;

[0070] Adaptive density control determines whether cloning or splitting is required based on the following gradient and scale conditions:

[0071] Gradient condition: If the gradient vector modulus of a Gaussian sphere ∥g i ∥Exceeds the specified threshold τ f , then the area needs further adjustment to improve detail coverage,

[0072] ∥g i ∥≥τ f

[0073] Scale condition: If the maximum scale of a Gaussian ball max(s i ) is less than, equal to, or greater than the product of the scene range and the scale factor,

[0074] max(s i )≤p f ·E

[0075] max(s i )>p f ·E

[0076] Based on the gradient and scale conditions, adaptive density control defines the following cloning and segmentation conditions:

[0077] Growing Conditions:

[0078] mask i =(∥g i ∥≥τ f )∧(max(s i )≤p f ·E)

[0079] Attenuation conditions:

[0080] mask i =(∥g i ∥≥τ f )∧(max(s i )>p f ·E)

[0081] in:

[0082] τ f : Indicates the frequency adjustment threshold;

[0083] p f : represents the frequency ratio;

[0084] This frequency control method allows the number and density of Gaussian spheres to be dynamically adjusted according to the needs of different regions, thereby ensuring the details of regional features while avoiding redundancy.

[0085] As a preferred solution, the entire training process of the adaptive strategy adjustment is divided into three stages: ep1, ep2 and ep3. The specific strategies are as follows:

[0086] Initial stage 0~ep1

[0087] Strategy: Only clone-separation strategy is applied;

[0088] Goal: Quickly establish a sparse initial distribution of Gaussian spheres to cover the main structures of the scene;

[0089] In this stage, the overall density is increased by cloning and separating Gaussian balls, quickly filling the basic outline of the scene and ensuring initial coverage in key areas. This strategy can effectively improve the density of the initial structure in a short period of time, laying the foundation for subsequent refinement;

[0090] Intermediate stage ep1~ep2

[0091] Strategy: Adaptively select clone-separation or decay-growth strategy according to gradient conditions;

[0092] Goal: Increase density or function frequency in detail-rich areas while maintaining sparsity in flat areas to achieve efficient distribution adjustment;

[0093] When the gradient is large, the clone-separation strategy is used first to increase the density. When the gradient is small, the decay-growth strategy is used to save resources by further adjusting the frequency of the Gaussian balls to reduce redundant Gaussian balls.

[0094] Convergence phase ep2~ep3

[0095] Strategy: Only the decay-growth strategy is applied;

[0096] Objective: Further optimize the model to make the Gaussian sphere distribution more consistent with scene requirements, reduce redundancy and improve the accuracy of detail reconstruction;

[0097] After the model gradually converges, the density of the Gaussian spheres is mainly optimized: in flat areas, the attenuation strategy is used to reduce unnecessary Gaussian spheres and release computing resources; in key detail areas, the growth strategy is used to further increase the density to make the reconstruction results more detailed and accurate. The adjustments at this stage make the Gaussian sphere distribution more suitable for the scene, and ultimately achieve accurate convergence of the model.

[0098] As a preferred solution, the probability density function formula of the three-dimensional Gaussian distribution is as follows:

[0099]

[0100] in:

[0101] n: represents the dimension of the random variable;

[0102] x: represents a column vector containing n elements, which is the observed value of the random variable;

[0103] μ: represents a column vector containing n elements, which is the mean of the distribution;

[0104] Σ: represents an n-order square matrix, which is the covariance matrix;

[0105] The Gaussian splash rendering formula is as follows:

[0106]

[0107] in:

[0108] K: represents the total number of Gaussian balls;

[0109] c k : The color of the kth Gaussian ball;

[0110] α k : The transparency of the kth Gaussian sphere;

[0111] Gaussian distribution value of the 2D projection of the k-th Gaussian ball at position x;

[0112] The cumulative opacity term, that is, the transmittance of all Gaussian balls before the kth Gaussian ball, represents the influence of the previous Gaussian balls on the current Gaussian ball.

