A three-dimensional gaussian radiation field optimization method and system based on energy field guidance and frequency domain perception
By constructing an energy field-guided three-dimensional Gaussian radiation field optimization method, the problems of insufficient density control and frequency domain optimization in three-dimensional Gaussian sputtering technology are solved, realizing intelligent allocation of computing resources and high-quality three-dimensional reconstruction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEAT UNIV OF SCI & TECH
- Filing Date
- 2026-05-13
- Publication Date
- 2026-07-14
AI Technical Summary
Existing 3D Gaussian sputtering technology lacks scene structure awareness in density control, resulting in uneven allocation of computing resources. Furthermore, the loss function lacks frequency domain awareness, leading to insufficient or blurry reconstruction results in areas with complex structures and rich details.
By constructing an energy field based on high-frequency point clouds, adaptive initialization and phased frequency domain loss optimization of a three-dimensional Gaussian sphere are performed. Combined with a hierarchical density control strategy guided by the energy field, computational resources and optimization weights are dynamically adjusted to improve detail fidelity.
It achieves adaptive optimization allocation of computing resources, improves reconstruction accuracy and resource utilization efficiency, ensures high-quality rendering effects from coarse to fine, and improves the stability and convergence efficiency of the training process.
Smart Images
Figure CN122391501A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision, 3D reconstruction and neural rendering, and in particular to a method and system for optimizing 3D Gaussian splatting based on energy field guidance and frequency domain perception. Background Technology
[0002] 3D scene reconstruction and novel perspective synthesis are core tasks in the field of computer vision, with wide applications in virtual reality, augmented reality, digital twins, and digital preservation of cultural heritage. In recent years, Neural Radiance Field (NeRF) has achieved photorealistic rendering effects through implicit neural representations; however, its volumetric rendering-based sampling strategy incurs enormous computational overhead, making training and inference speeds insufficient for real-time applications.
[0003] 3DGaussian Splatting (3DGS), as a next-generation explicit 3D representation method, achieves real-time, high-quality rendering by modeling the scene as a large number of optimizable 3D Gaussian spheres and combining it with tile-based rasterization technology. This method boasts advantages such as fast training speed, high rendering efficiency, and easy editing of explicit geometric representations. However, existing 3DGS technology still has the following key limitations:
[0004] First, density control strategies lack scene structure awareness. Traditional 3DGS relies solely on view space gradients to clone and split 3D Gaussian spheres, failing to effectively utilize the geometric complexity and texture detail distribution information of the scene. This results in insufficient distribution of 3D Gaussian spheres in structurally complex and detail-rich areas, while point redundancy exists in flat and simple areas, leading to an uneven distribution of computational resources.
[0005] Secondly, the loss function lacks frequency domain awareness. The L1 loss and SSIM loss used in standard 3DGS training have the same optimization weights for each frequency component in the frequency domain. They fail to focus on optimizing high-frequency details (such as edges and textures) that are more sensitive to human vision, which leads to blurring in the detailed areas of the reconstruction results, affecting the sharpness and fidelity of the rendered image.
[0006] To address the issues of blind density control and insufficient frequency domain optimization in the existing technologies, this invention proposes a three-dimensional Gaussian radiation field optimization method based on energy field guidance and frequency domain awareness, aiming to significantly improve the detail fidelity and overall rendering quality of three-dimensional reconstruction. Summary of the Invention
[0007] This invention addresses the problems of blind control and insufficient frequency domain optimization in existing technologies by providing a three-dimensional Gaussian radiation field optimization method and system based on energy field guidance and frequency domain awareness, aiming to significantly improve the detail fidelity and overall rendering quality of three-dimensional reconstruction.
[0008] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:
[0009] A three-dimensional Gaussian radiation field optimization method based on energy field guidance and frequency domain sensing includes the following steps:
[0010] Step 1: Energy field construction based on high-frequency point cloud. High-frequency point cloud is obtained from scene structure decomposition. The point cloud is converted into a three-dimensional voxel mesh through voxelization. Energy diffusion and pooling operations are performed using three-dimensional convolution kernels to construct a three-dimensional energy field that reflects the distribution of scene details.