[0113] The beneficial effects of an improved 3DGS algorithm based on a decay-growth adaptive strategy disclosed in the present invention are: sparse point clouds are initialized through COLMAP, and the density and frequency of the initialized Gaussian sphere are adjusted by adaptive density control and adaptive frequency control. Finally, corresponding training parameters are updated by Gaussian rendering splashing. By combining 3DGS and the decay-growth strategy, the reconstruction effect of the model on image details and overall structure is improved, which can meet the industry's demand for high-quality visual reconstruction, bring clearer and more natural images and video content, effectively improve the performance of intelligent visual systems, and improve user experience, especially in scenes that require realism and clarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 The present invention is a flowchart of an improved 3DGS algorithm based on a decay-growth adaptive strategy. DETAILED DESCRIPTION

[0115] The present invention will be further described and illustrated below in conjunction with specific embodiments and accompanying drawings:

[0116] Please refer to Figure 1 , an improved 3DGS algorithm based on decay-growth adaptive strategy, comprising the following steps:

[0117] Image initialization

[0118] Obtain image video frames through the shooting device, extract and match feature points from multi-view images through COLMAP to restore the 3D point cloud, and initialize the Gaussian set;

[0119] The image video frame is obtained through cameras such as monocular, binocular, RGBD, etc. The specific steps of image initialization are as follows:

[0120] Image import: obtain multi-view image data to be tested;

[0121] Feature matching: Use SI FT feature extraction algorithm to match the extracted feature points;

[0122] Sparse reconstruction: After feature matching, the position of 3D points is calculated by triangulation, and the camera parameters and 3D point positions are optimized in combination with the Bundle Adjustment algorithm to obtain a sparse point cloud.

[0123] Gaussian ball parameter initialization: obtain a sparse point set and initialize the Gaussian ball parameter set

[0124]

[0125] in:

[0126] μ i : represents the center position of the Gaussian sphere, that is, the position of the i-th 3D point extracted from the sparse point cloud; sh i : represents the color of the Gaussian sphere, expressed using spherical harmonic sparse expression;

[0127] r i : represents the rotation matrix of the Gaussian ball, which is used to control the direction of the Gaussian in three-dimensional space;

[0128] s i : represents the size parameter of the Gaussian ball, which is used to control the size or range of the Gaussian;

[0129] α i : Indicates the opacity of the Gaussian sphere, which is used to control the effect of the Gaussian on the image during rendering; β i ,γ i : Indicates the parameters that control the frequency adjustment of the Gaussian sphere;

[0130] N: represents the number of Gaussian balls;

[0131] Note: β i ,γ i are untrained parameters, and the rest are training parameters

[0132] Adaptive Density Control

[0133] After obtaining the sparsely reconstructed Gaussian sphere, the model can be trained. Since the initial number of Gaussians is small, the number and density of Gaussians need to be adjusted through adaptive density control to achieve the transition from the initial sparse Gaussian set to a denser set.

[0134] Adaptive density control focuses on two key areas:

[0135] Under-reconstructed regions: These are regions that lack sufficient geometric features. For these regions, adaptive density adjusts the Gaussian through a cloning operation to cover the newly created geometry.

[0136] Over-reconstructed areas: These are areas where the Gaussian spheres are too dense or too large in size. Adaptive Density adjusts the Gaussian through a cloning operation to cover the newly created geometry.

[0137] Adaptive density control determines whether cloning or splitting is required based on the following gradient and scale conditions:

[0138] Gradient condition: If the gradient vector modulus of a Gaussian sphere ∥g i ∥Exceeds the specified threshold τ d, then the area needs more Gaussian spheres to improve detail coverage;

[0139] ∥g i ∥≥τ d

[0140] Scale condition: If the maximum scale of a Gaussian ball max(s i ) is less than, equal to, or greater than the product of the scene extent and the scale factor;

[0141] max(s i )≤p d ·E

[0142] max(s i )>p d ·E

[0143] Based on the gradient and scale conditions, adaptive density control defines the following cloning and segmentation conditions:

[0144] Cloning conditions:

[0145] mask i =(∥g i ∥≥τ d )∧(max(s i )≤p d ·E)

[0146] Separation conditions:

[0147] mask i =(∥g i ∥≥τ g )∧(max(s i )>p d ·E)

[0148] in:

[0149] g i :The loss function for u i The gradient vector of

[0150] τ d : Density adjustment threshold;

[0151] s i : size vector;

[0152] E: scene range;

[0153] p d : Densification ratio;

[0154] mask i : A Gaussian sphere that satisfies the conditions.