[0011] Step 2: Initialization of the three-dimensional Gaussian sphere guided by the energy field. The initial three-dimensional Gaussian sphere is adaptively adjusted based on the energy field. The spatial points where the three-dimensional Gaussian sphere is located are divided into high-energy, medium-energy and low-energy regions according to the energy value. The Gaussian scale is adaptively initialized. Then, the three-dimensional Gaussian sphere is projected onto the imaging plane and an image is generated using the differentiable rendering module.
[0012] Step 3: Phased frequency domain loss optimization. Design a phased frequency domain perceptual loss function. In the early stage of training, focus on optimizing low-frequency components to ensure the stability of the overall structure. In the later stage of training, enhance the high-frequency weights to improve the sharpness of details. The optimization process from coarse to fine is achieved by dynamically switching high and low frequency weights.
[0013] Step 4: Energy field-guided hierarchical density control. During the training process, the spatial points where the three-dimensional Gaussian sphere is located are divided into high-energy, medium-energy, and low-energy regions based on the energy field. Differentiated density control strategies are implemented: a traditional gradient threshold strategy is used for the low-energy region, an exponential splitting strategy is used for the high-energy region, and a probabilistic uniform sampling splitting strategy is used for the medium-energy region.
[0014] Furthermore, the energy field construction in step one specifically includes: obtaining a set of high-frequency point cloud coordinates from scene structure decomposition, calculating the axial bounding box of the scene, projecting the point cloud onto a three-dimensional voxel grid based on the voxel size, performing energy diffusion and pooling operations on the voxel grid using a three-dimensional convolution kernel, and finally normalizing to obtain the energy field.
[0015] Furthermore, the energy field query in step one specifically includes: converting the coordinates of three-dimensional spatial points to a voxel grid index, checking whether the points are within the bounding box, directly querying the energy value of valid points through the voxel grid, and determining the energy region to which the point belongs based on an adaptive threshold.
[0016] Furthermore, the initialization of the three-dimensional Gaussian sphere in step two specifically includes: dividing the spatial points where the three-dimensional Gaussian sphere is located into a high-energy point set, a medium-energy point set, and a low-energy point set according to energy based on an adaptive threshold; and using a smaller initial scale for the three-dimensional Gaussian sphere in the high-energy region to improve the ability to express details.
[0017] Furthermore, the staged frequency domain loss function in step three is defined as: ,in, It is a low-frequency loss component The weight, It is a high-frequency loss component The weight.
[0018] Furthermore, the medium-energy region probability uniform sampling splitting strategy in step four specifically includes: calculating a probability threshold based on the energy value, and then obtaining a probability value through uniform sampling. When this probability value is greater than the probability threshold, the three-dimensional Gaussian sphere splits.
[0019] Furthermore, the three-dimensional convolutional pooling operation in step one uses a size of A uniform convolution kernel with a stride of 1. This expands the range of energy field perception while maintaining spatial resolution.
[0020] Furthermore, in step one, obtaining the high-frequency point cloud from the scene structure decomposition specifically involves performing wavelet transform on the initial point cloud generated by the motion recovery structure to separate and extract its high-frequency components, thereby obtaining the high-frequency point cloud.
[0021] Furthermore, the low-frequency loss in the staged frequency domain loss function and high frequency loss It is obtained by performing Fast Fourier Transform on the rendered image and the real image respectively, and calculating the difference in the corresponding frequency range in the frequency domain.
[0022] This invention also discloses a three-dimensional Gaussian radiation field optimization system based on energy field guidance and frequency domain sensing, used to implement the above method, the system comprising:
[0023] The energy field construction module is used to obtain high-frequency point clouds from scene structure decomposition and construct a three-dimensional energy field that reflects the distribution of scene details through voxelization and three-dimensional convolution operations.