[0155] Loss function:

[0156]

[0157] This loss function combines L1 loss and SSIM loss, and controls the relative balance between the two through weight parameters;

[0158] in:

[0159] λ: The ratio used to balance the L1 and DSSIM losses.

[0160] represents L1 loss, which calculates the absolute value of the pixel difference between the predicted image and the real image;

[0161] Represents the structural similarity index, which is used to measure the similarity of image structures.

[0162] Adaptive frequency control

[0163] By applying the frequency parameter to the size of the Gaussian ball, the frequency of the Gaussian can be adjusted approximately, and it can be adjusted dynamically according to the regional information without relying on repeated training;

[0164] β i =f(β i )*γ i

[0165] r i =(r g or r d )

[0166] γ i =γ i *r i

[0167] s i =s i *β i

[0168] f(x) = ReLU(2.0*sigmoid(x))

[0169] in:

[0170] β i : Frequency control parameter, used to adjust the frequency of the Gaussian ball, initialized to 0;

[0171] r i : Adjustment factor, select growth according to regional information (r g ) or attenuation (r d );

[0172] γ i : Adaptive density coefficient, initialized to 1; 0;

[0173] f(x): frequency control function, implemented using a combination of ReLU and sigmoid functions;

[0174] Adaptive frequency control focuses on two key areas:

[0175] High-frequency region: When the Gaussian ball is in the high-frequency region, its size tends to be smaller, and the frequency needs to be increased to enhance feature capture.

[0176] Low frequency region: When the Gaussian sphere is in the low frequency region, its size tends to be larger, and the frequency needs to be lowered to reduce over-calculation.

[0177] Adaptive density control determines whether cloning or splitting is required based on the following gradient and scale conditions:

[0178] Gradient condition: If the gradient vector modulus of a Gaussian sphere ∥g i ∥Exceeds the specified threshold τ f , then the area needs further adjustment to improve detail coverage,

[0179] ∥g i ∥≥τ f

[0180] Scale condition: If the maximum scale of a Gaussian ball max(s i ) is less than, equal to, or greater than the product of the scene range and the scale factor,

[0181] max(s i )≤p f ·E

[0182] max(s i )>p f ·E

[0183] Based on the gradient and scale conditions, adaptive density control defines the following cloning and segmentation conditions:

[0184] Growing Conditions:

[0185] mask i =(∥g i ∥≥τ f )∧(max(s i )≤p f ·E)

[0186] Attenuation conditions:

[0187] mask i =(∥g i ∥≥τ f )∧(max(s i )>p f ·E)

[0188] in:

[0189] τ f : Indicates the frequency adjustment threshold;

[0190] p f : represents the frequency ratio;

[0191] This frequency control method allows the number and density of Gaussian spheres to be dynamically adjusted according to the needs of different regions, thereby ensuring the details of regional features while avoiding redundancy.

[0192] Adaptive strategy adjustment

[0193] During the training process, in order to gradually optimize the number and density distribution of Gaussian balls, a phased adaptive strategy adjustment method was adopted. According to the gradient conditions at different stages, the clone-separation and decay-growth strategies were systematically applied to more effectively control the distribution density of Gaussian balls.

[0194] Initial stage 0~ep1

[0195] Strategy: Only clone-separation strategy is applied;

[0196] Goal: Quickly establish a sparse initial distribution of Gaussian spheres to cover the main structures of the scene;

[0197] In this stage, the overall density is increased by cloning and separating Gaussian balls, quickly filling the basic outline of the scene and ensuring initial coverage in key areas. This strategy can effectively improve the density of the initial structure in a short period of time, laying the foundation for subsequent refinement;

[0198] Intermediate stage ep1~ep2

[0199] Strategy: Adaptively select clone-separation or decay-growth strategy according to gradient conditions;

[0200] Goal: Increase density or function frequency in detail-rich areas while maintaining sparsity in flat areas to achieve efficient distribution adjustment;

[0201] When the gradient is large, the clone-separation strategy is used first to increase the density. When the gradient is small, the decay-growth strategy is used to save resources by further adjusting the frequency of the Gaussian balls to reduce redundant Gaussian balls.