[0024] The model initialization module is connected to the energy field construction module and is used to perform scale-adaptive initialization of the initial three-dimensional Gaussian sphere based on the three-dimensional energy field.
[0025] A differentiable rendering module, connected to the model initialization module, is used to project the initialized three-dimensional Gaussian sphere onto the imaging plane and perform differentiable rendering to generate an image.
[0026] The frequency domain optimization module, connected to the differentiable rendering module, is used to calculate the staged frequency domain perceptual loss between the rendered image and the real image, and optimize the three-dimensional Gaussian sphere parameters in the scene based on the loss.
[0027] The density control module is connected to the energy field construction module and the frequency domain optimization module, respectively, and is used to implement a hierarchical differentiated density control strategy for the three-dimensional Gaussian sphere based on the three-dimensional energy field during the optimization process.
[0028] Compared with the prior art, the advantages of the present invention are as follows:
[0029] 1. Adaptive optimization allocation of computing resources is achieved: By constructing a three-dimensional energy field that reflects the distribution of scene details, and implementing a layered differential density control strategy for the three-dimensional Gaussian sphere based on this (such as exponential splitting for high-energy regions), this invention can intelligently allocate computing and storage resources (i.e., the three-dimensional Gaussian sphere) to geometrically complex and textured detailed regions, thereby overcoming the problems of insufficient detail reconstruction in complex regions and point redundancy in simple regions in traditional methods, and significantly improving reconstruction accuracy and resource utilization efficiency.
[0030] 2. Achieving a high-quality reconstruction process from coarse to fine: By designing a staged frequency-domain perceptual loss function and dynamically switching the optimization weights of low-frequency and high-frequency components during training, this invention prioritizes ensuring the correctness and stability of the overall structure in the early stages of training, while focusing on enhancing the sharpness of high-frequency details in the later stages. This staged and focused optimization strategy effectively improves the visual fidelity of the final rendered image, especially in edge and texture areas.
[0031] 3. Improved stability and convergence efficiency during training: By introducing mechanisms such as gradient preservation, this invention optimizes training dynamics, helps alleviate instability during the optimization process, and promotes faster and more robust convergence of the model to a better solution, providing a reliable technical foundation for high-quality 3D reconstruction. Attached Figure Description
[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;
[0034] Figure 2 This is a schematic diagram of the energy field construction process according to an embodiment of the present invention;
[0035] Figure 3 This is a schematic diagram of the layered density control strategy according to an embodiment of the present invention;
[0036] Figure 4 This is a schematic diagram illustrating the phased frequency domain loss weight changes according to an embodiment of the present invention. Detailed Implementation
[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0038] like Figure 1 As shown, this embodiment of the invention provides a three-dimensional Gaussian radiation field optimization method based on energy field guidance and frequency domain sensing, the specific implementation of which includes the following steps:
[0039] S101: Constructing a three-dimensional energy field based on high-frequency point clouds.
[0040] This step aims to construct a 3D energy field that quantifies and reflects the spatial distribution of scene geometry and texture details to guide subsequent optimization processes. Specifically, in implementation:
[0041] 1. Obtaining High-Frequency Point Clouds: First, the initial sparse point cloud obtained by the Structure for Motion Restoration (SfM) technique is preprocessed. High-frequency components are extracted from the point cloud using frequency domain decomposition methods such as wavelet transform, resulting in a set of high-frequency point cloud coordinates that primarily contains detailed information about scene edges, corners, and textures.
[0042] 2. Voxelization and Energy Diffusion: The axial bounding box of the high-frequency point cloud is calculated to determine its extent in three-dimensional space. Then, a voxel size is set, and the entire bounding box space is divided into a regular three-dimensional voxel mesh. The high-frequency point cloud is projected onto this mesh, and each voxel containing a high-frequency point is assigned an initial energy value. To smooth the energy distribution and expand the perception range of local details, a voxel with a size of [missing information] is used. Uniform 3D convolution kernels with stride Perform convolution (energy diffusion) and pooling operations on the voxel mesh.
[0043] 3. Generate energy field: Normalize the voxel mesh after convolution and pooling to obtain a three-dimensional scalar field with values in the range [0,1] that reflects the strength of "detail energy" at various points in the scene, i.e., the three-dimensional energy field.
[0044] S102: Energy field-guided initialization and differentiable rendering of a 3D Gaussian sphere.
[0045] like Figure 2 As shown, this step uses the constructed energy field to intelligently initialize the initial model of three-dimensional Gaussian sputtering (3DGS).
[0046] 1. Energy Query and Region Division: For a set of three-dimensional Gaussian spheres generated from the initial SfM point cloud or randomly initialized, query the energy value corresponding to the spatial coordinates of each three-dimensional Gaussian sphere in the three-dimensional energy field constructed in S101. Based on a preset adaptive threshold (which can be dynamically calculated according to the global statistical distribution of energy field values, such as mean, quantiles, etc.), the three-dimensional Gaussian spheres are divided into two categories: a high-energy point set with energy values higher than the threshold (corresponding to complex regions) and a normal point set with energy values lower than or equal to the threshold (corresponding to flat or simple regions).
[0047] 2. Adaptive Scale Initialization: Different initialization strategies are adopted for different point sets obtained from the partitioning. For points in the high-energy point set, a smaller initial spatial scale (i.e., a "smaller" Gaussian sphere) is used so that it can better fit and express fine geometric and texture details in subsequent optimization.
[0048] 3. Projection and Rendering: After initialization, these three-dimensional Gaussian spheres with attributes such as position, scale, rotation, color, and opacity are projected onto a two-dimensional imaging plane according to the camera parameters, and the predicted rendered image is synthesized by a differentiable rasterizer renderer.
[0049] S103: Staged frequency domain loss optimization.
[0050] like Figure 4 As shown, this step designs a frequency-domain-aware loss function to drive model optimization and focuses on different frequency components in stages.
[0051] 1. Calculate the frequency domain loss: Perform Fast Fourier Transform (FFT) on the image rendered in step S102 and the corresponding ground truth image to transform them into the frequency domain. In the frequency domain, divide the spectrum into low-frequency and high-frequency components. Calculate the differences (e.g., L1 or L2 distance) between the rendered image and the ground truth image in the low-frequency and high-frequency bands respectively to obtain the low-frequency loss component. and high-frequency loss components .
[0052] 2. Definition and Optimization of the Staged Loss Function: The overall frequency domain sensing loss function is defined as follows: Among them, weight and It is dynamically changing. In the early stages of training (e.g., the first 50% of iterations), the settings are... (like , This allows the optimization process to focus on reconstructing the correct low-frequency structures, ensuring the stability of the overall scene layout and shape. Later in the training process (after reaching a specified number of iterations), a weight switch is performed, setting... (like This allows the optimization process to focus on enhancing high-frequency details, improving edge sharpness and texture clarity of the rendered image. The loss is then backpropagated to optimize the property parameters of all three-dimensional Gaussian spheres.
[0053] S104: Energy field-guided stratified density control.
[0054] like Figure 3 As shown, this step dynamically and differentially adjusts the density (i.e., quantity and distribution) of the three-dimensional Gaussian spheres based on the energy field during the training process.
[0055] 1. Three-layer region division: During training, in addition to the initial division as in step S102, more refined density control of the 3D Gaussian sphere is continuously performed based on the energy field. According to the energy value, the spatial region to which the 3D Gaussian sphere belongs is divided into three layers: a low-energy region, a medium-energy region, and a high-energy region. This can be achieved by setting two adaptive thresholds.