[0202] Convergence phase ep2~ep3

[0203] Strategy: Only the decay-growth strategy is applied;

[0204] Objective: Further optimize the model to make the Gaussian sphere distribution more consistent with scene requirements, reduce redundancy and improve the accuracy of detail reconstruction;

[0205] After the model gradually converges, the density of the Gaussian spheres is mainly optimized: in flat areas, the attenuation strategy is used to reduce unnecessary Gaussian spheres and release computing resources; in key detail areas, the growth strategy is used to further increase the density to make the reconstruction results more detailed and accurate. The adjustments at this stage make the Gaussian sphere distribution more suitable for the scene, and ultimately achieve accurate convergence of the model.

[0206] Initial stage 0~ep1

[0207] Strategy: Only clone-separation strategy is applied;

[0208] Goal: Quickly establish a sparse initial distribution of Gaussian spheres to cover the main structures of the scene;

[0209] In this stage, the overall density is increased by cloning and separating Gaussian balls, quickly filling the basic outline of the scene and ensuring initial coverage in key areas. This strategy can effectively improve the density of the initial structure in a short period of time, laying the foundation for subsequent refinement;

[0210] Intermediate stage ep1~ep2

[0211] Strategy: Adaptively select clone-separation or decay-growth strategy according to gradient conditions;

[0212] Goal: Increase density or function frequency in detail-rich areas while maintaining sparsity in flat areas to achieve efficient distribution adjustment;

[0213] When the gradient is large, the clone-separation strategy is used first to increase the density. When the gradient is small, the decay-growth strategy is used to save resources by further adjusting the frequency of the Gaussian balls to reduce redundant Gaussian balls.

[0214] Convergence phase ep2~ep3

[0215] Strategy: Only the decay-growth strategy is applied;

[0216] Objective: Further optimize the model to make the Gaussian sphere distribution more consistent with scene requirements, reduce redundancy and improve the accuracy of detail reconstruction;

[0217] After the model gradually converges, the density of the Gaussian spheres is mainly optimized: in flat areas, the attenuation strategy is used to reduce unnecessary Gaussian spheres and release computing resources; in key detail areas, the growth strategy is used to further increase the density to make the reconstruction results more detailed and accurate. The adjustments at this stage make the Gaussian sphere distribution more suitable for the scene, and ultimately achieve accurate convergence of the model.

[0218] Gaussian Splatter Rendering

[0219] Use Gaussian rendering splash to update the corresponding training parameters.

[0220] The probability density function formula of the three-dimensional Gaussian distribution is as follows:

[0221]

[0222] in:

[0223] n: represents the dimension of the random variable;

[0224] x: represents a column vector containing n elements, which is the observed value of the random variable;

[0225] μ: represents a column vector containing n elements, which is the mean of the distribution;

[0226] Σ: represents an n-order square matrix, which is the covariance matrix;

[0227] The Gaussian splash rendering formula is as follows:

[0228]

[0229] in:

[0230] K: represents the total number of Gaussian balls;

[0231] c k : The color of the kth Gaussian ball;

[0232] α k : The transparency of the kth Gaussian sphere;

[0233] Gaussian distribution value of the 2D projection of the k-th Gaussian ball at position x;

[0234] The cumulative opacity term, that is, the transmittance of all Gaussian balls before the kth Gaussian ball, represents the influence of the previous Gaussian balls on the current Gaussian ball.

[0235] The present invention provides an improved 3DGS algorithm based on a decay-growth adaptive strategy, and initializes a sparse point cloud through COLMAP. After the initialization, the density and frequency of the Gaussian sphere are adjusted by adaptive density control and adaptive frequency control. Finally, the corresponding training parameters are updated by Gaussian rendering splashing. By combining 3DGS and the decay-growth strategy, the reconstruction effect of the model on image details and overall structure is improved, which can meet the industry's demand for high-quality visual reconstruction, bring clearer and more natural images and video content, effectively improve the performance of the intelligent visual system, and improve the user experience, especially in scenes that require realism and clarity.