[0056] 2. Implement a differentiated splitting strategy:
[0057] For low-energy regions: a traditional cloning and splitting strategy based on a view-space location gradient threshold is adopted. Splitting only occurs when the gradient of the 3D Gaussian sphere at the current viewpoint exceeds the threshold, in order to control the number of points in simple regions and avoid redundancy.
[0058] For high-energy regions: an aggressive exponential splitting strategy is adopted. Once a three-dimensional Gaussian sphere is located in this region, the probability or number of its splits will be exponentially positively correlated with its energy value, ensuring that the density of the three-dimensional Gaussian sphere can be increased rapidly and sufficiently in the detailed regions.
[0059] For the mid-energy region: a probabilistic uniform sampling splitting strategy is adopted. Specifically, this involves calculating a probability threshold based on the energy value, and then obtaining a probability value through uniform sampling. The three-dimensional Gaussian sphere splits only when this probability value is greater than the previously calculated probability threshold.
[0060] By iteratively executing steps S102 (rendering), S103 (calculating loss and optimizing parameters), and S104 (adjusting the density of the 3D Gaussian sphere based on the optimized state and energy field), the model can efficiently and effectively reconstruct the 3D scene under the global guidance of the energy field and with the fine-grained drive of frequency domain loss.
[0061] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0062] In another embodiment, a three-dimensional Gaussian radiation field optimization system based on energy field guidance and frequency domain sensing is provided to implement the above method, the system comprising:
[0063] The energy field construction module is used to obtain high-frequency point clouds from scene structure decomposition and construct a three-dimensional energy field that reflects the distribution of scene details through voxelization and three-dimensional convolution operations.
[0064] The model initialization module is connected to the energy field construction module and is used to perform scale-adaptive initialization of the initial three-dimensional Gaussian sphere based on the three-dimensional energy field.
[0065] A differentiable rendering module, connected to the model initialization module, is used to project the initialized three-dimensional Gaussian sphere onto the imaging plane and perform differentiable rendering to generate an image.
[0066] The frequency domain optimization module, connected to the differentiable rendering module, is used to calculate the staged frequency domain perceptual loss between the rendered image and the real image, and optimize the three-dimensional Gaussian sphere parameters in the scene based on the loss.
[0067] The density control module is connected to the energy field construction module and the frequency domain optimization module, respectively, and is used to implement a hierarchical differentiated density control strategy for the three-dimensional Gaussian sphere based on the three-dimensional energy field during the optimization process.
[0068] For specific limitations regarding the three-dimensional Gaussian radiation field optimization system based on energy field guidance and frequency domain awareness, please refer to the limitations of the three-dimensional Gaussian radiation field optimization method based on energy field guidance and frequency domain awareness mentioned above, which will not be repeated here. Each module in the above system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0069] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0071] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A three-dimensional Gaussian radiation field optimization method based on energy field guidance and frequency domain sensing, characterized in that, Includes the following steps: Step 1: Energy field construction based on high-frequency point cloud. High-frequency point cloud is obtained from scene structure decomposition. The point cloud is converted into a three-dimensional voxel mesh through voxelization. Energy diffusion and pooling operations are performed using three-dimensional convolution kernels to construct a three-dimensional energy field that reflects the distribution of scene details. Step 2: Initialization of the three-dimensional Gaussian sphere guided by the energy field. The initial three-dimensional Gaussian sphere is adaptively adjusted based on the energy field. The spatial points where the three-dimensional Gaussian sphere is located are divided into high-energy, medium-energy and low-energy regions according to the energy value. The Gaussian scale is adaptively initialized. Then, the three-dimensional Gaussian sphere is projected onto the imaging plane and an image is generated using the differentiable rendering module. Step 3: Phased frequency domain loss optimization. Design a phased frequency domain perceptual loss function. In the early stage of training, focus on optimizing low-frequency components to ensure the stability of the overall structure. In the later stage of training, enhance the high-frequency weights to improve the sharpness of details. The optimization process from coarse to fine is achieved by dynamically switching high and low frequency weights. Step 4: Energy field-guided hierarchical density control. During the training process, the spatial points where the three-dimensional Gaussian sphere is located are divided into high-energy, medium-energy, and low-energy regions based on the energy field. Differentiated density control strategies are implemented: a traditional gradient threshold strategy is used for the low-energy region, an exponential splitting strategy is used for the high-energy region, and a probabilistic uniform sampling splitting strategy is used for the medium-energy region.
2. The method according to claim 1, characterized in that, The energy field construction in step one specifically includes: obtaining a set of high-frequency point cloud coordinates from scene structure decomposition, calculating the axial bounding box of the scene, projecting the point cloud onto a three-dimensional voxel grid based on the voxel size, performing energy diffusion and pooling operations on the voxel grid using a three-dimensional convolution kernel, and finally normalizing to obtain the energy field.
3. The method according to claim 1, characterized in that, The energy field query in step one specifically includes: converting the coordinates of three-dimensional spatial points to a voxel grid index, checking whether the points are within the bounding box, directly querying the energy value of valid points through the voxel grid, and determining the energy region to which the point belongs based on an adaptive threshold.
4. The method according to claim 1, characterized in that, The initialization of the three-dimensional Gaussian sphere in step two specifically includes: querying the energy value at each position of the initial three-dimensional Gaussian sphere, and dividing the spatial points where the three-dimensional Gaussian sphere is located into high-energy point sets, medium-energy point sets, and low-energy point sets according to the energy based on an adaptive threshold. The three-dimensional Gaussian sphere in the high-energy region adopts a smaller initial scale to improve the ability to express details.
5. The method according to claim 1, characterized in that, The staged frequency domain loss function in step three is defined as follows: ,in, It is a low-frequency loss component The weight, It is a high-frequency loss component The weight.
6. The method according to claim 1, characterized in that, The medium-energy region probability uniform sampling splitting strategy in step four specifically includes: calculating a probability threshold based on the energy value, and then obtaining a probability value through uniform sampling. When this probability value is greater than the probability threshold, the three-dimensional Gaussian sphere splits.
7. The method according to claim 1, characterized in that, The three-dimensional convolutional pooling operation in step one uses a size of... A uniform convolution kernel with a stride of 1. This expands the range of energy field perception while maintaining spatial resolution.
8. The method according to claim 1 or 2, characterized in that, In step one, the high-frequency point cloud is obtained from the scene structure decomposition. Specifically, the initial point cloud generated by the motion recovery structure is subjected to wavelet transform to separate and extract its high-frequency components, thereby obtaining the high-frequency point cloud.
9. The method according to claim 5, characterized in that, The low-frequency loss in the staged frequency domain loss function and high frequency loss It is obtained by performing Fast Fourier Transform on the rendered image and the real image respectively, and calculating the difference in the corresponding frequency range in the frequency domain.
10. A three-dimensional Gaussian radiation field optimization system based on energy field guidance and frequency domain sensing, used to implement the method as described in any one of claims 1-9, characterized in that, The system includes: The energy field construction module is used to obtain high-frequency point clouds from scene structure decomposition and construct a three-dimensional energy field that reflects the distribution of scene details through voxelization and three-dimensional convolution operations. The model initialization module is connected to the energy field construction module and is used to perform scale-adaptive initialization of the initial three-dimensional Gaussian sphere based on the three-dimensional energy field. A differentiable rendering module, connected to the model initialization module, is used to project the initialized three-dimensional Gaussian sphere onto the imaging plane and perform differentiable rendering to generate an image. The frequency domain optimization module, connected to the differentiable rendering module, is used to calculate the staged frequency domain perceptual loss between the rendered image and the real image, and optimize the three-dimensional Gaussian sphere parameters in the scene based on the loss. The density control module is connected to the energy field construction module and the frequency domain optimization module, respectively, and is used to implement a hierarchical differentiated density control strategy for the three-dimensional Gaussian sphere based on the three-dimensional energy field during the optimization process.