[0236] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. An improved 3DGS algorithm based on decay-growth adaptive strategy, characterized in that: The following steps are involved: Image initialization: obtain image video frames through the shooting device, extract and match feature points from multi-view images through COLMAP to restore the 3D point cloud, and initialize the Gaussian set; Adaptive density control: After obtaining the sparsely reconstructed Gaussian sphere, the model can be trained. Since the initial number of Gaussians is small, adaptive density control is needed to adjust the number and density of Gaussians, so as to achieve the transition from the initial sparse Gaussian set to a denser set. Adaptive frequency control: By applying the frequency parameter to the size of the Gaussian ball, the frequency of the Gaussian can be adjusted approximately. This can be adjusted dynamically based on regional information without relying on repeated training. Adaptive strategy adjustment. During the training process, in order to gradually optimize the number and density distribution of Gaussian balls, a phased adaptive strategy adjustment method is adopted. According to the gradient conditions at different stages, the clone-separation and decay-growth strategies are systematically applied to more effectively control the distribution density of Gaussian balls. Gaussian splash rendering, uses Gaussian rendering splash to update the corresponding training parameters.

2. The improved 3DGS algorithm based on decay-growth adaptive strategy as claimed in claim 1, characterized in that: The specific steps of image initialization are as follows: Image import: obtain multi-view image data to be tested; Feature matching: Use SIFT feature extraction algorithm to match the extracted feature points; Sparse reconstruction: After feature matching, the position of 3D points is calculated by triangulation, and the camera parameters and 3D point positions are optimized in combination with the BundleAdjustment algorithm to obtain a sparse point cloud. Gaussian ball parameter initialization: Get a sparse point set and initialize the Gaussian ball parameter set {μ i ,sh i ,r i ,s i ,α i ,β i ,γ i } N i＝1 ; in: μ i : represents the center position of the Gaussian sphere, that is, the position of the i-th 3D point extracted from the sparse point cloud; sh i : represents the color of the Gaussian sphere, expressed using spherical harmonic sparse expression; r i : represents the rotation matrix of the Gaussian ball, which is used to control the direction of the Gaussian in three-dimensional space; s i : represents the size parameter of the Gaussian ball, which is used to control the size or range of the Gaussian; α i : Indicates the opacity of the Gaussian sphere, which is used to control the effect of the Gaussian on the image during the rendering process; β i ,γ i : Indicates the parameters that control the frequency adjustment of the Gaussian sphere; N: represents the number of Gaussian balls; Note: β i ,γ i are untrained parameters, and the rest are training parameters.

3. The improved 3DGS algorithm based on decay-growth adaptive strategy as claimed in claim 1, characterized in that: The adaptive density control determines whether cloning or segmentation operations are required through the following gradient and scale conditions: Gradient condition: If the gradient vector modulus of a Gaussian sphere ∥g i ∥Exceeds the specified threshold τ d , then the area needs more Gaussian spheres to improve detail coverage; ∥g i ∥≥τ d Scale condition: If the maximum scale of a Gaussian ball max(s i ) is less than, equal to, or greater than the product of the scene extent and the scale factor; max(s i )≤p d ·E max(s i )>p d ·E Based on the gradient and scale conditions, adaptive density control defines the following cloning and segmentation conditions: Cloning conditions: mask i =(∥g i ∥≥τ d )∧(max(s i )≤p d ·E) Separation conditions: mask i =(∥g i ∥≥τ g )∧(max(s i )>p d ·E) in: g i :The loss function for u i The gradient vector of τ d : Density adjustment threshold; s i : size vector; E: scene range; p d : Densification ratio; mask i : A Gaussian sphere that satisfies the conditions.

4. The improved 3DGS algorithm based on decay-growth adaptive strategy as claimed in claim 3, characterized in that: The loss function: This loss function combines L1 loss and SSIM loss, and controls the relative balance between the two through weight parameters; in: λ: The ratio used to balance the L1 and DSSIM losses. represents L1 loss, which calculates the absolute value of the pixel difference between the predicted image and the real image; Represents the structural similarity index, which is used to measure the similarity of image structures.

5. The improved 3DGS algorithm based on decay-growth adaptive strategy as claimed in claim 1, characterized in that: The specific steps of the adaptive frequency control are: b i =f(β i )*c i r i ＝(r g or r d ) c i =c i *r i s i =s i *b i f(x) = ReLU(2.0*sigmoid(x)) in: β i : Frequency control parameter, used to adjust the frequency of the Gaussian ball, initialized to 0; r i : Adjustment factor, select growth according to regional information (r g ) or attenuation (r d ); γ i : Adaptive density coefficient, initialized to 1; 0; f(x): frequency control function, implemented using a combination of ReLU and sigmoid functions; Adaptive density control determines whether cloning or splitting is required based on the following gradient and scale conditions: Gradient condition: If the gradient vector modulus of a Gaussian sphere ∥g i ∥Exceeds the specified threshold τ f , then the area needs further adjustment to improve detail coverage, ∥g i ∥≥τ f Scale condition: If the maximum scale of a Gaussian ball max(s i ) is less than, equal to, or greater than the product of the scene range and the scale factor, max(s i )≤p f ·E max(s i )>p f ·E Based on the gradient and scale conditions, adaptive density control defines the following cloning and segmentation conditions: Growing Conditions: mask i =(∥g i ∥≥τ f )∧(max(s i )≤p f ·E) Attenuation conditions: mask i =(∥g i ∥≥τ f )∧(max(s i )>p f ·E) in: τ f : Indicates the frequency adjustment threshold; p f : represents the frequency ratio; This frequency control method allows the number and density of Gaussian spheres to be dynamically adjusted according to the needs of different regions, thereby ensuring the details of regional features while avoiding redundancy.

6. The improved 3DGS algorithm based on decay-growth adaptive strategy as claimed in claim 1, characterized in that: The entire training process of the adaptive strategy adjustment is divided into three stages: ep1, ep2 and ep3. The specific strategies are as follows: Initial stage 0~ep1 Strategy: Only clone-separation strategy is applied; Goal: Quickly establish a sparse initial distribution of Gaussian spheres to cover the main structures of the scene; In this stage, the overall density is increased by cloning and separating Gaussian balls, quickly filling the basic outline of the scene and ensuring initial coverage in key areas. This strategy can effectively improve the density of the initial structure in a short period of time, laying the foundation for subsequent refinement; Intermediate stage ep1~ep2 Strategy: Adaptively select clone-separation or decay-growth strategy according to gradient conditions; Goal: Increase density or function frequency in detail-rich areas while maintaining sparsity in flat areas to achieve efficient distribution adjustment; When the gradient is large, the clone-separation strategy is used first to increase the density. When the gradient is small, the decay-growth strategy is used to save resources by further adjusting the frequency of the Gaussian balls to reduce redundant Gaussian balls. Convergence phase ep2~ep3 Strategy: Only the decay-growth strategy is applied; Objective: Further optimize the model to make the Gaussian sphere distribution more consistent with scene requirements, reduce redundancy and improve the accuracy of detail reconstruction; After the model gradually converges, the density of the Gaussian spheres is mainly optimized: in flat areas, the attenuation strategy is used to reduce unnecessary Gaussian spheres and release computing resources; in key detail areas, the growth strategy is used to further increase the density to make the reconstruction results more detailed and accurate. The adjustments at this stage make the Gaussian sphere distribution more suitable for the scene, and ultimately achieve accurate convergence of the model.

7. The improved 3DGS algorithm based on decay-growth adaptive strategy as claimed in claim 1, characterized in that: The probability density function formula of the three-dimensional Gaussian distribution is as follows: in: n: represents the dimension of the random variable; x: represents a column vector containing n elements, which is the observed value of the random variable; μ: represents a column vector containing n elements, which is the mean of the distribution; Σ: represents an n-order square matrix, which is the covariance matrix; The Gaussian splash rendering formula is as follows: in: K: represents the total number of Gaussian balls; c k : The color of the kth Gaussian ball; α k : The transparency of the kth Gaussian sphere; Gaussian distribution value of the 2D projection of the k-th Gaussian ball at position x; The cumulative opacity term, that is, the transmittance of all Gaussian balls before the kth Gaussian ball, represents the influence of the previous Gaussian balls on the current Gaussian ball.

Images